LLM-Enabled Humanoid Interaction in Care Settings: A Two-Phase Study with Ameca

LLM-Enabled Humanoid Interaction in Care Settings: A Two-Phase Study with Ameca | 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 Article LLM-Enabled Humanoid Interaction in Care Settings: A Two-Phase Study with Ameca Charlotte Larmuseau, Mattice Rigole, Jens Krijgsman This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9212133/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 Loneliness among older adults is a growing public health concern, particularly in residential care settings where opportunities for sustained one-on-one interaction are limited. Recent advances in large language models (LLMs) enable socially assistive humanoid robots to engage in more natural dialogue, potentially supporting interaction-level experiences related to social presence and engagement, which are conceptually relevant to loneliness. However, evidence from real-world care deployments remains limited. This study evaluated the feasibility and user experience (UX) of interactions with Ameca (Engineered Arts, UK) in a Flemish residential-care setting and examined how baseline psychosocial factors relate to perceived interaction quality. We conducted a two-session, mixed-methods feasibility study. Thirty-one older adults (63–93 years) completed a 15–20 minute semi-structured Dutch conversation with Ameca (Session 1); twenty participants returned approximately one week later for Session 2 after a prompt-level re-finement of the conversational configuration. Pre-interaction measures assessed loneliness and well-being. Post-interaction measures captured user experience (UX): perceived enjoyment, empathy, social comfort, trust, and comprehensibility complemented by open-ended feedback. Conversa-tion transcripts were available for Session 2 and were analyzed for turn-taking balance, response patterns, and thematic content. UX ratings were generally positive and stable across sessions. Higher baseline loneliness was associated with less favorable UX evaluations, most notably reduced comprehensibility. Session 2 transcript analyses indicated balanced turn-taking but predominantly question–answer interaction patterns, with conversational depth varying across participants and themes centering on everyday life. Overall, findings support the feasibility and acceptability of LLM-enabled humanoid inter-action in residential care and highlight design priorities for future systems, including improved pacing, personalization, and context handling to better support social presence and engagement. Health sciences/Health care Biological sciences/Psychology Social science/Psychology humanoid robot large language models conversational AI user experience feasibility study older-adult care Figures Figure 1 Figure 2 Figure 3 1 Introduction Loneliness among older adults is a growing public health concern, particularly in residential care settings where reduced mobility, loss of social networks, and limited family contact can contribute to social isolation (Courtin & Knapp, 2017 ). Persistent loneliness has been linked to adverse health outcomes, including cognitive decline, depression, and increased mortality risk (Cacioppo & Cacioppo, 2014 ). Simultaneously, healthcare systems and long-term care facilities face persistent staff shortages, limiting the time available for meaningful one-on-one conversations and emotional support. Advances in social robotics offer new opportunities to address these challenges. Humanoid robots, such as Ameca , are capable of naturalistic speech, facial expressions, and responsive interaction, making them a potential tool to supplement human contact and provide moments of social engage-ment (Broadbent, 2016). The present work evaluates an LLM-enabled humanoid interaction frame-work deployed in a real residential-care context, focusing on feasibility and UX outcomes and how these relate to baseline psychosocial differences. Socially assistive robots may influence loneliness-related experiences through proximal interaction mechanisms rather than direct clinical effects. First, humanoid embodiment can increase perceived so-cial presence —the sense of “being with” a responsive other—through gaze behavior, facial expressive-ness, and responsive backchannels. Second, contingent responsiveness (timely, contextually appropriate replies) supports reciprocity and can reduce the experience of one-sided communication. Third, com-panionship cues (affiliative language, warmth, humor, and empathic acknowledgments) can increase approachability and momentary feelings of companionship. Fourth, structured prompts and remi-niscence topics can enhance engagement , encouraging conversational participation and self-disclosure. In this feasibility study, these mechanisms are operationalized through UX ratings (enjoyment, per-ceived empathy, social comfort, trust, and clarity) and conversational interaction metrics as proximal indicators that may precede longer-term psychosocial change. 