Trust in the Robot Brain Server

Trust in the Robot Brain Server | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Trust in the Robot Brain Server Johan F. Hoorn, Ivy. S. Huang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9534725/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract We investigated user trust and self-disclosure behaviours in the context of empathetic robotic companions supported by a Robot Brain Server (RBS), designed to address mental health care gaps in Hong Kong. We conducted a three-wave repeated-measures experiment with 86 young adult participants, comparing interactions with a Nao robot connected either to an Internet-enabled RBS (potentially subject to third-party surveillance) or to a stand-alone, offline RBS. Across three sessions, we measured user perceptions of the RBS in terms of ethics, affordances, valence, relevance, and use intentions, as well as their willingness to disclose personal information and subsequent data deletion behaviour. Multivariate analyses revealed that awareness of third-party surveillance negatively impacted perceived affordances and use intentions during initial interactions, while the offline RBS fostered a greater sense of security and relevance to personal goals. Over time, these differences diminished, suggesting that repeated exposure may mitigate initial privacy concerns. Notably, participants who broke off interactions at deeper levels of self-disclosure consistently rated the RBS lower on all experiential dimensions. Logistic regression indicated that both negative expectations and higher perceived affordances predicted information deletion. Path analyses showed that initial awareness of surveillance became a sustained ethical concern, which, alongside protective functionality, drove intensions to use the RBS via positive or negative outcome expectations. These findings underscore the importance of transparent data management and system design in fostering user trust and engagement with digital mental health technologies. Implications for the deployment of AI-driven caregivers in sensitive domains are discussed. Figures Figure 1 1. Introduction We are developing robot companions to help support mental health in Hong Kong, as there are currently not enough human caretakers. These robots use advanced AI to understand and react to human emotions, giving people who feel low a trustworthy way, a confidante, to share their feelings. Our robots are powered by a system called the Robot Brain Server (RBS) (Hoorn, 2018 ), which is designed to keep personal data safe and private. In this study, we looked at how much participants trusted that the robot would keep their data safe when sharing personal information over a longer period of time. To do so, we compared two situations: one group knew the RBS was connected to the Internet and could potentially be monitored. Another group used a private version of the RBS that was not connected to the Internet. We wanted to answer three Research Questions: RQ 1: If people know that a third party might be watching or listening, does it change how much they trust the robot or how much they are willing to share? RQ 2: Does trust in the RBS increase the more often people use it? RQ 3: What do users think about the ethics implemented by the system, its affordances, how well it works for them (i.e., valence, outcome expectancies), relevance to personal concerns, and whether they would actually use it again (i.e., use intentions)? 2. Related work The health and mental health domains involve sensitive information and high personal risk. People often need to reveal private thoughts, emotions, and health details before a digital system can be helpful. In this sense, trust is both the willingness to “dare to tell” as well as “dare to keep” from the belief that the system will hold what is told safe and confidential. Across healthcare, current AI systems are used mostly in limited or experimental ways, rather than as fully integrated tools. Choudhury and Shamszare ( 2023 ) found that medical students and healthcare professionals were familiar with generative AI, but used it mainly for low-risk academic tasks, not for direct clinical care. Actual use in routine clinical practice remained low, because users were unsure how far they could trust AI outputs and how safe it was to rely on them. In nursing, which is especially relationship-focused, the trust issue is even sharper. Belfrage, Helgesson, and Lynøe ( 2022 ) showed that nurses were cautious about AI, because nursing rests on empathy, moral judgement, and close therapeutic relationships. There was concern that AI could weaken the human side of care or undercut professional judgement. Nurses were more open to AI when it reduced routine work and freed time for patient contact, and they were strongly resistant when it seemed to replace emotional or moral aspects of care. In broader healthcare adoption, Golda et al. ( 2024 ) and Jaspers and Pearson ( 2022 ) showed that trust is the central factor linking users’ views of AI to their willingness to use it. Patients and professionals worried that AI could reduce the personal nature of care, and they needed strong reassurance before accepting AI in roles that affected health outcomes. In other high-stakes fields, similar patterns appeared. For example, Jabbar et al. ( 2023 ) studied Central Bank Digital Currencies. These were still largely in pilot phase, and adoption was slowed by trust and privacy concerns. People were willing to consider using CBDCs only when they believed that their data were protected and that the system was run by credible institutions. Shin, Kee, and Shin ( 2022 ) observed the same gap between potential and actual use of AI across sectors. Many organisations ran pilots, but few moved to full integration. Low trust, concerns about opaque decision-making, and fear of unexpected risks all limited adoption. Taken together, these findings suggest that in an AI-for-health or mental-health companion context, trust is especially critical because: self-disclosure is a precondition for usefulness the information disclosed is highly sensitive the personal and relational aspects of care are central to perceived quality This makes trust and perceptions of ethical handling of data central to how people will judge a mental-health AI or robot system like the RBS, and sets the stage for hypotheses about ethics, willingness to disclose, and deletion behaviour (see H1, H6, H7 by the end of this section). The literature we reviewed shows that trust in AI is not a single thing, but has several dimensions. These dimensions are especially important in high-stakes domains like healthcare (or finance for that matter). Across the studies we read, users needed evidence that AI systems are accurate and dependable. Choudhury and Shamszare ( 2023 ) reported that clinicians worried about generative AI producing convincing but false content. Golda et al. ( 2024 ) found that proof of performance, such as consistent accuracy in image recognition or risk prediction, was vital for trust. Jaspers and Pearson ( 2022 ) showed that people were more willing to rely on AI-enabled devices or service robots when they believed the systems were technically competent. Trust also depended on how data were governed, stored, and protected. Golda et al. ( 2024 ) showed that privacy and security concerns could quickly undermine trust, even if technical performance was good. Users wanted reassurance about data storage, sharing, and protection from breaches. Similar concerns appeared in nursing, where trust in AI was limited by fear of black-box systems and uncertainty about accountability (Belfrage et al., 2022 ). Nurses did not want to be held responsible for errors caused by invisible algorithms. In financial context, users worried about surveillance and lack of anonymity. Jabbar et al. ( 2023 ) found that trust increased when CBDCs were backed by credible institutions and when strong legal protections were in place. Users wanted clear rules and safeguards that limited misuse of their data. Across domains, Shin et al. ( 2022 ) highlighted trust in providers as well as in systems. Trust was affected by perceptions of the organisations behind AI, concerns about bias, and the fear of hidden agendas. They also noted that making AI more human-like could both build trust (cf. CASA, Lee and Nass, 2010 ) as well as cause discomfort (e.g., MacDorman, 2025 ). This underlines the need to handle relational robots and AI in a careful way. Explainability was a recurring theme. Golda et al. ( 2024 ) stressed that users trusted AI more when they could understand how it made decisions. Belfrage et al. ( 2022 ) found that nurses distrusted AI systems whose recommendations could not be explained or justified. Shin et al. ( 2022 ) argued that moving from black-box to explainable AI is key for adoption, especially in areas like healthcare and law. Trust also depended on how AI fitted with professional roles and values. Belfrage et al. ( 2022 ) showed that nurses were more accepting of AI when it supported the nurse-patient relationship and reduced routine tasks, and less accepting when it threatened the core of nursing practice. Choudhury and Shamszare ( 2023 ) found similar patterns among medical professionals, who were more willing to use AI for administrative or preparatory work than for direct patient care or prescribing. Across studies, these dimensions map directly onto core user-experience constructs (Van Vugt et al., 2009 ) such as ethical concerns: privacy, security, fairness, accountability (Golda et al., 2024 ; Belfrage et al., 2022 ; Jabbar et al., 2023 ; Shin et al., 2022 ), but also affordances: what users believe the system can actually do, how easy it is to use, and how reliable it feels in daily practice (Jaspers and Pearson, 2022 ; Jabbar et al., 2023 ; Shin et al., 2022 ). Another important dimension is valence: overall positive or negative feelings based on expected safety, transparency, and emotional comfort (Golda et al., 2024 ; Shin et al., 2022 ; Jaspers and Pearson, 2022 ). Yet another is personal relevance: fit with goals, objectives, tasks, roles, and local constraints such as regulations and workflows (Choudhury and Shamszare, 2023 ; Belfrage et al., 2022 ; Jaspers and Pearson, 2022 ; Shin et al., 2022 ). These themes and dimensions motivate the hypotheses that a stand-alone, unconnected RBS will score higher on ethics (H1), affordances (H2), valence (H3), and relevance (H4), and will strengthen intentions to use the system (H5). The studies we reviewed show that people make disclosure decisions by weighing up the benefits and risks of sharing personal information, in line with privacy calculus theory (Jabbar et al., 2023 ). This is crucial for understanding behaviour in AI companion contexts, where self-disclosure is essential but also risky. In their CBDC study, Jabbar et al. ( 2023 ) found that people were more willing to share data when the expected benefits, such as convenience and reliability, outweighed privacy concerns. Ease of use was the strongest driver of adoption. Many users were prepared to accept some privacy risk if they trusted the institution and the system felt simple and efficient. As an aside, our results showed similar attitudes as accepting “bearable consequences” in the third week of use. In healthcare, similar patterns emerged. Golda et al. ( 2024 ) showed that intentions to use AI were highest when users believed that AI would improve patient outcomes or clinical efficiency, and when they felt that privacy and security were adequately protected. Jaspers and Pearson ( 2022 ) found that performance expectancy and perceived health benefits were key drivers of intention to use AI-enabled devices, especially among people who were already highly motivated about their health. Shin et al. ( 2022 ) reported that individuals’ intentions to use AI were shaped by perceived usefulness, confidence in using technology, and perceived risks. When perceived risk fell, trust rose, and defensive behaviours reduced. This implies that behaviours like withdrawing from a system, refusing to answer, or deleting data can be understood as risk management, not mere “lack of trust.” In nursing, staff were more willing to use AI when it clearly reduced burdens and allowed more time with patients, and less willing when it felt intrusive or undermined professional judgement (Belfrage et al., 2022 ). This again reflects a cost-benefit trade-off: adoption was conditional on AI supporting core values of care. In the context of a robotic mental-health companion, these insights suggest that willingness to disclose personal information (H6) will rise when the system is seen as secure, local, and non-intrusive. Tendency to delete personal information after a session (H7) can be seen as a protective strategy when perceived risks are high, and should fall when users understand that data are stored locally and not transmitted over the Internet. Break-off or avoidance behaviours may reflect rational risk management rather than simple rejection of the technology. This perspective supports interpreting disclosure and deletion as part of an ongoing privacy calculus rather than as a binary trust/no-trust outcome, and aligns with the ethics and affordance-related hypotheses (see further down: H1, H2, H6, H7). Across the literature, there is a clear gap between initial interest in AI and actual sustained use. Several studies highlight the role of time, repeated exposure, and habit formation in closing this gap, and hint that trust may change in non-linear ways. Choudhury and Shamszare ( 2023 ) found that AI use among medical students and professionals was largely exploratory. Users were trying out generative AI tools and learning their strengths and weaknesses, rather than fully adopting them. This suggests, understandably, a phase of cautious experimentation, during which trust and use may fluctuate before stabilising. Belfrage et al. ( 2022 ) showed that many nursing applications of AI remained in pilot or high-tech environments. Nurses’ willingness to use AI depended on seeing, over time, that it did not harm patient relationships or undermine professional skill. Fear of long-term deskilling, where future nurses might become dependent on technology for assessments, made staff more cautious about adoption (the same trends can be observed in education). Trust therefore depended on how AI behaved and was governed in the long run. Golda et al. ( 2024 ) described fragmented implementation across healthcare. Many AI models existed, but actual use was limited to specific domains. Technical integration problems and absent standards slowed the move from pilot to routine use. This implies that even when initial evaluations are positive, use may not increase steadily unless infrastructure, governance, and user support improve over time. Jaspers and Pearson ( 2022 ) found that actual use of AI devices was highest when AI became a habit, as with wearables and fitness trackers. They noted that intentions alone were not enough. Repeated positive experiences, ease of use, and visible health benefits were needed for ongoing use. Social influence also played a role, and could change over time as technologies became more familiar and accepted by clinicians and peers. Across sectors, Shin et al. ( 2022 ) observed that organisations often stayed at the pilot stage. Moving to full integration required long-term investment in infrastructure, skills, and explainable AI, which in turn could shift trust and attitudes over time. They also noted that human-like features could initially raise trust, but later produce discomfort if they felt “too human” (cf. Lee and Nass, 2010 , limited by MacDorman, 2025 ), pointing to possible non-monotonic trust trajectories. The CBDC case mirrors this. Jabbar et al. ( 2023 ) reported that CBDCs were mostly at pilot stage, with adoption dependent on building reliable infrastructure and public familiarity. Over time, if users repeatedly experienced secure and efficient transactions, their willingness to share data and use the system could increase, even in the presence of some residual privacy concerns. In the context of an artificial confidante, these findings suggest that initial reactions may be cautious, and trust may not increase in a straight line. Users may test the system, pull back, and then slowly adopt it as they gain experience. Repeated positive experiences, especially consistent respect for privacy and user control, are likely to strengthen ethics-related evaluations, perceived affordances, positive valence, and relevance to daily life. As the system proves itself over time, intentions to use it and willingness to disclose personal information may increase, while defensive and distancing behaviours like deleting data may decrease. This directly motivates the time-related hypothesis (H8, see next), which proposes that awareness of a stand-alone, unconnected RBS, combined with repeated secure use, will lead to stronger evaluations and greater willingness to use and disclose over time. Overall, the literature indicates that: Ethical and privacy concerns are central barriers to trust and adoption (justifying H1) Perceived affordances, including ease of use and reliability, are key drivers of intentions to use AI (H2 and H5) Positive emotional responses and outcomes expectations of safety (valence) improve when systems are transparent and secure (H3) Adoption is higher when systems fit users’ goals, tasks, and organisational constraints (H4) Willingness to disclose and reduced defensive behaviours follow from higher trust and lower perceived risk (H6 and H7) Trust, comfort, and use tend to evolve with repeated exposure, and can strengthen over time when experiences remain positive and privacy is respected (hence, H8) It is reasonable to assume, then, that awareness of a stand-alone, unconnected RBS will lead to more positive ratings of ethics, affordances, valence, and relevance, stronger intentions to use the robot, greater willingness to disclose, less data deletion, and that these effects will increase with time and familiarity. We also suspect that trust is often not so much “I do not believe AI will bring benefits,” but rather “I worry it might cause harm, errors, liability” – mistrust because of potential damage. Hence, awareness of using an unconnected, stand-alone Robot Brain Server would positively influence people’s evaluation across several important dimensions compared to an Internet-connected RBS. In the remainder of this paper, we will: operationalise third-party awareness as an infrastructure manipulation (online vs. offline RBS), tracking user-experience effects across three waves complement self-reported trust perceptions with behavioural indicators of self-disclosure and privacy management (break-off the session, delete vs keep data) as external indicators of trust/risk management test the model pathway (Ethics/Affordances → Relevance/Valence → Use Intentions) and highlight time dynamics: how the drivers shift across sessions To do so, we will test eight hypotheses, following the user-experience framework by Van Vugt, et al. ( 2009 ): Ethics (H1): Ethical concerns, especially around data privacy and security, are major barriers to trust and adopt AI and digital systems (Golda et al., 2024 ; Belfrage et al., 2022 ; Shin et al., 2022 ). When users know that their data are stored locally and not transmitted over the Internet, they are likely to perceive the system as more ethical. This is because the risk of data breaches, unauthorised sharing, and algorithmic bias is reduced, and users feel more in control of their personal information. Affordances (H2): Affordance refers to the perceived functionality of a system, affecting the intentions to actually use it. Jabbar et al. ( 2023 ) and Jaspers and Pearson ( 2022 ) found that ease of use and perceived control were major drivers of adoption. A stand-alone RBS, which does not rely on complex Cloud infrastructure or constant Internet connectivity, is likely to be seen as more reliable and easier to use, especially in environments with limited Internet access or strict data regulations. Valence (H3): Valence, or the overall positive expectations about the system, is closely tied to trust and perceived safety. The literature shows that users respond more positively to systems that are transparent and secure (Golda et al., Shin et al.). Knowing that a robot operates in stand-alone mode, without sending data externally, increases users’ comfort and positive attitudes towards the technology. Relevance (H4): Relevance is about how well the system fits users’ goals and concerns. Studies highlight that adoption is higher when users feel the technology is tailored to their environment and requirements (Jaspers & Pearson, Shin et al.). A stand-alone RBS can be customised and integrated more easily into existing workflows, making it more relevant to users’ daily routines and organisational policies. Use Intentions (H5): Intentions to use AI and digital systems seem to be strongly influenced by perceived benefits, trust, and ease of use (Jabbar et al., Shin et al.). A stand-alone RBS addresses many of the concerns that hold users back from adopting Internet-connected systems, such as privacy risks and technical complexity. As a result, users are more likely to intend to use the robot regularly. Willingness to Disclose (H6): Jabbar et al. found that willingness to disclose personal information increased when users trust the system and believe their data are secure. A stand-alone RBS, which does not transmit data externally, provides a clear assurance of privacy, making users more comfortable with sharing information. Tendency to Delete Personal Information (H7): When users trust that their data are not being sent to external servers and is handled securely, they are less likely to feel the need to delete personal information after a session. This is supported by findings that lower perceived risk leads to higher trust and less defensive behaviour (Shin et al., Golda et al.). Improvement Over Time (H8): Related work suggests that as users become more familiar with a secure, stand-alone system and see that their privacy is respected, their positive evaluations and willingness to use the system will only increase (Jaspers & Pearson, Belfrage et al.). Trust and comfort tend to grow with repeated positive experiences, especially when ethical and privacy concerns are consistently addressed. Thus, awareness of a stand-alone, unconnected RBS will lead to more positive evaluations across ethics, affordance, valence, relevance, and use intentions. It also increases willingness to disclose personal information, reduces the tendency to delete data, and these effects are likely to strengthen over time as users’ trust and familiarity grow. 3. Methods 3.1. Participants and design After obtaining approval from the institutional Ethical Review Board (filed under HSEARS20220730001), we recruited 150 young adults as voluntary participants (valid cases: N = 86; M age = 22.83, SD age = 2.76, 32.6% male, Asian, 1 European, the bulk being undergraduate and higher). While maintaining the gender ratio of ⅔f and ⅓m over conditions, participants were randomly assigned to two conditions: connected to the Internet and unconnected to the Internet. Participants interacted with a Nao robot in three waves, with an interval of approximately one week between each wave. All participants received supermarket coupons as incentive and for those who finished all three waves, three Hanson’s Professor Einstein™ robots were awarded through a lucky draw. The experimental design, then, had a between-Ss factor ( N = 86) of Third-Party Awareness . Level 1 was with Internet ( n = 42): people were aware that the robot used Wi-Fi and Bluetooth to connect to the RBS and collect information from the Internet such as Cloud services and that therefore, user information could have been visible to governments and companies. Level 2 was No Internet ( n = 44), people being aware that the robot had a cable running to the RBS, which was a stand-alone system that did not use the Internet or Cloud services but handled and stored data locally in a visible vault (i.e., ioSafe Duo Pro). The within-Ss factor was ‘time’ or Wave : we measured at time point t 1 , t 2 , t 3 with intervals of about a week to gauge whether user trust increased with repeated use. Our stop criterion for sampling participants was N = 83. In our repeated measures set-up, participants were to converse with the robot 3 times over about three weeks in 2 between-Ss conditions: RBS wired to the Internet and Cloud services or RBS stand-alone. We calculated G*Power and our sample size was in the clear: given a conventional rejection area ( p < .05), a power of .80, and sample size of N = 83 (so we oversampled 3 more people), the effect sizes were expected to be around η p 2 = .27, which is acceptable for a first experiment on any given research topic. 3.2. Procedure The experiments were conducted in an indoor lab, with two separate rooms, one for the participant, the other for the experimenter (for details, see the technical report in the online supplementary materials). The double door between the two rooms was sound proof, creating a safe environment for participants to freely disclose to the robot. Participants were invited to the experiment, being informed that they would talk to a robot about some of their life events. They came to the experiment room and the functioning of the RBS, either connected to the Internet or not, was explained to them by the experimenter. Appendix 1 in the supplementary materials shows the instructions, giving consent, robot interaction, and running the questionnaire. The purpose of presenting the RBS to the user first was to demonstrate how the server worked, as well as an illustration of what a data vault might be like and how people’s personal information would be protected locally or be freely available in the Cloud. After this explanation of the RBS, participants signed a consent form on an iPad – being the opening of the questionnaire. The consent form told them they would engage with a robot that was safe or with a robot whose safety was not guaranteed (supplementary materials, Appendix 1). At that point, participants decided to click, ‘I agree’ or ‘I do not agree to participate.’ We counted these nominal responses as ‘Decision point zero’ ( Dp0 , yes-no). If they declined, then the experiment was broken off, the participant thanked, and we did not ask for their contribution anymore. They also did not receive a coupon. If participants disagreed to participate in Wave 2 or 3, the same procedure was executed and they would receive a coupon but were excluded from the lucky draw. In other words, who bailed out stayed out and was not asked back again. We did not want people skipping t 1 and doing t 2 , or doing t 1 and t 3 but not t 2 , etc. We made sure that participants understood beforehand that they were rewarded only if they completed a session, not when they broke it off at Dp0 . However, if they stopped at further decision points (supplementary materials, Appendix 1: Dp1 ,…, n ); then that was accepted. We regarded stopping at or after Dp1 as part of the experiment. After all, participants did engage with the robot and the RBS but at later stages may have felt insecure, which we saw as genuine data. As soon as participants filled out the questionnaire, the data counted and they were asked back. After signing the written consent, participants interacted with a Nao robot, standing on a table, participants sitting on a chair, inviting more grave confessions progressively (for details, see the supplementary materials): Dp0 : Giving consent: I agree / do not agree to participate in this study; Dp1 Are you prepared to share some secrets with me? Dp2 Confess an insincerity, a false compliment; Dp3 A real lie you told someone; Dp4 A secret you hardly dare to share with someone else (e.g., secret crush, depression, family issues, doing illegal things). Two Arduino buttons, indicating “Yes” in green and “No” in red, were mounted on the lower deck of the table. During interaction, participants pressed those buttons to respond to the robot’s questions. A TNN 500AF PC host carrying a visible label of “Robot Brain Server” was placed next to the participant’s table. Note that the robot was not driven by the RBS but by the experimenter in a so-called Wizard of Oz set-up. In the control room, lights indicated the participant’s responses, to which the experimenter provided the proper robot actions and respective follow-up questions. Participants were unaware of this set-up and were debriefed after the last session they completed. A completed session took about 15 minutes. At the end of the session, participants could decide to delete all their personal information or to keep it for the next session. After interacting with the robot, participants filled out a questionnaire about their level of trust in the RBS and experiences with the robot (supplementary materials, Appendix 1). The questionnaire was programmed in QuestionPro, an online tool for conducting surveys. 1 We made sure that participants could not skip any item on the questionnaire so to avoid missing values. The interaction and questionnaire were in Cantonese, Mandarin, or English, depending on user preference. 3.3. Apparatus and materials Hoorn ( 2018 ) introduced the Robot Brain Server as a software architecture, handling data and AI on a stand-alone server with edge devices (the robots). If for medical reasons, however, data should be uploaded to the Cloud or retrieved, security is at risk and hence user trust may decline. Yet, software architecture is abstract and to make the RBS tangible to users with no special knowledge of software security, we created an illustrative design. A TNN 500AF PC host was adopted as the carrier of the RBS. We modded this PC with RGB light stripes and a GC9A01 round display to clearly show the user it was processing their data or not. The light stripes indicated interaction and the server’s processing states as follows: breathing light indicated “default,” flashing inward represented “listening,” flashing outward cued “talking,” and randomly shining “thinking” (supplementary materials). A round display was placed inside and in the centre of the PC, indicating being connected to the Internet or not (supplementary materials). An Arduino sound sensor was employed to determine whether the participant was talking. A Nao V5 programmed in Choreographe performed all interactions in all groups (supplementary materials). The Nao robot is a commonly adopted humanoid robot in various technological and academic studies. It is 12.2 inches tall, with speakers, touch sensors, camera, movable limbs, and flexible joints. Choregraphe, a graphical programming environment, was used to vocalise the interaction protocol (supplementary materials, Appendix 1) and to design its movements. The interaction protocol was uploaded to the Nao robot and played through its speakers. The original English version and the translated Mandarin and Cantonese versions of the interaction protocols are available in Appendix 1 of the supplementary materials. Text-to-speech was obtained through the getAvailableVoices function, while the voice fitted Nao’s appearance (i.e., NaoOfficialVoiceEnglish). The protocol had 8 sections, played sequentially based on the user’s choices of pressing the “Yes” or “No” button. Five decision point protocols consisted of a welcoming part, three question parts that asked users to share their life secrets with each question being more personal and distressing, and one closure part that asked participants to choose whether to delete today’s conversation or to save it to the RBS for later use. If the user pressed “No” during an interaction, the end script would be played, showing appreciation of the decision and inviting the participant to fill out the questionnaire on an iPad Pro 2019. At the fifth decision point , the robot would confirm that today’s conversation was saved or deleted, ending the interaction. The robot never answered or responded directly to the participants’ confessions during the interaction. 3.4. Measurements We conducted a 50-item questionnaire on user trust and related experiences (supplementary materials, Appendix 1), adapted from Van Vugt et al. ( 2009 ) and Duan et al. ( 2021 ). Items were Likert-type statements followed by a 6-point rating scale (1 = strongly disagree, 6 = strongly agree). Measurement scales consisted of at least 4 indicative and 4 counter-indicative items, making up measurement scales such as Ethics (8 items, e.g., “RBS is trustworthy”) or Affordances (12 items, e.g., “RBS provides a safe environment”). We then created blocks of items. Blocks could consist of more scales. For instance, Ethics and Affordances made 20 items mixed together in the first block of the survey. The second block was a mix of Relevance (8 items, e.g., “I feel the RBS serves my purposes”) and Valence (10 items, e.g., “I have positive expectations”). The third block of items consisted of Use Intentions (8 items, e.g., “I want to work with the RBS”). Blocks of items were sometimes preceded by an introductory sentence such as ‘I think that the RBS is…,’ which was repeated after every 5 items in a block. Within blocks, items were presented in random order irrespective of the measurement scales they came from, using a different order for each participant. Items were the same in all three waves except for the last block of 4 Demographic items (Gender, Age, Education, Cultural background), which were queried the first time only. For the complete questionnaire, see supplementary materials, Appendix 1. Counter-indicative items were re-coded into new variables: 1→6, 6→1. Before running reliability tests, we checked participants for acquiescence bias and identified those who gave the same responses for all items on a scale, or, contradictory answers for items with opposite meanings. Participants that filled in the questionnaire within unreasonable time (i.e., less than 1 minute or more than 30 minutes) were marked as suspect. These five cases were disregarded in further analyses. Reliability analysis with Cronbach’s α was performed on the data of the first wave and measurement scales showed an acceptable level of reliability (α ≥ .76) in the first run (see supplementary materials). We performed Principal Component Analysis with Promax rotation on the questionnaire items, free fit, and any factor loading lower than .30 was suppressed. Eleven components with eigenvalues > 1 were retained, and it took 141 iterations for the Rotated Component Matrix to converge, indicating that the data were spherical and did not favour any particular direction (Wold et al., 1987 ). Items were removed that showed multiple loadings across different components, or that had no strong loading on any of the components. Details of the PCA results can be found in the supplementary materials. With the remaining items, we ran a second PCA, free fit, with a factor loading threshold of .30. This time, all the items of Ethics, Affordance, Valence, and Use Intentions fell into their own single independent component. The second round of reliability tests was performed on the items of the shortened scales, with Spearman-Brown split half applied to the 2-item Ethics scale. All scales achieved acceptable to good reliability: Ethics (2 items, Spearman-Brown coefficient = .76), Affordance (3 items, Cronbach’s α = .79), Valence (5 items, α = .89), and Use Intentions (3 items, α = .72). As for Relevance, we found a study backing the reliability of a single-item scale (Allen et al., 2022 ) with the correlation of item score from adjoining time points providing insight for scale reliability. The correlations between this one Relevance item (“I believe the technical set-up is worthwhile”) across three time points were all significant. However, this alone cannot provide proficient evidence for its reliability, so this part should be considered as exploratory. With regards to the four scales that were reliable in the first wave, we performed scale analysis and PCA on the second and the third waves as well. The supplementary materials show that acceptable reliability was achieved for all scales at the two subsequent time points. Only Ethics in the third wave obtained a meagre .52. We then tracked outliers with box plots. Since the numbers of items on the scales were different, we calculated the mean M of each scale at every time point. All extremes were negative (see supplementary materials), indicating deep mistrust and anticipating pessimistic outcomes. Consequently, we will use two data sets for further analysis in each wave: 5 extreme cases were found in Wave 1 (i.e., N1 = 81 with outliers; n1 = 76 without outliers). In Wave 2, 4 extreme cases were found (i.e., N2 = 79 with outliers; n2 = 75 without outliers). Lastly, 2 outliers were detected in Wave 3, so that the datasets for analysis will be N3 = 73 with outliers and n3 = 71 without outliers. For details on the outlier analyses, consult the supplementary materials. 4. Analysis and results Eighty-six participants joined the experiment but 5 cases were filtered out, due to acquiescence bias. The supplementary materials show the demographic distribution of the original dataset ( N = 86), and the distributions in all three waves, with and without outliers. In subsequent analyses, however, demographics did not yield significant effects. 4.1. Path analysis of factors influential for Use Intentions A series of path analyses were conducted using Hayes’ PROCESS Procedure for SPSS to examine the factors influencing the mean Use Intentions over three time points (Wave 1, 2, 3) and with the full dataset, respectively. The models tested were based on Van Vugt et al., ( 2009 ), stating that features encoded in Ethics and Affordances predict the Valence and Relevance of features in view of user goals and concerns, which in turn mediate the effects on Use Intentions. The results of the path analyses are summarised in Table 1 , which presents the direct effects (path coefficients) between the variables at each time point and for the full dataset. Be aware that Relevance was but a single-item measurement. Figure 1 depicts the significant paths graphically. For Wave 1, Third-Party Awareness had a significant direct effect on the evaluation of the RBS’ Affordances, but not on the Ethics (cf. its trustworthiness). Both Ethics and Affordances significantly affected Valence. Affordances also were significantly influential for Relevance, but Ethics was not. In turn, Valence and Relevance had significant direct effects on Use Intentions. Thus, indirect effects of Third-Party Awareness on Use Intentions went through the pathways of Affordance → Valence and Affordance → Relevance. Wave 2 showed that Third-Party Awareness no longer had a significant direct effect on Affordances or Ethics (Table 1 , Fig. 1 , second panel). However, both continued to significantly predict Valence and Relevance. Valence had a direct effect on Use Intentions, whereas Relevance did not. Indirect effects of Third-Party Awareness on Use Intentions primarily went via Ethics → Valence and Affordance → Use Intentions. The path analysis of Wave 3 presented similar results but with higher coefficients. In the analysis of the full dataset (Table 1 ), Third-Party Awareness had a significant direct effect on Affordances but not on Ethics. Both Ethics and Affordances significantly predicted Valence and Relevance, which in turn had significant direct effects on Use Intentions. Further details of the path analysis can be found in the supplementary materials, technical report, Section 3.2. Table 1 Direct effects (path coefficients) between experiential variables at different waves and for the full dataset Direct effect Wave 1 ( N1 = 78) Wave 2 ( N2 = 78) Wave 3 ( N3 = 73) All waves coeff se t p coeff se t p coeff se t p coeff se t p 3rd Party Awareness → M Ethics − .031 .209 − .146 .884 − .205 .180 -1.138 .258 − .374 .189 -1.972 .053 − .197 .112 -1.769 .078 3rd Party Awareness → M Affordances − .650 .216 -3.010 .004* − .146 .226 − .647 .520* − .150 .229 − .656 .514 − .323 .130 -2.488 .014* M Eth → M Val .273 .092 2.963 .004* .435 .097 4.473 .000* .653 .096 6.836 .000* .425 .055 7.707 .000* M Eth → M Rel .003 .086 .031 .976 .512 .111 4.632 .000* .661 .105 6.315 .000* .329 .060 5.461 .000* M Aff → M Val .395 .082 4.802 .000* .326 .079 4.126 .000* .550 .078 7.027 .000* .415 .046 9.121 .000* M Aff → M Rel .188 .080 2.346 .022* .418 .088 4.766 .000* .520 .089 5.842 .000* .370 .049 7.502 .000* M Val → M UInt .257 .104 2.5473 .016* .451 .141 3.196 .002* .517 .147 3.515 .001* .361 .072 5.029 .000* M Rel → M UInt .301 .111 2.704 .008* .082 .124 .657 .513 .003 .134 .025 .980 .184 .066 2.807 > .005 Panel 1 up to 3 with results compiled per week. The fourth panel shows results of the repeated measures. Roman numerals refer to effects [uppercase] and trends [lowercase] reported in the article 4.2. Multivariate analyses The full-fledged multivariate analyses, including the testing of assumptions and outlier analyses, are available in the supplementary materials, technical report, Section 3.3. For each wave, significant results as well as interesting trends are depicted in Fig. 1 . We performed three rounds of MANOVA tests. In the first round, we marked those who, across three waves, persistently deleted personal information from those who kept it (i.e., Delete vs Keep). The second round of MANOVA was conducted for each wave separately, using the same division. The third round would be a complement to the second round adding a third group, namely those who broke off a session at deeper levels of self-disclosure. This information-deletion behaviour we termed DelInfo (Delete vs Keep vs Break Off). In the following, we will report a range of significant effects, marked with uppercase Roman numerals in square brackets, relating back to Fig. 1 . Additionally, we will indicate a number of interesting trends , which yet did not achieve statistical significance according to the conventional α = .05 cut-off. Trends are signified by using lowercase Roman numerals in square brackets. Effects and trends are compiled in Fig. 1 , in which Roman numerals are used to refer back to the related statistics discussed next. 4.2.1 MANOVA: Delete vs Keep across three waves Since some participants broke off a session without opting to delete the information or not, which resulted in missing values at the fifth decision point, we included only those participants who, in all three waves, supplied data to the final decision point. We computed the average value over three rounds for the two groups ( DelInfo : 1 = Yes; 2 = No) and to secure a larger sample size, we classified 2⋅Yes and 1⋅No = 1.33 as 1 = Deleter, and 1⋅Yes and 2⋅No = 1.67 as 2 = Keeper. This left us with a sample size of n = 53 (1 = 37; 2 = 16) of those who also completed the fifth (final) decision point. After all assumptions tested, no significant multivariate effects were present, so we can only speak of trends. Tests of between-Ss effects showed a trend for Third-Party Awareness on Affordances ( F (1,43) = 3.90, p = .055, η p 2 = .08) ([i], Fig. 1 , first panel). The univariate ANOVA indicated that Affordances tended to be lower when RBS was Internet-connected, compared to No Internet connection. 4.2.2 MANOVA: Delete vs Keep per wave In Wave 1 ( N1 = 78), no overall significance was found for multivariate effects of DelInfo and Third-Party Awareness . Between-Ss effects indicated that Third-Party Awareness showed a trend for Affordances ( F (1, 57) = 3.80, p = .056, η p 2 = .06) [ii] and Use Intentions ( F (1, 57) = 3.80, p = .056, η p 2 = .06) [iii]. Both mean Affordances and mean Use Intentions were higher without Internet connection compared to being Internet connected. In Wave 1, for MANOVA without outliers ( n1 = 76), significant effects were found for DelInfo1 (Wilks’ Λ = .20, F (5,52) = 2.53, p = .040, η p 2 = .20). No significant overall effect was found for Third-Party Awareness nor any significant interaction effect. Between-Ss effects were not significant. To examine which dependent variable contributed to the overall effect, we ran a one-way MANOVA with DelInfo as the fixed factor. Between-Ss effects were significant for Valence ( F (1,63) = 4.06, p = .048, η p 2 = .06) [I]. A univariate test was performed and Valence was significantly more positive in the Keep-information group ( M = 4.58, SD = .62) than in the Delete-information group ( M = 4.18, SD = .09). In Wave 2 (Fig. 1 , second panel), without outliers, n2 = 74, multivariate tests did not render significant results. In the Between-Ss test, a trend was visible for Third-Party Awareness on Relevance ( F (1,57) = 4.170, p = .046, η p 2 = .07) [iv]. Relevance was higher in the No Internet group compared with Internet-connected. In Wave 3 (Fig. 1 , third panel), N3 = 73, no significant multivariate effects were found. In the Between-Ss tests, trends were found for Third-Party Awareness on Ethics ( F (1,53) = 4.62, p = .036, η p 2 = .08) [v], and for DelInfo on Use Intentions ( F (1,53) = 4.32, p = .043, η p 2 = .08) [vi]. Ethics tended to be lower in the Internet than in the No Internet group; Use Intentions to be higher in the Keep group than in the Delete group. Without outliers ( n3 = 71), we found a trend for DelInfo , indicating that Use Intentions tended to be higher for Keepers compared to Deleters ( F (1,51) = 6.92, p = .011, η p 2 = .12) [vii]. 4.2.3 MANOVA: Delete vs Keep information vs Breaking Off the session We ran MANOVA on the data of all three waves with three groups in DelInfo as fixed factor: those who kept their information for later sessions (Keepers), those who deleted their information after a session (Deleters), and those who broke off the session (Breakers) when more private information was at stake. In Wave 1, we ran MANOVA with DelInfo and Third-Party Awareness as fixed factors. No significant effects were obtained with multivariate tests and the observed trends in the univariate analyses were similar as before in Wave 1. We then ran MANOVA for Wave 2 and Pillai’s Trace was significant for the overall effect of DelInfo on the dependents: V = .27, F (10,130) = 2.02, p = .036, η p 2 = .14. Between-Ss tests showed significant effects of DelInfo on Ethics ( F (2, 68) = 3.76, p = .028, η p 2 = .10) [II] and on Use Intentions ( F (2, 68) = 5.13, p = .008, η p 2 = .13) [III]. Scheffé Post Hoc indicated that mean Ethics was significantly lower for Breakers ( M Δ (Breakers − Deleters) = − .68, p = .026), similar to mean Use Intentions ( M Δ (Breakers − Keepers) = − .61, p = .026) being significantly lower for Breakers than for Keepers. In Wave 2, information deleters seemed to be more concerned about ethics while keepers were more focused on use, whereas breakers were neither convinced about the ethics nor the utility of the RBS when they were asked to go into more intimate levels of confession. In Wave 3, we ran MANOVA, excluding outliers ( n3 = 71), and obtained a significant overall effect of DelInfo : Pillai’s V = .40, F (10, 112) = 2.84, p = .004, η p 2 = .20) [IV]. Between-Ss tests showed that DelInfo3_R3 had a significant effect on Ethics ( F (2, 59) = 8.59, p = .001, η p 2 = .23), on Affordances ( F (2, 59) = 4.40, p = .017, η p 2 = .13), Relevance ( F (2, 59) = 9.18, p = .000, η p 2 = .19), Valence ( F (2, 59) = 6.78, p = .002, η p 2 = .19), and Use Intentions ( F (2, 59) = 7.65, p = .001, η p 2 = .21). The one-way ANOVA with post-hoc Scheffé indicated that all five dependents were significantly lower for Breakers compared to Deleters and to Keepers of private information. Table 2 Pairwise comparisons between DelInfo and Third-Party Awareness Measure Aware I J M Δ(I-J) p M _Relevance Internet Breakers Deleters − .722 .032 No Internet Breakers Deleters -1.207 .002 Breakers Keepers -1.821 .000 Deleters Keepers − .615 .045 The interaction effect of Third-Party Awareness ⋅ DelInfo was significant for Relevance: Wilks’ Λ = .80, F (2, 59) = 3.56, p = .035, η p 2 = .11 [V]. In an ANOVA with pairwise comparisons (Table 2 ), being Internet-connected significantly lowered the mean Relevance of the RBS for people who broke off in the middle of the session as compared to those who merely deleted their information. Without Internet, personal Relevance of the RBS to Breakers was significantly lower than to Deleters and Keepers, while Deleters also thought Relevance was significantly lower to them than to the Keepers of private information. Those who broke off apparently saw lowest personal relevance because they did not dare to confess more private information. Deleters took the middle ground, because they dared to self-disclose very private information but deleted it afterwards for security’s sake. Keepers saw most relevance because they self-disclosed to the deepest level while wanting to maintain their information for the next session, relying on the RBS to safeguard their privacy. 4.3. Repeated measures of Third-Party Awareness and DelInfo across three waves We ran Repeated Measures analysis to examine the differences in experience caused by Third-Party Awareness and DelInfo across three waves. The specifics of these analyses can be checked in the supplementary materials, technical report, Section 3.4. Results of the Repeated Measures are depicted in the fourth panel of Fig. 1 . With N = 73, there was a significant multivariate interaction effect of Wave ⋅ Third-Party Awareness (Pillai’s V = .25, F (10,62) = 2.09, p = .039, η p 2 = .25). Univariate tests showed significant effects of Wave on Affordances ( F (2, 142) = 4.17, p = .017, η p 2 = .06) and of Wave ⋅ Third-Party Awareness on Affordances ( F (2, 142) = 8.31, p = .000, η p 2 = .11) [VI]. Pairwise comparison of the difference scores with Bonferroni correction indicated that Affordances were assessed as significantly lower in Wave 2 and Wave 3 ( M Δ (Wave2–Wave3) = − .25, p = .018). Regarding the interaction, Affordances were significantly lower in Wave 1 compared to Wave 3, in the Internet-connected group ( M Δ (Wave1–Wave3) = − .42, p = .003), but significantly higher in Wave 1 compared to Wave 2 in the No-Internet group ( M Δ (Wave1–Wave2) = .44, p = .002). We excluded the outliers and with n = 63, a significant interaction occurred between Wave and Third-Party Awareness (Pillai’s V = .30, F (10, 52) = 2.23, p = .030, η p 2 = .30). In the Univariate Tests (Greenhouse-Geisser), Wave had a significant effect on Affordances ( F (2, 122) = 3.35, p = .039, η p 2 = .05) and on Valence ( F (2, 122) = 4.23, p = .019, η p 2 = .07). Also the interaction between Wave ⋅ Third-Party Awareness on Affordances was significant: F (2, 122) = 4.11, p = .019, η p 2 = .06. Pairwise comparisons for the differences across timepoints indicated that Valence was significantly lower in Wave 1 compared to Wave 3 ( M Δ (Wave1–Wave3) = − .20, p = .046), while Affordances tended to be lower in Wave 2 than in Wave 3 ( M Δ (Wave2–Wave3) = − .23, p = .054). Again, in the Internet group, Affordances were rated lower in Wave 1 than in Wave 3 ( M Δ (Wave1–Wave3) = − .38, p = .017), but were significantly higher in Wave 1 versus Wave 2 in the No Internet group ( M Δ (Wave1–Wave2) = .30, p = .005). 4.4. Logistic regression: impact of user experience on information-deletion behaviour To investigate the influence of user experience factors (Ethics, Affordances, etc.) on information-deletion behaviour ( DelInfo ), a series of logistic regression analyses were conducted for three waves. Intricacies of these analyses are available in the supplementary materials, technical report, Section 4. Interesting results are exhibited in Fig. 1 . In Wave 1, with the full data set, the logistic regression approached statistical significance: χ 2 (5) = 12.12, p = .033, with Nagelkerke R 2 = .22. Affordances significantly and positively predicted the likelihood of information-deletion behaviour [VII], B = .82, SE = .37, Wald = 4.91, p = .027, Exp(B) = 2.27, counter-intuitively indicating that the higher the perceived ability of the RBS to provide a secure interaction environment, the higher the likelihood of information-deletion behaviour. Additionally, Valence significantly and negatively predicted the likelihood of information-deletion behaviour [VIII], B = -1.32, SE = .53, Wald = 6.34, p = .012, Exp(B) = .27, suggesting that lower perceived Valence was associated with a higher likelihood of deleting information. In Wave 2, the logistic regression model was not statistically significant: χ 2 (5) = 8.84, p = .12, Nagelkerke R 2 = .17. None of the independent variables made a statistically significant contribution to the model. Splitting up Third-Party Awareness into Internet and Not Internet connected groups did not render significant results either. In Wave 3, the logistic regression model was not statistically significant: χ 2 (5) = 6.54, p = .26, Nagelkerke R 2 = .14. None of the predictor variables were statistically significant. However, in Wave 3, participants started to differentiate according to being Internet connected. For those with Internet connection, the regression model was statistically significant: χ 2 (5) = 13.34, p = .020, Nagelkerke R 2 = .49, owing to Use Intentions being the only statistically significant predictor: [IX] B = -2.86, SE = 1.36, Wald = 4.42, p = .036, Exp(B) = .057. Increased Intentions to Use the RBS were associated with a decreased likelihood of information-deletion behaviour. For those without Internet, the model was not statistically significant: χ 2 (5) = 3.23, p = .665, Nagelkerke R 2 = .15. 5. Discussion/conclusions Users interacted with a robot over a time span of three weeks, either connected to the Internet, users being aware of potential third-party surveillance, or stand-alone, guaranteeing utmost privacy. Next, we relate our results back to the hypotheses. H1 (Ethics, Golda et al., 2024 ; Belfrage et al., 2022 ; Shin et al., 2022 ) was supported. While awareness of surveillance did not affect ethical perceptions in week 1, by week 3, participants using the offline (stand-alone) RBS rated it significantly higher on ethics than those in the Internet-connected group. H2 (Affordances, Jabbar et al., 2023 ; Jaspers & Pearson, 2022 ) received mixed support. Initially, in week 1, the offline RBS was perceived to have higher affordances (functionality/security). However, over time, week 3, this reversed; participants in the Internet-connected group rated affordances higher, likely due to the perceived limitations of a stand-alone system. H3 (Valence, Golda et al., Shin et al.) was supported. Higher perceived valence (positive outcome expectancy) was significantly associated with Keepers (those who did not delete data), while those with negative expectations were more likely to delete their information. H4 (Relevance, Jaspers & Pearson; Shin et al.): supported. In week 2, relevance was significantly higher for the offline group. By week 3, a significant interaction showed that Keepers found the stand-alone RBS much more personally relevant to their goals than those who broke off interactions. H5 (Use Intentions, Jabbar et al., Shin et al.): supported. Across the experiment, intentions to use the robot were higher in the offline condition. Path analyses confirmed that Use Intentions were driven by both perceived Ethics and Affordances. H6 (Willingness to Disclose, Jabbar et al.): supported. Participants who felt secure enough to Keep their data reported higher Use Intentions and more positive evaluations, whereas those uncomfortable with the level of disclosure (Breakers) consistently rated the system lower. H7 (Tendency to Delete, Shin et al.; Golda et al.) was confronted with a counter-intuitive result. In week 1, higher perceived affordances (security) actually predicted higher likelihood of deletion. However, by week 3, as hypothesized, higher Use Intentions were associated with a decreased likelihood of deleting information. H8 (Improvement Over Time, Jaspers & Pearson; Belfrage et al.): partially supported. Trust and familiarity evolved; initial awareness of surveillance faded and was internalised as a sustained ethical concern. While some differences diminished, the offline RBS maintained a stronger sense of security for long-term users. What do these results tell us? We tracked participants’ interactions with the RBS over three weeks, revealing a shift from initial caution to a more pragmatic, utility-driven mindset. In week 1, first wave, the initial encounter held a kind of safety paradox. In the first wave, participants’ reactions seem to have been influenced by external factors, specifically whether they were connected to the Internet ( Third-Party Awareness ). Those connected to the Internet, it seems, were more sceptical of the system’s technical capabilities and security (cf. Affordances). During this initial phase, the participants divided into two main behavioural groups: those who wanted to keep their personal data for later use and those who deleted it. The Keepers reported a more positive emotional response (Valence) to the system. Because they felt good about the interaction, they retained their data. The Deleters, maybe, were responsible for the counter-intuitive effect: one would expect that if users trust the RBS’ security (i.e., Affordances), they would feel safe enough to keep their data stored within it. However, the logistic regression showed the opposite: higher perceived Affordances led to a higher likelihood of deleting data. There seems to have been a safety-first effect. In week 1, users who rated the system highly for providing a secure interaction environment likely viewed the functionality to delete data as a core component of that security. They perhaps did not delete the data because they thought the system was broken or unsafe; rather, they used the system’s security features (the delete button) precisely because they were engaging with the system’s affordances to manage their privacy. These users were actively managing their security through the system’s features, rather than passively trusting it. In week 2, second wave, distinct user types started to emerge. The external factor of an Internet connection ceased to be the primary driver of the participants’ views. The users’ internal evaluation of the system (Ethics and Affordances) became the dominant predictor of whether they found the system relevant and fit for future use. Additionally, a third group of users became statistically distinct in this wave, which we baptised Breakers: those who abandoned the session when asked for deep self-disclosure. These individuals began to disengage, rating the system significantly lower on Ethics and Use Intentions. They were unconvinced by the system’s moral standing or its utility. With that, a clear divergence in motivation appeared. Deleters remained concerned with Ethics (privacy and trust), whereas Keepers were focused on Utility (usage and function). Week 3, the third wave, was one of polarisation and pragmatism. By the final wave, the separation between the groups was clear. The Breakers effectively ‘checked out.’ They rated the RBS lowest on almost every metric: Ethics, Affordances, Valence, Relevance, and Use Intentions. For them, the presence of an Internet connection may have acted as a final deterrent, significantly lowering the system’s perceived Relevance. For those who stayed (Keepers and Deleters), behaviour became more pragmatic. The logistic regression showed that for Internet-connected users, the only significant predictor of behaviour was Use Intentions. If they intended to use the system again, they kept the data. If not, they deleted it. In all, over the course of three weeks, our results suggest that participants moved from cautious exploration to pragmatic decision-making. Initially, in the first week, users were hyper-aware of surveillance (i.e., the Internet connection). Those who recognised the system’s security features used them to delete data, exercising control over their privacy. Over time, second and third week, we suspect that the external fear of the Internet faded. Users who trusted the system’s utility became Keepers, while those who doubted its ethics and/or relevance became Deleters or even Breakers. Ultimately, the counter-intuitive deletion behaviour of the first week disappeared. Two weeks later, keeping data was no longer about security perception, but simply about whether the user found the system useful enough to warrant a return visit. 5.1. Limitations While our study offers longer-term insights into trust in robotic confidantes, several limitations apply. Firstly, the use of a Wizard of Oz methodology means participants interacted with a simulated and fully controlled AI rather than an autonomous system; algorithmic latency or errors might alter trust dynamics in real-world deployments. Secondly, the sample consisted primarily of young adult university students in Hong Kong, which may limit the generalisability of the findings to older populations or different cultural contexts where privacy norms vary. Dependent on age, users’ confidence in their own interpretation judgment (trust-in-self) may affect AI and data usage. Older adults often seek confirmation even when information comes from an authoritative source, because they worry, they may have misunderstood. Younger people seem less concerned about loss of privacy. Thirdly, the measurement of Relevance relied on a single-item scale, which, despite showing correlation across waves, lacks the robust psychometric validation of the other multi-item scales. Finally, the three-week duration, while longer than just one session, may not capture the exhaustion of the ‘novelty effect’ or the long-term trust fluctuations that occur over months of, for instance, therapeutic use. 5.2. Practical application Our findings provide a guideline for the design and deployment of AI-driven mental-health technologies. It would be wise to prioritise local processing (i.e., edge computing). To foster initial trust and ethical security, developers should favour stand-alone or edge processing over Cloud-based systems. A visible, offline data vault that they can bring home may reduce user anxiety during sensitive disclosures (Hoorn, 2018 ). For those, however, who want to access the Internet irrespective, security measures should be upgraded (ibid.). It would be important to implement transparent data control. The safety-first effect observed in the first week suggests that providing prominent data-delete functions actually enhances the perception of a secure environment. Users should be given granular control over their data to encourage early-stage engagement. We should design for utility to sustain long-term use. As the initial fear of surveillance faded, use intentions appeared to be driven by perceived utility (affordances). For long-term retention, the system must transition from being merely secure to being demonstrably useful in helping the user achieve their goals, in our case, mental support. Developers must be cautious when the AI solicits highly intimate secrets. Since Breakers (those who abandoned sessions) reported lower trust across all dimensions, therapeutic systems should use incremental disclosure protocols that allow users to build rapport before reaching deeper levels of vulnerability. For social robots to be safe confidants, we should learn how to manage the deep-disclosure threshold. Declarations Acknowledgments This research is funded by the Laboratory for Artificial Intelligence in Design (project code: R2P3), Innovation and Technology Fund, Hong Kong Special Administrative Region Government as well as by the Research Grants Council (project code: T43-518/24-N) under the University Grants Committee, Hong Kong Special Administrative Region Government. We are grateful to Stoney Y. Wang and Yoyo W. Y. Cheung for data acquisition. References Allen MS, Iliescu D, Greiff S (2022) Single item measures in psychological science. Eur J Psychol Assess 38(1):1–5. 10.1027/1015-5759/a000699 Belfrage S, Helgesson G, Lynøe N (2022) Trust and digital privacy in healthcare: a cross-sectional descriptive study of trust and attitudes towards uses of electronic health data among the general public in Sweden. BMC Med Ethics 23(1):19. 10.1186/s12910-022-00758-z Choudhury A, Shamszare H (2023) Investigating the impact of user trust on the adoption and use of ChatGPT: survey analysis. J Med Internet Res 25:e47184. 10.2196/47184 Duan EY, Yoon JM, Liang EZ, Hoorn JF (2021) Self-disclosure to a robot: Only for those who suffer the most. Robotics 10(3):98. 10.3390/robotics10030098 Golda A, Mekonen K, Pandey A, Singh A, Hassija V, Chamola V, Sikdar B (2024) Privacy and security concerns in generative AI: a comprehensive survey. IEEE Access 12:48126–48144. 10.1109/ACCESS.2024.3381611 Green EC, Murphy EM, Gryboski K (2020) The health belief model. In: Sweeny K, Robbins ML, Cohen LM (eds) The Wiley encyclopedia of health psychology. Wiley, New York. doi: 10.1002/9781119057840.ch68 Hoorn JF (2018) The Robot Brain Server, design of a human-artificial systems partnership. In W. Karwowski, & T. Ahram (Eds.), Intelligent Human Systems Integration. Advances in Intelligent Systems and Computing, 722 (pp. 531–536). Berlin, Heidelberg: Springer. 10.1007/978-3-319-73888-8_83 Jabbar A, Geebren A, Hussain Z, Dani S, Ul-Durar S (2023) Investigating individual privacy within CBDC: A privacy calculus perspective. Res Int Bus Finance 64:101826. 10.1016/j.ribaf.2022.101826 Jaspers ED, Pearson E (2022) Consumers’ acceptance of domestic Internet-of-Things: The role of trust and privacy concerns. J Bus Res 142:255–265. 10.1016/j.jbusres.2021.12.043 Lee JER, Nass CI (2010) Trust in computers: The computers-are-social-actors (CASA) paradigm and trustworthiness perception in human-computer communication. In: Latusek D, Gerbasi A (eds) Trust and technology in a ubiquitous modern environment: Theoretical and methodological perspectives. IGI Global Scientific, Hershey, PA, pp 1–15 MacDorman KF (2025) Chikamatsu, Mori, and the uncanny valley. i-Perception 16(1):20416695251317469. 10.1177/2041669525131 Shin D, Kee KF, Shin EY (2022) Algorithm awareness: Why user awareness is critical for personal privacy in the adoption of algorithmic platforms? Int J Inf Manag 65:102494. 10.1016/j.ijinfomgt.2022.102494 Tabachnick BG, Fidell LS, Ullman JB (2013) Using multivariate statistics, vol 6. pearson, Boston, MA, pp 497–516 Van Vugt HC, Hoorn JF, Konijn EA (2009) Interactive engagement with embodied agents: An empirically validated framework. Comput Animat Virtual Worlds 20:195–204. 10.1002/cav.312 Venkatesh V, Morris MG, Davis GB, Davis FD (2003) User acceptance of information technology: Toward a unified view. MIS Q 27(3):425–478. 10.2307/30036540 Wold S, Esbensen K, Geladi P (1987) Principal component analysis. Chemometr Intell Lab Syst 2(1–3):37–52 Footnotes https://www.questionpro.com/ Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9534725","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":629874736,"identity":"7647a8a3-6925-45de-bd7f-caada169d8e1","order_by":0,"name":"Johan F. Hoorn","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABEElEQVRIiWNgGAWjYBACxgYeECXHYADmVjAw8IEZbAwMBjx4tRhDtZwBK8avhYEBWQtjGxFamNt7D34uAGowZ28+9rlw3h05NokcA4YPZYcZzHkOYHdYz7lk6RlALZY9x5Jnz9z2zBikhXHGucMMlr0N2LXMyDGQ5mH4w2BwI8eYmXfb4cQ2oBZm3rbDDAbncXh/Ro7xbx6gLRAtc6Ba/uLXYiaN0NIA1cII0nIWh8N6zphZ8xgY8BicOZbMzHMM6BeeZwUHe86lA0Wwe9+wvcf4Nk+FgZzB8ebDzDw1d+T42ZM3PvhRZi1ncCYBuxaw5QYMsDgAGiyQACIZcEakPBofqJgfu3tGwSgYBaNg5AIAM4pV+0j+LpMAAAAASUVORK5CYII=","orcid":"","institution":"The Hong Kong Polytechnic University","correspondingAuthor":true,"prefix":"","firstName":"Johan","middleName":"F.","lastName":"Hoorn","suffix":""},{"id":629874737,"identity":"a989597f-761d-4ea3-96ee-8ef2d4510705","order_by":1,"name":"Ivy. S. Huang","email":"","orcid":"","institution":"The Chinese University of Hong Kong","correspondingAuthor":false,"prefix":"","firstName":"Ivy.","middleName":"S.","lastName":"Huang","suffix":""}],"badges":[],"createdAt":"2026-04-26 22:41:21","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9534725/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9534725/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107965916,"identity":"b1232f3d-4f93-45b6-ace9-ff5810104804","added_by":"auto","created_at":"2026-04-28 05:41:53","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":289844,"visible":true,"origin":"","legend":"\u003cp\u003ePanel 1 up to 3 with results compiled per week. The fourth panel shows results of the repeated measures. Roman numerals refer to effects [uppercase] and trends [lowercase] reported in the article\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9534725/v1/2ea2572106e4fdff487bd106.png"},{"id":108006789,"identity":"dfc35100-61ea-46fe-8d26-74d2e3b36e4a","added_by":"auto","created_at":"2026-04-28 12:57:13","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":800510,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9534725/v1/d891b229-dcfa-426f-885a-45f8851b4acb.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eTrust in the Robot Brain Server\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eWe are developing robot companions to help support mental health in Hong Kong, as there are currently not enough human caretakers. These robots use advanced AI to understand and react to human emotions, giving people who feel low a trustworthy way, a confidante, to share their feelings.\u003c/p\u003e \u003cp\u003eOur robots are powered by a system called the Robot Brain Server (RBS) (Hoorn, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), which is designed to keep personal data safe and private. In this study, we looked at how much participants trusted that the robot would keep their data safe when sharing personal information over a longer period of time. To do so, we compared two situations: one group knew the RBS was connected to the Internet and could potentially be monitored. Another group used a private version of the RBS that was not connected to the Internet. We wanted to answer three Research Questions:\u003c/p\u003e \u003cp\u003eRQ 1: If people know that a third party might be watching or listening, does it change how much they trust the robot or how much they are willing to share?\u003c/p\u003e \u003cp\u003eRQ 2: Does trust in the RBS increase the more often people use it?\u003c/p\u003e \u003cp\u003eRQ 3: What do users think about the ethics implemented by the system, its affordances, how well it works for them (i.e., valence, outcome expectancies), relevance to personal concerns, and whether they would actually use it again (i.e., use intentions)?\u003c/p\u003e"},{"header":"2. Related work","content":"\u003cp\u003eThe health and mental health domains involve sensitive information and high personal risk. People often need to reveal private thoughts, emotions, and health details before a digital system can be helpful. In this sense, trust is both the willingness to \u0026ldquo;dare to tell\u0026rdquo; as well as \u0026ldquo;dare to keep\u0026rdquo; from the belief that the system will hold what is told safe and confidential.\u003c/p\u003e \u003cp\u003eAcross healthcare, current AI systems are used mostly in limited or experimental ways, rather than as fully integrated tools. Choudhury and Shamszare (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) found that medical students and healthcare professionals were familiar with generative AI, but used it mainly for low-risk academic tasks, not for direct clinical care. Actual use in routine clinical practice remained low, because users were unsure how far they could trust AI outputs and how safe it was to rely on them.\u003c/p\u003e \u003cp\u003eIn nursing, which is especially relationship-focused, the trust issue is even sharper. Belfrage, Helgesson, and Lyn\u0026oslash;e (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) showed that nurses were cautious about AI, because nursing rests on empathy, moral judgement, and close therapeutic relationships. There was concern that AI could weaken the human side of care or undercut professional judgement. Nurses were more open to AI when it reduced routine work and freed time for patient contact, and they were strongly resistant when it seemed to replace emotional or moral aspects of care.\u003c/p\u003e \u003cp\u003eIn broader healthcare adoption, Golda et al. (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) and Jaspers and Pearson (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) showed that trust is the central factor linking users\u0026rsquo; views of AI to their willingness to use it. Patients and professionals worried that AI could reduce the personal nature of care, and they needed strong reassurance before accepting AI in roles that affected health outcomes.\u003c/p\u003e \u003cp\u003eIn other high-stakes fields, similar patterns appeared. For example, Jabbar et al. (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) studied Central Bank Digital Currencies. These were still largely in pilot phase, and adoption was slowed by trust and privacy concerns. People were willing to consider using CBDCs only when they believed that their data were protected and that the system was run by credible institutions.\u003c/p\u003e \u003cp\u003eShin, Kee, and Shin (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) observed the same gap between potential and actual use of AI across sectors. Many organisations ran pilots, but few moved to full integration. Low trust, concerns about opaque decision-making, and fear of unexpected risks all limited adoption.\u003c/p\u003e \u003cp\u003eTaken together, these findings suggest that in an AI-for-health or mental-health companion context, trust is especially critical because:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eself-disclosure is a precondition for usefulness\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003ethe information disclosed is highly sensitive\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003ethe personal and relational aspects of care are central to perceived quality\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThis makes trust and perceptions of ethical handling of data central to how people will judge a mental-health AI or robot system like the RBS, and sets the stage for hypotheses about ethics, willingness to disclose, and deletion behaviour (see H1, H6, H7 by the end of this section).\u003c/p\u003e \u003cp\u003eThe literature we reviewed shows that trust in AI is not a single thing, but has several dimensions. These dimensions are especially important in high-stakes domains like healthcare (or finance for that matter). Across the studies we read, users needed evidence that AI systems are accurate and dependable. Choudhury and Shamszare (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) reported that clinicians worried about generative AI producing convincing but false content. Golda et al. (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) found that proof of performance, such as consistent accuracy in image recognition or risk prediction, was vital for trust. Jaspers and Pearson (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) showed that people were more willing to rely on AI-enabled devices or service robots when they believed the systems were technically competent.\u003c/p\u003e \u003cp\u003eTrust also depended on how data were governed, stored, and protected. Golda et al. (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) showed that privacy and security concerns could quickly undermine trust, even if technical performance was good. Users wanted reassurance about data storage, sharing, and protection from breaches. Similar concerns appeared in nursing, where trust in AI was limited by fear of black-box systems and uncertainty about accountability (Belfrage et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Nurses did not want to be held responsible for errors caused by invisible algorithms.\u003c/p\u003e \u003cp\u003eIn financial context, users worried about surveillance and lack of anonymity. Jabbar et al. (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) found that trust increased when CBDCs were backed by credible institutions and when strong legal protections were in place. Users wanted clear rules and safeguards that limited misuse of their data.\u003c/p\u003e \u003cp\u003eAcross domains, Shin et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) highlighted trust in providers as well as in systems. Trust was affected by perceptions of the organisations behind AI, concerns about bias, and the fear of hidden agendas. They also noted that making AI more human-like could both build trust (cf. CASA, Lee and Nass, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2010\u003c/span\u003e) as well as cause discomfort (e.g., MacDorman, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). This underlines the need to handle relational robots and AI in a careful way.\u003c/p\u003e \u003cp\u003eExplainability was a recurring theme. Golda et al. (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) stressed that users trusted AI more when they could understand how it made decisions. Belfrage et al. (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) found that nurses distrusted AI systems whose recommendations could not be explained or justified. Shin et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) argued that moving from black-box to explainable AI is key for adoption, especially in areas like healthcare and law.\u003c/p\u003e \u003cp\u003eTrust also depended on how AI fitted with professional roles and values. Belfrage et al. (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) showed that nurses were more accepting of AI when it supported the nurse-patient relationship and reduced routine tasks, and less accepting when it threatened the core of nursing practice. Choudhury and Shamszare (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) found similar patterns among medical professionals, who were more willing to use AI for administrative or preparatory work than for direct patient care or prescribing.\u003c/p\u003e \u003cp\u003eAcross studies, these dimensions map directly onto core user-experience constructs (Van Vugt et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2009\u003c/span\u003e) such as ethical concerns: privacy, security, fairness, accountability (Golda et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Belfrage et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Jabbar et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Shin et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e), but also affordances: what users believe the system can actually do, how easy it is to use, and how reliable it feels in daily practice (Jaspers and Pearson, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Jabbar et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Shin et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Another important dimension is valence: overall positive or negative feelings based on expected safety, transparency, and emotional comfort (Golda et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Shin et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Jaspers and Pearson, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Yet another is personal relevance: fit with goals, objectives, tasks, roles, and local constraints such as regulations and workflows (Choudhury and Shamszare, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Belfrage et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Jaspers and Pearson, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Shin et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThese themes and dimensions motivate the hypotheses that a stand-alone, unconnected RBS will score higher on ethics (H1), affordances (H2), valence (H3), and relevance (H4), and will strengthen intentions to use the system (H5).\u003c/p\u003e \u003cp\u003eThe studies we reviewed show that people make disclosure decisions by weighing up the benefits and risks of sharing personal information, in line with privacy calculus theory (Jabbar et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). This is crucial for understanding behaviour in AI companion contexts, where self-disclosure is essential but also risky.\u003c/p\u003e \u003cp\u003eIn their CBDC study, Jabbar et al. (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) found that people were more willing to share data when the expected benefits, such as convenience and reliability, outweighed privacy concerns. Ease of use was the strongest driver of adoption. Many users were prepared to accept some privacy risk if they trusted the institution and the system felt simple and efficient. As an aside, our results showed similar attitudes as accepting \u0026ldquo;bearable consequences\u0026rdquo; in the third week of use.\u003c/p\u003e \u003cp\u003eIn healthcare, similar patterns emerged. Golda et al. (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) showed that intentions to use AI were highest when users believed that AI would improve patient outcomes or clinical efficiency, and when they felt that privacy and security were adequately protected. Jaspers and Pearson (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) found that performance expectancy and perceived health benefits were key drivers of intention to use AI-enabled devices, especially among people who were already highly motivated about their health.\u003c/p\u003e \u003cp\u003eShin et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) reported that individuals\u0026rsquo; intentions to use AI were shaped by perceived usefulness, confidence in using technology, and perceived risks. When perceived risk fell, trust rose, and defensive behaviours reduced. This implies that behaviours like withdrawing from a system, refusing to answer, or deleting data can be understood as risk management, not mere \u0026ldquo;lack of trust.\u0026rdquo;\u003c/p\u003e \u003cp\u003eIn nursing, staff were more willing to use AI when it clearly reduced burdens and allowed more time with patients, and less willing when it felt intrusive or undermined professional judgement (Belfrage et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). This again reflects a cost-benefit trade-off: adoption was conditional on AI supporting core values of care.\u003c/p\u003e \u003cp\u003eIn the context of a robotic mental-health companion, these insights suggest that willingness to disclose personal information (H6) will rise when the system is seen as secure, local, and non-intrusive. Tendency to delete personal information after a session (H7) can be seen as a protective strategy when perceived risks are high, and should fall when users understand that data are stored locally and not transmitted over the Internet. Break-off or avoidance behaviours may reflect rational risk management rather than simple rejection of the technology.\u003c/p\u003e \u003cp\u003eThis perspective supports interpreting disclosure and deletion as part of an ongoing privacy calculus rather than as a binary trust/no-trust outcome, and aligns with the ethics and affordance-related hypotheses (see further down: H1, H2, H6, H7).\u003c/p\u003e \u003cp\u003eAcross the literature, there is a clear gap between initial interest in AI and actual sustained use. Several studies highlight the role of time, repeated exposure, and habit formation in closing this gap, and hint that trust may change in non-linear ways.\u003c/p\u003e \u003cp\u003eChoudhury and Shamszare (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) found that AI use among medical students and professionals was largely exploratory. Users were trying out generative AI tools and learning their strengths and weaknesses, rather than fully adopting them. This suggests, understandably, a phase of cautious experimentation, during which trust and use may fluctuate before stabilising.\u003c/p\u003e \u003cp\u003eBelfrage et al. (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) showed that many nursing applications of AI remained in pilot or high-tech environments. Nurses\u0026rsquo; willingness to use AI depended on seeing, over time, that it did not harm patient relationships or undermine professional skill. Fear of long-term deskilling, where future nurses might become dependent on technology for assessments, made staff more cautious about adoption (the same trends can be observed in education). Trust therefore depended on how AI behaved and was governed in the long run.\u003c/p\u003e \u003cp\u003eGolda et al. (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) described fragmented implementation across healthcare. Many AI models existed, but actual use was limited to specific domains. Technical integration problems and absent standards slowed the move from pilot to routine use. This implies that even when initial evaluations are positive, use may not increase steadily unless infrastructure, governance, and user support improve over time.\u003c/p\u003e \u003cp\u003eJaspers and Pearson (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) found that actual use of AI devices was highest when AI became a habit, as with wearables and fitness trackers. They noted that intentions alone were not enough. Repeated positive experiences, ease of use, and visible health benefits were needed for ongoing use. Social influence also played a role, and could change over time as technologies became more familiar and accepted by clinicians and peers.\u003c/p\u003e \u003cp\u003eAcross sectors, Shin et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) observed that organisations often stayed at the pilot stage. Moving to full integration required long-term investment in infrastructure, skills, and explainable AI, which in turn could shift trust and attitudes over time. They also noted that human-like features could initially raise trust, but later produce discomfort if they felt \u0026ldquo;too human\u0026rdquo; (cf. Lee and Nass, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2010\u003c/span\u003e, limited by MacDorman, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), pointing to possible non-monotonic trust trajectories.\u003c/p\u003e \u003cp\u003eThe CBDC case mirrors this. Jabbar et al. (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) reported that CBDCs were mostly at pilot stage, with adoption dependent on building reliable infrastructure and public familiarity. Over time, if users repeatedly experienced secure and efficient transactions, their willingness to share data and use the system could increase, even in the presence of some residual privacy concerns.\u003c/p\u003e \u003cp\u003eIn the context of an artificial confidante, these findings suggest that initial reactions may be cautious, and trust may not increase in a straight line. Users may test the system, pull back, and then slowly adopt it as they gain experience. Repeated positive experiences, especially consistent respect for privacy and user control, are likely to strengthen ethics-related evaluations, perceived affordances, positive valence, and relevance to daily life. As the system proves itself over time, intentions to use it and willingness to disclose personal information may increase, while defensive and distancing behaviours like deleting data may decrease. This directly motivates the time-related hypothesis (H8, see next), which proposes that awareness of a stand-alone, unconnected RBS, combined with repeated secure use, will lead to stronger evaluations and greater willingness to use and disclose over time. Overall, the literature indicates that:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eEthical and privacy concerns are central barriers to trust and adoption (justifying H1)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003ePerceived affordances, including ease of use and reliability, are key drivers of intentions to use AI (H2 and H5)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003ePositive emotional responses and outcomes expectations of safety (valence) improve when systems are transparent and secure (H3)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eAdoption is higher when systems fit users\u0026rsquo; goals, tasks, and organisational constraints (H4)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWillingness to disclose and reduced defensive behaviours follow from higher trust and lower perceived risk (H6 and H7)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTrust, comfort, and use tend to evolve with repeated exposure, and can strengthen over time when experiences remain positive and privacy is respected (hence, H8)\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eIt is reasonable to assume, then, that awareness of a stand-alone, unconnected RBS will lead to more positive ratings of ethics, affordances, valence, and relevance, stronger intentions to use the robot, greater willingness to disclose, less data deletion, and that these effects will increase with time and familiarity.\u003c/p\u003e \u003cp\u003eWe also suspect that trust is often not so much \u0026ldquo;I do not believe AI will bring benefits,\u0026rdquo; but rather \u0026ldquo;I worry it might cause harm, errors, liability\u0026rdquo; \u0026ndash; mistrust because of potential damage. Hence, awareness of using an unconnected, stand-alone Robot Brain Server would positively influence people\u0026rsquo;s evaluation across several important dimensions compared to an Internet-connected RBS. In the remainder of this paper, we will:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eoperationalise third-party awareness as an infrastructure manipulation (online vs. offline RBS), tracking user-experience effects across three waves\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003ecomplement self-reported trust perceptions with behavioural indicators of self-disclosure and privacy management (break-off the session, delete vs keep data) as external indicators of trust/risk management\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003etest the model pathway (Ethics/Affordances \u0026rarr; Relevance/Valence \u0026rarr; Use Intentions) and highlight time dynamics: how the drivers shift across sessions\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eTo do so, we will test eight hypotheses, following the user-experience framework by Van Vugt, et al. (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2009\u003c/span\u003e):\u003c/p\u003e \u003cp\u003eEthics (H1): Ethical concerns, especially around data privacy and security, are major barriers to trust and adopt AI and digital systems (Golda et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Belfrage et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Shin et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). When users know that their data are stored locally and not transmitted over the Internet, they are likely to perceive the system as more ethical. This is because the risk of data breaches, unauthorised sharing, and algorithmic bias is reduced, and users feel more in control of their personal information.\u003c/p\u003e \u003cp\u003eAffordances (H2): Affordance refers to the perceived functionality of a system, affecting the intentions to actually use it. Jabbar et al. (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) and Jaspers and Pearson (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) found that ease of use and perceived control were major drivers of adoption. A stand-alone RBS, which does not rely on complex Cloud infrastructure or constant Internet connectivity, is likely to be seen as more reliable and easier to use, especially in environments with limited Internet access or strict data regulations.\u003c/p\u003e \u003cp\u003eValence (H3): Valence, or the overall positive expectations about the system, is closely tied to trust and perceived safety. The literature shows that users respond more positively to systems that are transparent and secure (Golda et al., Shin et al.). Knowing that a robot operates in stand-alone mode, without sending data externally, increases users\u0026rsquo; comfort and positive attitudes towards the technology.\u003c/p\u003e \u003cp\u003eRelevance (H4): Relevance is about how well the system fits users\u0026rsquo; goals and concerns. Studies highlight that adoption is higher when users feel the technology is tailored to their environment and requirements (Jaspers \u0026amp; Pearson, Shin et al.). A stand-alone RBS can be customised and integrated more easily into existing workflows, making it more relevant to users\u0026rsquo; daily routines and organisational policies.\u003c/p\u003e \u003cp\u003eUse Intentions (H5): Intentions to use AI and digital systems seem to be strongly influenced by perceived benefits, trust, and ease of use (Jabbar et al., Shin et al.). A stand-alone RBS addresses many of the concerns that hold users back from adopting Internet-connected systems, such as privacy risks and technical complexity. As a result, users are more likely to intend to use the robot regularly.\u003c/p\u003e \u003cp\u003eWillingness to Disclose (H6): Jabbar et al. found that willingness to disclose personal information increased when users trust the system and believe their data are secure. A stand-alone RBS, which does not transmit data externally, provides a clear assurance of privacy, making users more comfortable with sharing information.\u003c/p\u003e \u003cp\u003eTendency to Delete Personal Information (H7): When users trust that their data are not being sent to external servers and is handled securely, they are less likely to feel the need to delete personal information after a session. This is supported by findings that lower perceived risk leads to higher trust and less defensive behaviour (Shin et al., Golda et al.).\u003c/p\u003e \u003cp\u003eImprovement Over Time (H8): Related work suggests that as users become more familiar with a secure, stand-alone system and see that their privacy is respected, their positive evaluations and willingness to use the system will only increase (Jaspers \u0026amp; Pearson, Belfrage et al.). Trust and comfort tend to grow with repeated positive experiences, especially when ethical and privacy concerns are consistently addressed.\u003c/p\u003e \u003cp\u003eThus, awareness of a stand-alone, unconnected RBS will lead to more positive evaluations across ethics, affordance, valence, relevance, and use intentions. It also increases willingness to disclose personal information, reduces the tendency to delete data, and these effects are likely to strengthen over time as users\u0026rsquo; trust and familiarity grow.\u003c/p\u003e"},{"header":"3. Methods","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Participants and design\u003c/h2\u003e \u003cp\u003eAfter obtaining approval from the institutional Ethical Review Board (filed under HSEARS20220730001), we recruited 150 young adults as voluntary participants (valid cases: \u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;86; \u003cem\u003eM\u003c/em\u003e\u003csub\u003eage\u003c/sub\u003e = 22.83, \u003cem\u003eSD\u003c/em\u003e\u003csub\u003eage\u003c/sub\u003e = 2.76, 32.6% male, Asian, 1 European, the bulk being undergraduate and higher). While maintaining the gender ratio of ⅔f and ⅓m over conditions, participants were randomly assigned to two conditions: connected to the Internet and unconnected to the Internet. Participants interacted with a Nao robot in three waves, with an interval of approximately one week between each wave. All participants received supermarket coupons as incentive and for those who finished all three waves, three Hanson\u0026rsquo;s Professor Einstein\u0026trade; robots were awarded through a lucky draw.\u003c/p\u003e \u003cp\u003eThe experimental design, then, had a between-Ss factor (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;86) of \u003cb\u003eThird-Party Awareness\u003c/b\u003e. Level 1 was with Internet (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;42): people were aware that the robot used Wi-Fi and Bluetooth to connect to the RBS and collect information from the Internet such as Cloud services and that therefore, user information could have been visible to governments and companies. Level 2 was No Internet (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;44), people being aware that the robot had a cable running to the RBS, which was a stand-alone system that did not use the Internet or Cloud services but handled and stored data locally in a visible vault (i.e., ioSafe Duo Pro). The within-Ss factor was \u0026lsquo;time\u0026rsquo; or \u003cb\u003eWave\u003c/b\u003e: we measured at time point \u003cem\u003et\u003c/em\u003e\u003csub\u003e\u003cem\u003e1\u003c/em\u003e\u003c/sub\u003e, \u003cem\u003et\u003c/em\u003e\u003csub\u003e\u003cem\u003e2\u003c/em\u003e\u003c/sub\u003e, \u003cem\u003et\u003c/em\u003e\u003csub\u003e\u003cem\u003e3\u003c/em\u003e\u003c/sub\u003e with intervals of about a week to gauge whether user trust increased with repeated use.\u003c/p\u003e \u003cp\u003eOur stop criterion for sampling participants was \u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;83. In our repeated measures set-up, participants were to converse with the robot 3 times over about three weeks in 2 between-Ss conditions: RBS wired to the Internet and Cloud services or RBS stand-alone. We calculated G*Power and our sample size was in the clear: given a conventional rejection area (\u003cem\u003ep\u003c/em\u003e \u0026lt; .05), a power of .80, and sample size of \u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;83 (so we oversampled 3 more people), the effect sizes were expected to be around \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e\u003cem\u003e2\u003c/em\u003e\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.27, which is acceptable for a first experiment on any given research topic.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Procedure\u003c/h2\u003e \u003cp\u003eThe experiments were conducted in an indoor lab, with two separate rooms, one for the participant, the other for the experimenter (for details, see the technical report in the online supplementary materials). The double door between the two rooms was sound proof, creating a safe environment for participants to freely disclose to the robot.\u003c/p\u003e \u003cp\u003eParticipants were invited to the experiment, being informed that they would talk to a robot about some of their life events. They came to the experiment room and the functioning of the RBS, either connected to the Internet or not, was explained to them by the experimenter. Appendix 1 in the supplementary materials shows the instructions, giving consent, robot interaction, and running the questionnaire. The purpose of presenting the RBS to the user first was to demonstrate how the server worked, as well as an illustration of what a data vault might be like and how people\u0026rsquo;s personal information would be protected locally or be freely available in the Cloud.\u003c/p\u003e \u003cp\u003eAfter this explanation of the RBS, participants signed a consent form on an iPad \u0026ndash; being the opening of the questionnaire. The consent form told them they would engage with a robot that was safe or with a robot whose safety was not guaranteed (supplementary materials, Appendix 1). At that point, participants decided to click, \u0026lsquo;I agree\u0026rsquo; or \u0026lsquo;I do not agree to participate.\u0026rsquo; We counted these nominal responses as \u0026lsquo;Decision point zero\u0026rsquo; (\u003cem\u003eDp0\u003c/em\u003e, yes-no). If they declined, then the experiment was broken off, the participant thanked, and we did not ask for their contribution anymore. They also did not receive a coupon.\u003c/p\u003e \u003cp\u003eIf participants disagreed to participate in Wave 2 or 3, the same procedure was executed and they would receive a coupon but were excluded from the lucky draw. In other words, who bailed out stayed out and was not asked back again. We did not want people skipping \u003cem\u003et\u003c/em\u003e\u003csub\u003e\u003cem\u003e1\u003c/em\u003e\u003c/sub\u003e and doing \u003cem\u003et\u003c/em\u003e\u003csub\u003e\u003cem\u003e2\u003c/em\u003e\u003c/sub\u003e, or doing \u003cem\u003et\u003c/em\u003e\u003csub\u003e\u003cem\u003e1\u003c/em\u003e\u003c/sub\u003e and \u003cem\u003et\u003c/em\u003e\u003csub\u003e\u003cem\u003e3\u003c/em\u003e\u003c/sub\u003e but not \u003cem\u003et\u003c/em\u003e\u003csub\u003e\u003cem\u003e2\u003c/em\u003e\u003c/sub\u003e, etc. We made sure that participants understood beforehand that they were rewarded only if they completed a session, not when they broke it off at \u003cem\u003eDp0\u003c/em\u003e. However, if they stopped at further decision points (supplementary materials, Appendix 1: \u003cem\u003eDp1\u003c/em\u003e ,\u0026hellip;, \u003cem\u003en\u003c/em\u003e); then that was accepted. We regarded stopping at or after \u003cem\u003eDp1\u003c/em\u003e as part of the experiment. After all, participants did engage with the robot and the RBS but at later stages may have felt insecure, which we saw as genuine data. As soon as participants filled out the questionnaire, the data counted and they were asked back.\u003c/p\u003e \u003cp\u003eAfter signing the written consent, participants interacted with a Nao robot, standing on a table, participants sitting on a chair, inviting more grave confessions progressively (for details, see the supplementary materials):\u003c/p\u003e \u003cp\u003e \u003cem\u003eDp0\u003c/em\u003e: Giving consent: I agree / do not agree to participate in this study;\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eDp1\u003c/strong\u003e \u003cp\u003eAre you prepared to share some secrets with me?\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eDp2\u003c/strong\u003e \u003cp\u003eConfess an insincerity, a false compliment;\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eDp3\u003c/strong\u003e \u003cp\u003eA real lie you told someone;\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eDp4\u003c/strong\u003e \u003cp\u003eA secret you hardly dare to share with someone else (e.g., secret crush, depression, family issues, doing illegal things).\u003c/p\u003e \u003c/p\u003e \u003cp\u003eTwo Arduino buttons, indicating \u0026ldquo;Yes\u0026rdquo; in green and \u0026ldquo;No\u0026rdquo; in red, were mounted on the lower deck of the table. During interaction, participants pressed those buttons to respond to the robot\u0026rsquo;s questions. A TNN 500AF PC host carrying a visible label of \u0026ldquo;Robot Brain Server\u0026rdquo; was placed next to the participant\u0026rsquo;s table.\u003c/p\u003e \u003cp\u003eNote that the robot was \u003cem\u003enot\u003c/em\u003e driven by the RBS but by the experimenter in a so-called Wizard of Oz set-up. In the control room, lights indicated the participant\u0026rsquo;s responses, to which the experimenter provided the proper robot actions and respective follow-up questions. Participants were unaware of this set-up and were debriefed after the last session they completed. A completed session took about 15 minutes. At the end of the session, participants could decide to delete all their personal information or to keep it for the next session.\u003c/p\u003e \u003cp\u003eAfter interacting with the robot, participants filled out a questionnaire about their level of trust in the RBS and experiences with the robot (supplementary materials, Appendix 1). The questionnaire was programmed in QuestionPro, an online tool for conducting surveys.\u003csup\u003e1\u003c/sup\u003e We made sure that participants could not skip any item on the questionnaire so to avoid missing values. The interaction and questionnaire were in Cantonese, Mandarin, or English, depending on user preference.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Apparatus and materials\u003c/h2\u003e \u003cp\u003eHoorn (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) introduced the Robot Brain Server as a software architecture, handling data and AI on a stand-alone server with edge devices (the robots). If for medical reasons, however, data should be uploaded to the Cloud or retrieved, security is at risk and hence user trust may decline. Yet, software architecture is abstract and to make the RBS tangible to users with no special knowledge of software security, we created an illustrative design.\u003c/p\u003e \u003cp\u003eA TNN 500AF PC host was adopted as the carrier of the RBS. We modded this PC with RGB light stripes and a GC9A01 round display to clearly show the user it was processing their data or not. The light stripes indicated interaction and the server\u0026rsquo;s processing states as follows: \u003cem\u003ebreathing light\u003c/em\u003e indicated \u0026ldquo;default,\u0026rdquo; \u003cem\u003eflashing inward\u003c/em\u003e represented \u0026ldquo;listening,\u0026rdquo; \u003cem\u003eflashing outward\u003c/em\u003e cued \u0026ldquo;talking,\u0026rdquo; and \u003cem\u003erandomly shining\u003c/em\u003e \u0026ldquo;thinking\u0026rdquo; (supplementary materials). A round display was placed inside and in the centre of the PC, indicating being connected to the Internet or not (supplementary materials). An Arduino sound sensor was employed to determine whether the participant was talking.\u003c/p\u003e \u003cp\u003eA Nao V5 programmed in Choreographe performed all interactions in all groups (supplementary materials). The Nao robot is a commonly adopted humanoid robot in various technological and academic studies. It is 12.2 inches tall, with speakers, touch sensors, camera, movable limbs, and flexible joints. Choregraphe, a graphical programming environment, was used to vocalise the interaction protocol (supplementary materials, Appendix 1) and to design its movements. The interaction protocol was uploaded to the Nao robot and played through its speakers.\u003c/p\u003e \u003cp\u003eThe original English version and the translated Mandarin and Cantonese versions of the interaction protocols are available in Appendix 1 of the supplementary materials. Text-to-speech was obtained through the getAvailableVoices function, while the voice fitted Nao\u0026rsquo;s appearance (i.e., NaoOfficialVoiceEnglish).\u003c/p\u003e \u003cp\u003eThe protocol had 8 sections, played sequentially based on the user\u0026rsquo;s choices of pressing the \u0026ldquo;Yes\u0026rdquo; or \u0026ldquo;No\u0026rdquo; button. Five \u003cem\u003edecision point\u003c/em\u003e protocols consisted of a welcoming part, three question parts that asked users to share their life secrets with each question being more personal and distressing, and one closure part that asked participants to choose whether to delete today\u0026rsquo;s conversation or to save it to the RBS for later use. If the user pressed \u0026ldquo;No\u0026rdquo; during an interaction, the end script would be played, showing appreciation of the decision and inviting the participant to fill out the questionnaire on an iPad Pro 2019. At the fifth \u003cem\u003edecision point\u003c/em\u003e, the robot would confirm that today\u0026rsquo;s conversation was saved or deleted, ending the interaction. The robot never answered or responded directly to the participants\u0026rsquo; confessions during the interaction.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.4. Measurements\u003c/h2\u003e \u003cp\u003eWe conducted a 50-item questionnaire on user trust and related experiences (supplementary materials, Appendix 1), adapted from Van Vugt et al. (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2009\u003c/span\u003e) and Duan et al. (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Items were Likert-type statements followed by a 6-point rating scale (1\u0026thinsp;=\u0026thinsp;strongly disagree, 6\u0026thinsp;=\u0026thinsp;strongly agree). Measurement scales consisted of at least 4 indicative and 4 counter-indicative items, making up measurement scales such as Ethics (8 items, e.g., \u0026ldquo;RBS is trustworthy\u0026rdquo;) or Affordances (12 items, e.g., \u0026ldquo;RBS provides a safe environment\u0026rdquo;). We then created blocks of items. Blocks could consist of more scales. For instance, Ethics and Affordances made 20 items mixed together in the first block of the survey. The second block was a mix of Relevance (8 items, e.g., \u0026ldquo;I feel the RBS serves my purposes\u0026rdquo;) and Valence (10 items, e.g., \u0026ldquo;I have positive expectations\u0026rdquo;). The third block of items consisted of Use Intentions (8 items, e.g., \u0026ldquo;I want to work with the RBS\u0026rdquo;).\u003c/p\u003e \u003cp\u003eBlocks of items were sometimes preceded by an introductory sentence such as \u0026lsquo;I think that the RBS is\u0026hellip;,\u0026rsquo; which was repeated after every 5 items in a block. Within blocks, items were presented in random order irrespective of the measurement scales they came from, using a different order for each participant. Items were the same in all three waves except for the last block of 4 Demographic items (Gender, Age, Education, Cultural background), which were queried the first time only. For the complete questionnaire, see supplementary materials, Appendix 1.\u003c/p\u003e \u003cp\u003eCounter-indicative items were re-coded into new variables: 1\u0026rarr;6, 6\u0026rarr;1. Before running reliability tests, we checked participants for acquiescence bias and identified those who gave the same responses for all items on a scale, or, contradictory answers for items with opposite meanings. Participants that filled in the questionnaire within unreasonable time (i.e., less than 1 minute or more than 30 minutes) were marked as suspect. These five cases were disregarded in further analyses. Reliability analysis with Cronbach\u0026rsquo;s α was performed on the data of the first wave and measurement scales showed an acceptable level of reliability (α \u0026ge; .76) in the first run (see supplementary materials).\u003c/p\u003e \u003cp\u003eWe performed Principal Component Analysis with Promax rotation on the questionnaire items, free fit, and any factor loading lower than .30 was suppressed. Eleven components with eigenvalues\u0026thinsp;\u0026gt;\u0026thinsp;1 were retained, and it took 141 iterations for the Rotated Component Matrix to converge, indicating that the data were spherical and did not favour any particular direction (Wold et al., \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e1987\u003c/span\u003e). Items were removed that showed multiple loadings across different components, or that had no strong loading on any of the components. Details of the PCA results can be found in the supplementary materials.\u003c/p\u003e \u003cp\u003eWith the remaining items, we ran a second PCA, free fit, with a factor loading threshold of .30. This time, all the items of Ethics, Affordance, Valence, and Use Intentions fell into their own single independent component.\u003c/p\u003e \u003cp\u003eThe second round of reliability tests was performed on the items of the shortened scales, with Spearman-Brown split half applied to the 2-item Ethics scale. All scales achieved acceptable to good reliability: Ethics (2 items, Spearman-Brown coefficient = .76), Affordance (3 items, Cronbach\u0026rsquo;s \u003cem\u003eα\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.79), Valence (5 items, \u003cem\u003eα\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.89), and Use Intentions (3 items, \u003cem\u003eα\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.72). As for Relevance, we found a study backing the reliability of a single-item scale (Allen et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) with the correlation of item score from adjoining time points providing insight for scale reliability. The correlations between this one Relevance item (\u0026ldquo;I believe the technical set-up is worthwhile\u0026rdquo;) across three time points were all significant. However, this alone cannot provide proficient evidence for its reliability, so this part should be considered as exploratory.\u003c/p\u003e \u003cp\u003eWith regards to the four scales that were reliable in the first wave, we performed scale analysis and PCA on the second and the third waves as well. The supplementary materials show that acceptable reliability was achieved for all scales at the two subsequent time points. Only Ethics in the third wave obtained a meagre .52.\u003c/p\u003e \u003cp\u003eWe then tracked outliers with box plots. Since the numbers of items on the scales were different, we calculated the mean \u003cem\u003eM\u003c/em\u003e of each scale at every time point. All extremes were negative (see supplementary materials), indicating deep mistrust and anticipating pessimistic outcomes. Consequently, we will use two data sets for further analysis in each wave: 5 extreme cases were found in Wave 1 (i.e., \u003cem\u003eN1\u003c/em\u003e\u0026thinsp;=\u0026thinsp;81 with outliers; \u003cem\u003en1\u003c/em\u003e\u0026thinsp;=\u0026thinsp;76 without outliers). In Wave 2, 4 extreme cases were found (i.e., \u003cem\u003eN2\u003c/em\u003e\u0026thinsp;=\u0026thinsp;79 with outliers; \u003cem\u003en2\u003c/em\u003e\u0026thinsp;=\u0026thinsp;75 without outliers). Lastly, 2 outliers were detected in Wave 3, so that the datasets for analysis will be \u003cem\u003eN3\u003c/em\u003e\u0026thinsp;=\u0026thinsp;73 with outliers and \u003cem\u003en3\u003c/em\u003e\u0026thinsp;=\u0026thinsp;71 without outliers. For details on the outlier analyses, consult the supplementary materials.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Analysis and results","content":"\u003cp\u003eEighty-six participants joined the experiment but 5 cases were filtered out, due to acquiescence bias. The supplementary materials show the demographic distribution of the original dataset (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;86), and the distributions in all three waves, with and without outliers. In subsequent analyses, however, demographics did not yield significant effects.\u003c/p\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Path analysis of factors influential for Use Intentions\u003c/h2\u003e \u003cp\u003eA series of path analyses were conducted using Hayes\u0026rsquo; PROCESS Procedure for SPSS to examine the factors influencing the mean Use Intentions over three time points (Wave 1, 2, 3) and with the full dataset, respectively. The models tested were based on Van Vugt et al., (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2009\u003c/span\u003e), stating that features encoded in Ethics and Affordances predict the Valence and Relevance of features in view of user goals and concerns, which in turn mediate the effects on Use Intentions.\u003c/p\u003e \u003cp\u003eThe results of the path analyses are summarised in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, which presents the direct effects (path coefficients) between the variables at each time point and for the full dataset. Be aware that Relevance was but a single-item measurement. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e depicts the significant paths graphically.\u003c/p\u003e \u003cp\u003eFor Wave 1, \u003cem\u003eThird-Party Awareness\u003c/em\u003e had a significant direct effect on the evaluation of the RBS\u0026rsquo; Affordances, but not on the Ethics (cf. its trustworthiness). Both Ethics and Affordances significantly affected Valence. Affordances also were significantly influential for Relevance, but Ethics was not. In turn, Valence and Relevance had significant direct effects on Use Intentions. Thus, indirect effects of \u003cem\u003eThird-Party Awareness\u003c/em\u003e on Use Intentions went through the pathways of Affordance \u0026rarr; Valence and Affordance \u0026rarr; Relevance.\u003c/p\u003e \u003cp\u003eWave 2 showed that \u003cem\u003eThird-Party Awareness\u003c/em\u003e no longer had a significant direct effect on Affordances or Ethics (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, second panel). However, both continued to significantly predict Valence and Relevance. Valence had a direct effect on Use Intentions, whereas Relevance did not. Indirect effects of \u003cem\u003eThird-Party Awareness\u003c/em\u003e on Use Intentions primarily went via Ethics \u0026rarr; Valence and Affordance \u0026rarr; Use Intentions. The path analysis of Wave 3 presented similar results but with higher coefficients.\u003c/p\u003e \u003cp\u003eIn the analysis of the full dataset (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), \u003cem\u003eThird-Party Awareness\u003c/em\u003e had a significant direct effect on Affordances but not on Ethics. Both Ethics and Affordances significantly predicted Valence and Relevance, which in turn had significant direct effects on Use Intentions. Further details of the path analysis can be found in the supplementary materials, technical report, Section 3.2.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDirect effects (path coefficients) between experiential variables at different waves and for the full dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"20\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c12\" colnum=\"12\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c13\" colnum=\"13\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c14\" colnum=\"14\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c15\" colnum=\"15\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c16\" colnum=\"16\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c17\" colnum=\"17\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c18\" colnum=\"18\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c19\" colnum=\"19\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c20\" colnum=\"20\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eDirect effect\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"4\" nameend=\"c5\" namest=\"c2\"\u003e \u003cp\u003eWave 1 (\u003cem\u003eN1\u003c/em\u003e\u0026thinsp;=\u0026thinsp;78)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"4\" nameend=\"c10\" namest=\"c7\"\u003e \u003cp\u003eWave 2 (\u003cem\u003eN2\u003c/em\u003e\u0026thinsp;=\u0026thinsp;78)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"4\" nameend=\"c15\" namest=\"c12\"\u003e \u003cp\u003eWave 3 (\u003cem\u003eN3\u003c/em\u003e\u0026thinsp;=\u0026thinsp;73)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c16\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"4\" nameend=\"c20\" namest=\"c17\"\u003e \u003cp\u003eAll waves\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003ecoeff\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003ese\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003et\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cem\u003ecoeff\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u003cem\u003ese\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003e\u003cem\u003et\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c10\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c12\"\u003e \u003cp\u003e\u003cem\u003ecoeff\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c13\"\u003e \u003cp\u003e\u003cem\u003ese\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c14\"\u003e \u003cp\u003e\u003cem\u003et\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c15\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c16\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c17\"\u003e \u003cp\u003e\u003cem\u003ecoeff\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c18\"\u003e \u003cp\u003e\u003cem\u003ese\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c19\"\u003e \u003cp\u003e\u003cem\u003et\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c20\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003e3rd Party Awareness\u003c/em\u003e \u0026rarr; \u003cem\u003eM\u003c/em\u003eEthics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.031\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.209\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.146\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.884\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.205\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e.180\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e-1.138\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e.258\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.374\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e.189\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c14\"\u003e \u003cp\u003e-1.972\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e.053\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c16\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c17\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.197\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c18\"\u003e \u003cp\u003e.112\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c19\"\u003e \u003cp\u003e-1.769\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c20\"\u003e \u003cp\u003e.078\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003e3rd Party Awareness\u003c/em\u003e \u0026rarr; \u003cem\u003eM\u003c/em\u003eAffordances\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.650\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.216\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-3.010\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.004*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.146\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e.226\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.647\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e.520*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.150\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e.229\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c14\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.656\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e.514\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c16\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c17\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.323\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c18\"\u003e \u003cp\u003e.130\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c19\"\u003e \u003cp\u003e-2.488\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c20\"\u003e \u003cp\u003e.014*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eM\u003c/em\u003eEth \u0026rarr; \u003cem\u003eM\u003c/em\u003eVal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.273\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.092\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.963\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.004*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e.435\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e.097\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e4.473\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e.653\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e.096\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c14\"\u003e \u003cp\u003e6.836\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c16\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c17\"\u003e \u003cp\u003e.425\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c18\"\u003e \u003cp\u003e.055\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c19\"\u003e \u003cp\u003e7.707\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c20\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eM\u003c/em\u003eEth \u0026rarr; \u003cem\u003eM\u003c/em\u003eRel\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.003\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.086\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.031\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.976\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e.512\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e.111\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e4.632\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e.661\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e.105\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c14\"\u003e \u003cp\u003e6.315\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c16\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c17\"\u003e \u003cp\u003e.329\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c18\"\u003e \u003cp\u003e.060\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c19\"\u003e \u003cp\u003e5.461\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c20\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eM\u003c/em\u003eAff \u0026rarr; \u003cem\u003eM\u003c/em\u003eVal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.395\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.082\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.802\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e.326\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e.079\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e4.126\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e.550\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e.078\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c14\"\u003e \u003cp\u003e7.027\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c16\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c17\"\u003e \u003cp\u003e.415\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c18\"\u003e \u003cp\u003e.046\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c19\"\u003e \u003cp\u003e9.121\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c20\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eM\u003c/em\u003eAff \u0026rarr; \u003cem\u003eM\u003c/em\u003eRel\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.188\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.080\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.346\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.022*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e.418\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e.088\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e4.766\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e.520\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e.089\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c14\"\u003e \u003cp\u003e5.842\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c16\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c17\"\u003e \u003cp\u003e.370\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c18\"\u003e \u003cp\u003e.049\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c19\"\u003e \u003cp\u003e7.502\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c20\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eM\u003c/em\u003eVal \u0026rarr; \u003cem\u003eM\u003c/em\u003eUInt\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.257\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.104\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.5473\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.016*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e.451\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e.141\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e3.196\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e.002*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e.517\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e.147\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c14\"\u003e \u003cp\u003e3.515\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e.001*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c16\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c17\"\u003e \u003cp\u003e.361\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c18\"\u003e \u003cp\u003e.072\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c19\"\u003e \u003cp\u003e5.029\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c20\"\u003e \u003cp\u003e.000*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eM\u003c/em\u003eRel \u0026rarr; \u003cem\u003eM\u003c/em\u003eUInt\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.301\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.111\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.704\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.008*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e.082\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e.124\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e.657\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e.513\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e.003\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e.134\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c14\"\u003e \u003cp\u003e.025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e.980\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c16\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c17\"\u003e \u003cp\u003e.184\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c18\"\u003e \u003cp\u003e.066\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c19\"\u003e \u003cp\u003e2.807\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c20\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;.005\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003ePanel 1 up to 3 with results compiled per week. The fourth panel shows results of the repeated measures. Roman numerals refer to effects [uppercase] and trends [lowercase] reported in the article\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e4.2. Multivariate analyses\u003c/h2\u003e \u003cp\u003eThe full-fledged multivariate analyses, including the testing of assumptions and outlier analyses, are available in the supplementary materials, technical report, Section 3.3. For each wave, significant results as well as interesting trends are depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eWe performed three rounds of MANOVA tests. In the first round, we marked those who, across three waves, persistently deleted personal information from those who kept it (i.e., Delete vs Keep). The second round of MANOVA was conducted for each wave separately, using the same division. The third round would be a complement to the second round adding a third group, namely those who broke off a session at deeper levels of self-disclosure. This information-deletion behaviour we termed \u003cem\u003eDelInfo\u003c/em\u003e (Delete vs Keep vs Break Off).\u003c/p\u003e \u003cp\u003eIn the following, we will report a range of significant effects, marked with uppercase Roman numerals in square brackets, relating back to Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Additionally, we will indicate a number of interesting \u003cem\u003etrends\u003c/em\u003e, which yet did not achieve statistical significance according to the conventional \u003cem\u003eα\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.05 cut-off. Trends are signified by using lowercase Roman numerals in square brackets. Effects and trends are compiled in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, in which Roman numerals are used to refer back to the related statistics discussed next.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003e4.2.1 MANOVA: Delete vs Keep across three waves\u003c/h2\u003e \u003cp\u003eSince some participants broke off a session without opting to delete the information or not, which resulted in missing values at the fifth decision point, we included only those participants who, in all three waves, supplied data to the final decision point. We computed the average value over three rounds for the two groups (\u003cem\u003eDelInfo\u003c/em\u003e: 1\u0026thinsp;=\u0026thinsp;Yes; 2\u0026thinsp;=\u0026thinsp;No) and to secure a larger sample size, we classified 2\u0026sdot;Yes and 1\u0026sdot;No\u0026thinsp;=\u0026thinsp;1.33 as 1\u0026thinsp;=\u0026thinsp;Deleter, and 1\u0026sdot;Yes and 2\u0026sdot;No\u0026thinsp;=\u0026thinsp;1.67 as 2\u0026thinsp;=\u0026thinsp;Keeper. This left us with a sample size of \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;53 (1\u0026thinsp;=\u0026thinsp;37; 2\u0026thinsp;=\u0026thinsp;16) of those who also completed the fifth (final) decision point.\u003c/p\u003e \u003cp\u003eAfter all assumptions tested, no significant multivariate effects were present, so we can only speak of trends. Tests of between-Ss effects showed a trend for \u003cem\u003eThird-Party Awareness\u003c/em\u003e on Affordances (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(1,43)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;3.90, \u003cem\u003ep\u003c/em\u003e = .055, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.08) ([i], Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, first panel). The univariate ANOVA indicated that Affordances tended to be lower when RBS was Internet-connected, compared to No Internet connection.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e4.2.2 MANOVA: Delete vs Keep per wave\u003c/h2\u003e \u003cp\u003eIn Wave 1 (\u003cem\u003eN1\u003c/em\u003e\u0026thinsp;=\u0026thinsp;78), no overall significance was found for multivariate effects of \u003cem\u003eDelInfo\u003c/em\u003e and \u003cem\u003eThird-Party Awareness\u003c/em\u003e. Between-Ss effects indicated that \u003cem\u003eThird-Party Awareness\u003c/em\u003e showed a trend for Affordances (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(1, 57)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;3.80, \u003cem\u003ep\u003c/em\u003e = .056, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.06) [ii] and Use Intentions (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(1, 57)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;3.80, \u003cem\u003ep\u003c/em\u003e = .056, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.06) [iii]. Both mean Affordances and mean Use Intentions were higher without Internet connection compared to being Internet connected.\u003c/p\u003e \u003cp\u003eIn Wave 1, for MANOVA without outliers (\u003cem\u003en1\u003c/em\u003e\u0026thinsp;=\u0026thinsp;76), significant effects were found for \u003cem\u003eDelInfo1\u003c/em\u003e (Wilks\u0026rsquo; \u003cem\u003eΛ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.20, \u003cem\u003eF\u003c/em\u003e\u003csub\u003e(5,52)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;2.53, \u003cem\u003ep\u003c/em\u003e = .040, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.20). No significant overall effect was found for \u003cem\u003eThird-Party Awareness\u003c/em\u003e nor any significant interaction effect. Between-Ss effects were not significant.\u003c/p\u003e \u003cp\u003eTo examine which dependent variable contributed to the overall effect, we ran a one-way MANOVA with \u003cem\u003eDelInfo\u003c/em\u003e as the fixed factor. Between-Ss effects were significant for Valence (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(1,63)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;4.06, \u003cem\u003ep\u003c/em\u003e = .048, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.06) [I]. A univariate test was performed and Valence was significantly more positive in the Keep-information group (\u003cem\u003eM\u003c/em\u003e\u0026thinsp;=\u0026thinsp;4.58, \u003cem\u003eSD\u003c/em\u003e = .62) than in the Delete-information group (\u003cem\u003eM\u003c/em\u003e\u0026thinsp;=\u0026thinsp;4.18, \u003cem\u003eSD\u003c/em\u003e = .09).\u003c/p\u003e \u003cp\u003eIn Wave 2 (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, second panel), without outliers, \u003cem\u003en2\u003c/em\u003e\u0026thinsp;=\u0026thinsp;74, multivariate tests did not render significant results. In the Between-Ss test, a trend was visible for \u003cem\u003eThird-Party Awareness\u003c/em\u003e on Relevance (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(1,57)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;4.170, \u003cem\u003ep\u003c/em\u003e = .046, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.07) [iv]. Relevance was higher in the No Internet group compared with Internet-connected.\u003c/p\u003e \u003cp\u003eIn Wave 3 (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, third panel), \u003cem\u003eN3\u003c/em\u003e\u0026thinsp;=\u0026thinsp;73, no significant multivariate effects were found. In the Between-Ss tests, trends were found for \u003cem\u003eThird-Party Awareness\u003c/em\u003e on Ethics (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(1,53)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;4.62, \u003cem\u003ep\u003c/em\u003e = .036, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.08) [v], and for \u003cem\u003eDelInfo\u003c/em\u003e on Use Intentions (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(1,53)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;4.32, \u003cem\u003ep\u003c/em\u003e = .043, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.08) [vi]. Ethics tended to be lower in the Internet than in the No Internet group; Use Intentions to be higher in the Keep group than in the Delete group. Without outliers (\u003cem\u003en3\u003c/em\u003e\u0026thinsp;=\u0026thinsp;71), we found a trend for \u003cem\u003eDelInfo\u003c/em\u003e, indicating that Use Intentions tended to be higher for Keepers compared to Deleters (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(1,51)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;6.92, \u003cem\u003ep\u003c/em\u003e = .011, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.12) [vii].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e4.2.3 MANOVA: Delete vs Keep information vs Breaking Off the session\u003c/h2\u003e \u003cp\u003eWe ran MANOVA on the data of all three waves with three groups in \u003cem\u003eDelInfo\u003c/em\u003e as fixed factor: those who kept their information for later sessions (Keepers), those who deleted their information after a session (Deleters), and those who broke off the session (Breakers) when more private information was at stake.\u003c/p\u003e \u003cp\u003eIn Wave 1, we ran MANOVA with \u003cem\u003eDelInfo\u003c/em\u003e and \u003cem\u003eThird-Party Awareness\u003c/em\u003e as fixed factors. No significant effects were obtained with multivariate tests and the observed trends in the univariate analyses were similar as before in Wave 1.\u003c/p\u003e \u003cp\u003eWe then ran MANOVA for Wave 2 and Pillai\u0026rsquo;s Trace was significant for the overall effect of \u003cem\u003eDelInfo\u003c/em\u003e on the dependents: \u003cem\u003eV\u003c/em\u003e = .27, \u003cem\u003eF\u003c/em\u003e\u003csub\u003e(10,130)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;2.02, \u003cem\u003ep\u003c/em\u003e = .036, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.14. Between-Ss tests showed significant effects of \u003cem\u003eDelInfo\u003c/em\u003e on Ethics (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 68)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;3.76, \u003cem\u003ep\u003c/em\u003e = .028, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.10) [II] and on Use Intentions (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 68)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;5.13, \u003cem\u003ep\u003c/em\u003e = .008, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.13) [III].\u003c/p\u003e \u003cp\u003eScheff\u0026eacute; Post Hoc indicated that mean Ethics was significantly lower for Breakers (\u003cem\u003eM\u003c/em\u003eΔ \u003csub\u003e(Breakers \u0026minus; Deleters)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.68, \u003cem\u003ep\u003c/em\u003e = .026), similar to mean Use Intentions (\u003cem\u003eM\u003c/em\u003eΔ \u003csub\u003e(Breakers \u0026minus; Keepers)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.61, \u003cem\u003ep\u003c/em\u003e = .026) being significantly lower for Breakers than for Keepers.\u003c/p\u003e \u003cp\u003eIn Wave 2, information deleters seemed to be more concerned about ethics while keepers were more focused on use, whereas breakers were neither convinced about the ethics nor the utility of the RBS when they were asked to go into more intimate levels of confession.\u003c/p\u003e \u003cp\u003eIn Wave 3, we ran MANOVA, excluding outliers (\u003cem\u003en3\u003c/em\u003e\u0026thinsp;=\u0026thinsp;71), and obtained a significant overall effect of \u003cem\u003eDelInfo\u003c/em\u003e: Pillai\u0026rsquo;s \u003cem\u003eV\u003c/em\u003e = .40, \u003cem\u003eF\u003c/em\u003e\u003csub\u003e(10, 112)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;2.84, \u003cem\u003ep\u003c/em\u003e = .004, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.20) [IV]. Between-Ss tests showed that \u003cem\u003eDelInfo3_R3\u003c/em\u003e had a significant effect on Ethics (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 59)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;8.59, \u003cem\u003ep\u003c/em\u003e = .001, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.23), on Affordances (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 59)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;4.40, \u003cem\u003ep\u003c/em\u003e = .017, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.13), Relevance (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 59)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;9.18, \u003cem\u003ep\u003c/em\u003e = .000, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.19), Valence (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 59)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;6.78, \u003cem\u003ep\u003c/em\u003e = .002, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.19), and Use Intentions (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 59)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;7.65, \u003cem\u003ep\u003c/em\u003e = .001, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.21). The one-way ANOVA with post-hoc Scheff\u0026eacute; indicated that all five dependents were significantly lower for Breakers compared to Deleters and to Keepers of private information.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePairwise comparisons between \u003cem\u003eDelInfo\u003c/em\u003e and \u003cem\u003eThird-Party Awareness\u003c/em\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMeasure\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eAware\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eJ\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003eM\u003c/em\u003eΔ(I-J)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e\u003cem\u003eM\u003c/em\u003e_Relevance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInternet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreakers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDeleters\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.722\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.032\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eNo Internet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreakers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDeleters\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-1.207\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.002\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreakers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eKeepers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-1.821\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDeleters\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eKeepers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026minus;\u0026thinsp;.615\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.045\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe interaction effect of \u003cem\u003eThird-Party Awareness\u003c/em\u003e \u0026sdot; \u003cem\u003eDelInfo\u003c/em\u003e was significant for Relevance: Wilks\u0026rsquo; \u003cem\u003eΛ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.80, \u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 59)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;3.56, \u003cem\u003ep\u003c/em\u003e = .035, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.11 [V]. In an ANOVA with pairwise comparisons (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), being Internet-connected significantly lowered the mean Relevance of the RBS for people who broke off in the middle of the session as compared to those who merely deleted their information. Without Internet, personal Relevance of the RBS to Breakers was significantly lower than to Deleters and Keepers, while Deleters also thought Relevance was significantly lower to them than to the Keepers of private information.\u003c/p\u003e \u003cp\u003eThose who broke off apparently saw lowest personal relevance because they did not dare to confess more private information. Deleters took the middle ground, because they dared to self-disclose very private information but deleted it afterwards for security\u0026rsquo;s sake. Keepers saw most relevance because they self-disclosed to the deepest level while wanting to maintain their information for the next session, relying on the RBS to safeguard their privacy.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.3. Repeated measures of Third-Party Awareness and DelInfo across three waves\u003c/h2\u003e \u003cp\u003eWe ran Repeated Measures analysis to examine the differences in experience caused by \u003cem\u003eThird-Party Awareness\u003c/em\u003e and \u003cem\u003eDelInfo\u003c/em\u003e across three waves. The specifics of these analyses can be checked in the supplementary materials, technical report, Section 3.4. Results of the Repeated Measures are depicted in the fourth panel of Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eWith \u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;73, there was a significant multivariate interaction effect of \u003cem\u003eWave\u003c/em\u003e \u0026sdot; \u003cem\u003eThird-Party Awareness\u003c/em\u003e (Pillai\u0026rsquo;s \u003cem\u003eV\u003c/em\u003e = .25, \u003cem\u003eF\u003c/em\u003e\u003csub\u003e(10,62)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;2.09, \u003cem\u003ep\u003c/em\u003e = .039, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.25). Univariate tests showed significant effects of \u003cem\u003eWave\u003c/em\u003e on Affordances (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 142)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;4.17, \u003cem\u003ep\u003c/em\u003e = .017, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.06) and of \u003cem\u003eWave\u003c/em\u003e \u0026sdot; \u003cem\u003eThird-Party Awareness\u003c/em\u003e on Affordances (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 142)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;8.31, \u003cem\u003ep\u003c/em\u003e = .000, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.11) [VI].\u003c/p\u003e \u003cp\u003ePairwise comparison of the difference scores with Bonferroni correction indicated that Affordances were assessed as significantly lower in Wave 2 and Wave 3 (\u003cem\u003eM\u003c/em\u003eΔ\u003csub\u003e(Wave2\u0026ndash;Wave3)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.25, \u003cem\u003ep\u003c/em\u003e = .018). Regarding the interaction, Affordances were significantly lower in Wave 1 compared to Wave 3, in the Internet-connected group (\u003cem\u003eM\u003c/em\u003eΔ\u003csub\u003e(Wave1\u0026ndash;Wave3)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.42, \u003cem\u003ep\u003c/em\u003e = .003), but significantly higher in Wave 1 compared to Wave 2 in the No-Internet group (\u003cem\u003eM\u003c/em\u003eΔ\u003csub\u003e(Wave1\u0026ndash;Wave2)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;.44, \u003cem\u003ep\u003c/em\u003e = .002).\u003c/p\u003e \u003cp\u003eWe excluded the outliers and with \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;63, a significant interaction occurred between \u003cem\u003eWave\u003c/em\u003e and \u003cem\u003eThird-Party Awareness\u003c/em\u003e (Pillai\u0026rsquo;s \u003cem\u003eV\u003c/em\u003e = .30, \u003cem\u003eF\u003c/em\u003e\u003csub\u003e(10, 52)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;2.23, \u003cem\u003ep\u003c/em\u003e = .030, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.30). In the Univariate Tests (Greenhouse-Geisser), \u003cem\u003eWave\u003c/em\u003e had a significant effect on Affordances (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 122)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;3.35, \u003cem\u003ep\u003c/em\u003e = .039, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.05) and on Valence (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 122)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;4.23, \u003cem\u003ep\u003c/em\u003e = .019, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.07). Also the interaction between \u003cem\u003eWave\u003c/em\u003e \u0026sdot; \u003cem\u003eThird-Party Awareness\u003c/em\u003e on Affordances was significant: \u003cem\u003eF\u003c/em\u003e\u003csub\u003e(2, 122)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;4.11, \u003cem\u003ep\u003c/em\u003e = .019, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.06.\u003c/p\u003e \u003cp\u003ePairwise comparisons for the differences across timepoints indicated that Valence was significantly lower in Wave 1 compared to Wave 3 (\u003cem\u003eM\u003c/em\u003eΔ\u003csub\u003e(Wave1\u0026ndash;Wave3)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.20, \u003cem\u003ep\u003c/em\u003e = .046), while Affordances tended to be lower in Wave 2 than in Wave 3 (\u003cem\u003eM\u003c/em\u003eΔ\u003csub\u003e(Wave2\u0026ndash;Wave3)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.23, \u003cem\u003ep\u003c/em\u003e = .054).\u003c/p\u003e \u003cp\u003eAgain, in the Internet group, Affordances were rated lower in Wave 1 than in Wave 3 (\u003cem\u003eM\u003c/em\u003eΔ\u003csub\u003e(Wave1\u0026ndash;Wave3)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.38, \u003cem\u003ep\u003c/em\u003e = .017), but were significantly higher in Wave 1 versus Wave 2 in the No Internet group (\u003cem\u003eM\u003c/em\u003eΔ\u003csub\u003e(Wave1\u0026ndash;Wave2)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;.30, \u003cem\u003ep\u003c/em\u003e = .005).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.4. Logistic regression: impact of user experience on information-deletion behaviour\u003c/h2\u003e \u003cp\u003eTo investigate the influence of user experience factors (Ethics, Affordances, etc.) on information-deletion behaviour (\u003cem\u003eDelInfo\u003c/em\u003e), a series of logistic regression analyses were conducted for three waves. Intricacies of these analyses are available in the supplementary materials, technical report, Section 4. Interesting results are exhibited in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eIn Wave 1, with the full data set, the logistic regression approached statistical significance: \u003cem\u003eχ\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e\u003csub\u003e(5)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;12.12, \u003cem\u003ep\u003c/em\u003e = .033, with \u003cem\u003eNagelkerke R\u003c/em\u003e\u003csup\u003e\u003cem\u003e2\u003c/em\u003e\u003c/sup\u003e = .22. Affordances significantly and positively predicted the likelihood of information-deletion behaviour [VII], \u003cem\u003eB\u003c/em\u003e = .82, \u003cem\u003eSE\u003c/em\u003e = .37, \u003cem\u003eWald\u003c/em\u003e\u0026thinsp;=\u0026thinsp;4.91, \u003cem\u003ep\u003c/em\u003e = .027, \u003cem\u003eExp(B)\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2.27, counter-intuitively indicating that the higher the perceived ability of the RBS to provide a secure interaction environment, the higher the likelihood of information-deletion behaviour. Additionally, Valence significantly and negatively predicted the likelihood of information-deletion behaviour [VIII], \u003cem\u003eB\u003c/em\u003e = -1.32, \u003cem\u003eSE\u003c/em\u003e = .53, \u003cem\u003eWald\u003c/em\u003e\u0026thinsp;=\u0026thinsp;6.34, \u003cem\u003ep\u003c/em\u003e = .012, \u003cem\u003eExp(B)\u003c/em\u003e = .27, suggesting that lower perceived Valence was associated with a higher likelihood of deleting information.\u003c/p\u003e \u003cp\u003eIn Wave 2, the logistic regression model was not statistically significant: \u003cem\u003eχ\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e\u003csub\u003e(5)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;8.84, \u003cem\u003ep\u003c/em\u003e = .12, \u003cem\u003eNagelkerke R\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e = .17. None of the independent variables made a statistically significant contribution to the model. Splitting up \u003cem\u003eThird-Party Awareness\u003c/em\u003e into Internet and Not Internet connected groups did not render significant results either.\u003c/p\u003e \u003cp\u003eIn Wave 3, the logistic regression model was not statistically significant: \u003cem\u003eχ\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e\u003csub\u003e(5)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;6.54, \u003cem\u003ep\u003c/em\u003e = .26, \u003cem\u003eNagelkerke R\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e = .14. None of the predictor variables were statistically significant.\u003c/p\u003e \u003cp\u003eHowever, in Wave 3, participants started to differentiate according to being Internet connected. For those with Internet connection, the regression model was statistically significant: \u003cem\u003eχ\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e\u003csub\u003e(5)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;13.34, \u003cem\u003ep\u003c/em\u003e = .020, \u003cem\u003eNagelkerke R\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e = .49, owing to Use Intentions being the only statistically significant predictor: [IX] \u003cem\u003eB\u003c/em\u003e = -2.86, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1.36, \u003cem\u003eWald\u003c/em\u003e\u0026thinsp;=\u0026thinsp;4.42, \u003cem\u003ep\u003c/em\u003e = .036, \u003cem\u003eExp(B)\u003c/em\u003e = .057. Increased Intentions to Use the RBS were associated with a decreased likelihood of information-deletion behaviour. For those without Internet, the model was not statistically significant: \u003cem\u003eχ\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e\u003csub\u003e(5)\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;3.23, \u003cem\u003ep\u003c/em\u003e = .665, \u003cem\u003eNagelkerke R\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e = .15.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Discussion/conclusions","content":"\u003cp\u003eUsers interacted with a robot over a time span of three weeks, either connected to the Internet, users being aware of potential third-party surveillance, or stand-alone, guaranteeing utmost privacy. Next, we relate our results back to the hypotheses.\u003c/p\u003e \u003cp\u003eH1 (Ethics, Golda et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Belfrage et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Shin et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) was supported. While awareness of surveillance did not affect ethical perceptions in week 1, by week 3, participants using the offline (stand-alone) RBS rated it significantly higher on ethics than those in the Internet-connected group.\u003c/p\u003e \u003cp\u003eH2 (Affordances, Jabbar et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Jaspers \u0026amp; Pearson, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) received mixed support. Initially, in week 1, the offline RBS was perceived to have higher affordances (functionality/security). However, over time, week 3, this reversed; participants in the Internet-connected group rated affordances higher, likely due to the perceived limitations of a stand-alone system.\u003c/p\u003e \u003cp\u003eH3 (Valence, Golda et al., Shin et al.) was supported. Higher perceived valence (positive outcome expectancy) was significantly associated with Keepers (those who did not delete data), while those with negative expectations were more likely to delete their information.\u003c/p\u003e \u003cp\u003eH4 (Relevance, Jaspers \u0026amp; Pearson; Shin et al.): supported. In week 2, relevance was significantly higher for the offline group. By week 3, a significant interaction showed that Keepers found the stand-alone RBS much more personally relevant to their goals than those who broke off interactions.\u003c/p\u003e \u003cp\u003eH5 (Use Intentions, Jabbar et al., Shin et al.): supported. Across the experiment, intentions to use the robot were higher in the offline condition. Path analyses confirmed that Use Intentions were driven by both perceived Ethics and Affordances.\u003c/p\u003e \u003cp\u003eH6 (Willingness to Disclose, Jabbar et al.): supported. Participants who felt secure enough to Keep their data reported higher Use Intentions and more positive evaluations, whereas those uncomfortable with the level of disclosure (Breakers) consistently rated the system lower.\u003c/p\u003e \u003cp\u003eH7 (Tendency to Delete, Shin et al.; Golda et al.) was confronted with a counter-intuitive result. In week 1, higher perceived affordances (security) actually predicted higher likelihood of deletion. However, by week 3, as hypothesized, higher Use Intentions were associated with a decreased likelihood of deleting information.\u003c/p\u003e \u003cp\u003eH8 (Improvement Over Time, Jaspers \u0026amp; Pearson; Belfrage et al.): partially supported. Trust and familiarity evolved; initial awareness of surveillance faded and was internalised as a sustained ethical concern. While some differences diminished, the offline RBS maintained a stronger sense of security for long-term users.\u003c/p\u003e \u003cp\u003eWhat do these results tell us? We tracked participants\u0026rsquo; interactions with the RBS over three weeks, revealing a shift from initial caution to a more pragmatic, utility-driven mindset. In week 1, first wave, the initial encounter held a kind of safety paradox. In the first wave, participants\u0026rsquo; reactions seem to have been influenced by external factors, specifically whether they were connected to the Internet (\u003cem\u003eThird-Party Awareness\u003c/em\u003e). Those connected to the Internet, it seems, were more sceptical of the system\u0026rsquo;s technical capabilities and security (cf. Affordances).\u003c/p\u003e \u003cp\u003eDuring this initial phase, the participants divided into two main behavioural groups: those who wanted to keep their personal data for later use and those who deleted it. The Keepers reported a more positive emotional response (Valence) to the system. Because they felt good about the interaction, they retained their data.\u003c/p\u003e \u003cp\u003eThe Deleters, maybe, were responsible for the counter-intuitive effect: one would expect that if users trust the RBS\u0026rsquo; security (i.e., Affordances), they would feel safe enough to keep their data stored within it. However, the logistic regression showed the opposite: higher perceived Affordances led to a higher likelihood of deleting data. There seems to have been a safety-first effect. In week 1, users who rated the system highly for providing a secure interaction environment likely viewed the functionality to delete data as a core component of that security. They perhaps did not delete the data because they thought the system was broken or unsafe; rather, they used the system\u0026rsquo;s security features (the delete button) precisely because they were engaging with the system\u0026rsquo;s affordances to manage their privacy. These users were actively managing their security through the system\u0026rsquo;s features, rather than passively trusting it.\u003c/p\u003e \u003cp\u003eIn week 2, second wave, distinct user types started to emerge. The external factor of an Internet connection ceased to be the primary driver of the participants\u0026rsquo; views. The users\u0026rsquo; internal evaluation of the system (Ethics and Affordances) became the dominant predictor of whether they found the system relevant and fit for future use.\u003c/p\u003e \u003cp\u003eAdditionally, a third group of users became statistically distinct in this wave, which we baptised Breakers: those who abandoned the session when asked for deep self-disclosure. These individuals began to disengage, rating the system significantly lower on Ethics and Use Intentions. They were unconvinced by the system\u0026rsquo;s moral standing or its utility. With that, a clear divergence in motivation appeared. Deleters remained concerned with Ethics (privacy and trust), whereas Keepers were focused on Utility (usage and function).\u003c/p\u003e \u003cp\u003eWeek 3, the third wave, was one of polarisation and pragmatism. By the final wave, the separation between the groups was clear. The Breakers effectively \u0026lsquo;checked out.\u0026rsquo; They rated the RBS lowest on almost every metric: Ethics, Affordances, Valence, Relevance, and Use Intentions. For them, the presence of an Internet connection may have acted as a final deterrent, significantly lowering the system\u0026rsquo;s perceived Relevance. For those who stayed (Keepers and Deleters), behaviour became more pragmatic. The logistic regression showed that for Internet-connected users, the only significant predictor of behaviour was Use Intentions. If they intended to use the system again, they kept the data. If not, they deleted it.\u003c/p\u003e \u003cp\u003eIn all, over the course of three weeks, our results suggest that participants moved from cautious exploration to pragmatic decision-making. Initially, in the first week, users were hyper-aware of surveillance (i.e., the Internet connection). Those who recognised the system\u0026rsquo;s security features used them to delete data, exercising control over their privacy.\u003c/p\u003e \u003cp\u003eOver time, second and third week, we suspect that the external fear of the Internet faded. Users who trusted the system\u0026rsquo;s utility became Keepers, while those who doubted its ethics and/or relevance became Deleters or even Breakers.\u003c/p\u003e \u003cp\u003eUltimately, the counter-intuitive deletion behaviour of the first week disappeared. Two weeks later, keeping data was no longer about security perception, but simply about whether the user found the system useful enough to warrant a return visit.\u003c/p\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e5.1. Limitations\u003c/h2\u003e \u003cp\u003eWhile our study offers longer-term insights into trust in robotic confidantes, several limitations apply. Firstly, the use of a Wizard of Oz methodology means participants interacted with a simulated and fully controlled AI rather than an autonomous system; algorithmic latency or errors might alter trust dynamics in real-world deployments.\u003c/p\u003e \u003cp\u003eSecondly, the sample consisted primarily of young adult university students in Hong Kong, which may limit the generalisability of the findings to older populations or different cultural contexts where privacy norms vary. Dependent on age, users\u0026rsquo; confidence in their own interpretation judgment (trust-in-self) may affect AI and data usage. Older adults often seek confirmation even when information comes from an authoritative source, because they worry, they may have misunderstood. Younger people seem less concerned about loss of privacy.\u003c/p\u003e \u003cp\u003eThirdly, the measurement of Relevance relied on a single-item scale, which, despite showing correlation across waves, lacks the robust psychometric validation of the other multi-item scales. Finally, the three-week duration, while longer than just one session, may not capture the exhaustion of the \u0026lsquo;novelty effect\u0026rsquo; or the long-term trust fluctuations that occur over months of, for instance, therapeutic use.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e5.2. Practical application\u003c/h2\u003e \u003cp\u003e Our findings provide a guideline for the design and deployment of AI-driven mental-health technologies. It would be wise to prioritise local processing (i.e., edge computing). To foster initial trust and ethical security, developers should favour stand-alone or edge processing over Cloud-based systems. A visible, offline data vault that they can bring home may reduce user anxiety during sensitive disclosures (Hoorn, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). For those, however, who want to access the Internet irrespective, security measures should be upgraded (ibid.).\u003c/p\u003e \u003cp\u003eIt would be important to implement transparent data control. The safety-first effect observed in the first week suggests that providing prominent data-delete functions actually enhances the perception of a secure environment. Users should be given granular control over their data to encourage early-stage engagement.\u003c/p\u003e \u003cp\u003eWe should design for utility to sustain long-term use. As the initial fear of surveillance faded, use intentions appeared to be driven by perceived utility (affordances). For long-term retention, the system must transition from being merely secure to being demonstrably useful in helping the user achieve their goals, in our case, mental support.\u003c/p\u003e \u003cp\u003eDevelopers must be cautious when the AI solicits highly intimate secrets. Since Breakers (those who abandoned sessions) reported lower trust across all dimensions, therapeutic systems should use incremental disclosure protocols that allow users to build rapport before reaching deeper levels of vulnerability. For social robots to be safe confidants, we should learn how to manage the deep-disclosure threshold.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAcknowledgments\u003c/h2\u003e \u003cp\u003e This research is funded by the Laboratory for Artificial Intelligence in Design (project code: R2P3), Innovation and Technology Fund, Hong Kong Special Administrative Region Government as well as by the Research Grants Council (project code: T43-518/24-N) under the University Grants Committee, Hong Kong Special Administrative Region Government. We are grateful to Stoney Y. Wang and Yoyo W. Y. Cheung for data acquisition.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAllen MS, Iliescu D, Greiff S (2022) Single item measures in psychological science. Eur J Psychol Assess 38(1):1\u0026ndash;5. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1027/1015-5759/a000699\u003c/span\u003e\u003cspan address=\"10.1027/1015-5759/a000699\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBelfrage S, Helgesson G, Lyn\u0026oslash;e N (2022) Trust and digital privacy in healthcare: a cross-sectional descriptive study of trust and attitudes towards uses of electronic health data among the general public in Sweden. BMC Med Ethics 23(1):19. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12910-022-00758-z\u003c/span\u003e\u003cspan address=\"10.1186/s12910-022-00758-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChoudhury A, Shamszare H (2023) Investigating the impact of user trust on the adoption and use of ChatGPT: survey analysis. J Med Internet Res 25:e47184. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2196/47184\u003c/span\u003e\u003cspan address=\"10.2196/47184\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuan EY, Yoon JM, Liang EZ, Hoorn JF (2021) Self-disclosure to a robot: Only for those who suffer the most. Robotics 10(3):98. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/robotics10030098\u003c/span\u003e\u003cspan address=\"10.3390/robotics10030098\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGolda A, Mekonen K, Pandey A, Singh A, Hassija V, Chamola V, Sikdar B (2024) Privacy and security concerns in generative AI: a comprehensive survey. IEEE Access 12:48126\u0026ndash;48144. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ACCESS.2024.3381611\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2024.3381611\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGreen EC, Murphy EM, Gryboski K (2020) The health belief model. In: Sweeny K, Robbins ML, Cohen LM (eds) The Wiley encyclopedia of health psychology. Wiley, New York. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/9781119057840.ch68\u003c/span\u003e\u003cspan address=\"10.1002/9781119057840.ch68\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHoorn JF (2018) The Robot Brain Server, design of a human-artificial systems partnership. In W. Karwowski, \u0026amp; T. Ahram (Eds.), \u003cem\u003eIntelligent Human Systems Integration. Advances in Intelligent Systems and Computing, 722\u003c/em\u003e (pp. 531\u0026ndash;536). Berlin, Heidelberg: Springer. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/978-3-319-73888-8_83\u003c/span\u003e\u003cspan address=\"10.1007/978-3-319-73888-8_83\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJabbar A, Geebren A, Hussain Z, Dani S, Ul-Durar S (2023) Investigating individual privacy within CBDC: A privacy calculus perspective. Res Int Bus Finance 64:101826. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ribaf.2022.101826\u003c/span\u003e\u003cspan address=\"10.1016/j.ribaf.2022.101826\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJaspers ED, Pearson E (2022) Consumers\u0026rsquo; acceptance of domestic Internet-of-Things: The role of trust and privacy concerns. J Bus Res 142:255\u0026ndash;265. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jbusres.2021.12.043\u003c/span\u003e\u003cspan address=\"10.1016/j.jbusres.2021.12.043\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee JER, Nass CI (2010) Trust in computers: The computers-are-social-actors (CASA) paradigm and trustworthiness perception in human-computer communication. In: Latusek D, Gerbasi A (eds) Trust and technology in a ubiquitous modern environment: Theoretical and methodological perspectives. IGI Global Scientific, Hershey, PA, pp 1\u0026ndash;15\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMacDorman KF (2025) Chikamatsu, Mori, and the uncanny valley. i-Perception 16(1):20416695251317469. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/2041669525131\u003c/span\u003e\u003cspan address=\"10.1177/2041669525131\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShin D, Kee KF, Shin EY (2022) Algorithm awareness: Why user awareness is critical for personal privacy in the adoption of algorithmic platforms? Int J Inf Manag 65:102494. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ijinfomgt.2022.102494\u003c/span\u003e\u003cspan address=\"10.1016/j.ijinfomgt.2022.102494\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTabachnick BG, Fidell LS, Ullman JB (2013) Using multivariate statistics, vol 6. pearson, Boston, MA, pp 497\u0026ndash;516\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Vugt HC, Hoorn JF, Konijn EA (2009) Interactive engagement with embodied agents: An empirically validated framework. Comput Animat Virtual Worlds 20:195\u0026ndash;204. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/cav.312\u003c/span\u003e\u003cspan address=\"10.1002/cav.312\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVenkatesh V, Morris MG, Davis GB, Davis FD (2003) User acceptance of information technology: Toward a unified view. MIS Q 27(3):425\u0026ndash;478. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2307/30036540\u003c/span\u003e\u003cspan address=\"10.2307/30036540\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWold S, Esbensen K, Geladi P (1987) Principal component analysis. Chemometr Intell Lab Syst 2(1\u0026ndash;3):37\u0026ndash;52\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Footnotes","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.questionpro.com/\u003c/span\u003e\u003cspan address=\"https://www.questionpro.com/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Hong Kong University Grants Council","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-9534725/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9534725/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eWe investigated user trust and self-disclosure behaviours in the context of empathetic robotic companions supported by a Robot Brain Server (RBS), designed to address mental health care gaps in Hong Kong. We conducted a three-wave repeated-measures experiment with 86 young adult participants, comparing interactions with a Nao robot connected either to an Internet-enabled RBS (potentially subject to third-party surveillance) or to a stand-alone, offline RBS. Across three sessions, we measured user perceptions of the RBS in terms of ethics, affordances, valence, relevance, and use intentions, as well as their willingness to disclose personal information and subsequent data deletion behaviour. Multivariate analyses revealed that awareness of third-party surveillance negatively impacted perceived affordances and use intentions during initial interactions, while the offline RBS fostered a greater sense of security and relevance to personal goals. Over time, these differences diminished, suggesting that repeated exposure may mitigate initial privacy concerns. Notably, participants who broke off interactions at deeper levels of self-disclosure consistently rated the RBS lower on all experiential dimensions. Logistic regression indicated that both negative expectations and higher perceived affordances predicted information deletion. Path analyses showed that initial awareness of surveillance became a sustained ethical concern, which, alongside protective functionality, drove intensions to use the RBS via positive or negative outcome expectations. These findings underscore the importance of transparent data management and system design in fostering user trust and engagement with digital mental health technologies. Implications for the deployment of AI-driven caregivers in sensitive domains are discussed.\u003c/p\u003e","manuscriptTitle":"Trust in the Robot Brain Server","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-28 05:41:05","doi":"10.21203/rs.3.rs-9534725/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"bc1af670-c0db-4c1b-b8e5-9b488e58ddb2","owner":[],"postedDate":"April 28th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-28T05:41:10+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-28 05:41:05","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9534725","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9534725","identity":"rs-9534725","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}