2 Theoretical Framework 2.1 Loneliness in Residential Care as a Socio-Technical Challenge: Impli-cations for Scalable Care Technologies Loneliness in later life refers to a perceived mismatch between the quantity or quality of social re-lationships people have and those they desire, reflecting not merely social contact but unmet social needs (De Jong-Gierveld & Van Tilburg, 1999 ). In 2024, global life expectancy has risen to 73.3 years, and the population aged 60 and over is expected to grow from 1.1 billion in 2023 to 1.4 billion by 2030 (World Health Organization, 2025 ). Importantly, older adults living in residential care and long-term care facilities experience particularly high levels of loneliness compared with community-dwelling peers, with systematic reviews estimating that approximately 61% of residents experience moderate loneliness and around 35% severe loneliness in care homes (Gardiner, Laud, Heaton, & Gott, 2020 ). Loneliness in congregate care settings is associated with adverse mental and physical health outcomes, emphasizing the unique social and structural challenges faced within institutional care con-texts (Lapane et al., 2022 ; Donovan & Blazer, 2020 ). Within residential care environments, these risks are further amplified by practical constraints: staff shortages, time pressure, and increasing care demands often limit opportunities for extended one-on-one conversations and emotionally meaning-ful interactions (Boamah, Weldrick, Lee, & Taylor, 2021 ; Carson, Johansson, Schaumberg, & Hurtig, 2024 ). While traditional interventions such as animal-assisted therapy or organized group activities can temporarily reduce loneliness, they are typically resource-intensive, difficult to scale, and challenging to sustain over time (Banks, Willoughby, & Banks, 2008 ; Yen, Huang, Chiu, & Jin, 2024 ). Social isolation has diverse consequences. People with limited social contact show higher rates of physical health conditions (e.g., hypertension, arthritis, and heart disease) and increased risk of cognitive decline and dementia, alongside reduced emotional resilience (Lal, 2025 ). Lower social contact can delay diagnosis of health problems and is associated with increased mortality risk (Smith et al., 2021 ). Furthermore, social disconnection is linked to poorer perceived health and increased symptoms of depression and anxiety (Santini et al., 2020; Tuneu et al., 2023 ). Given these wide-ranging effects, there is a growing need for innovative, scalable approaches that can complement human care and support social and emotional well-being in later life. 2.2 Large Language Models, care applications, and social interaction Large Language Models (LLMs) are software systems designed to generate text based on prompts, pro-ducing contextually appropriate responses through probabilistic next-word prediction (Campbell IV, Chick, Shin, & Makary, 2024). In mental health and older-adult care, LLM-based systems are increas-ingly explored for conversational support, including psycho-educational explanations, motivational coaching, guided reminiscence, companionship-style dialogue, and practical assistance (e.g., reminders or informational support), typically under safety constraints that prohibit medical diagnosis or sensi-tive content. In these settings, the primary promise is not to replace caregivers, but to provide scalable conversational availability and adaptive dialogue that may support engagement and perceived social presence when human time for extended one-on-one interaction is limited. Related feasibility work on LLM-augmented social robots further illustrates this development. Van ’t Klooster et al. ( 2026 ) report on the design and pilot evaluation of a GPT-reinforced social robot intended to support patient–provider communication in clinical practice. The robot was connected to whitelisted, validated clinical sources and engaged users in natural dialogue to convey disease-and treatment-related information. User experience (UX) was assessed using the UX Questionnaire (UEQ) and semi-structured interviews in both a laboratory pilot and a hospital deployment with osteoarthritis patients and healthcare professionals. While the study focused on informational and clinical communication rather than emotionally oriented dialogue, it demonstrated the feasibility of LLM-enhanced social robots in applied healthcare contexts and highlighted the importance of inter-action quality, clarity, and pacing for user acceptance (van ’t Klooster et al., 2026 ). Building on such feasibility work, Pinto-Bernal, Biondina, and Belpaeme ( 2025 ) specifically examine the integration of LLMs into a physically embodied social robot Pepper with the aim of supporting engaging, socially oriented interaction. Their work describes a user-centered LLM architecture that combines real-time speech processing, persona conditioning, adaptive turn-taking, and lightweight memory mechanisms to sustain conversational coherence over time. Through iterative evaluations, including exploratory interactions with older adults, the authors show that features such as memory continuity and adap-tive dialogue pacing contribute to higher perceived trust, naturalness, and social presence, while also noting persistent challenges related to latency and robustness. This study underscores the importance of interaction design and UX evaluation when deploying LLM-enabled humanoid robots for socially meaningful conversations with older adults (Pinto-Bernal, Biondina, & Belpaeme, 2025 ). Within social robotics research more broadly, LLMs have been studied for enabling adaptive turn-taking, culturally attuned conversational style, and more natural dialogue flow, which can shape user perceptions of trust, naturalness, and social presence (Pinto-Bernal et al., 2025 ; Verhelst & Belpaeme, 2025 ). However, the effectiveness and acceptability of such systems depend not only on language generation capabilities, but also on how users respond to embodied agents and how interaction design supports comfortable pacing, clarity, and rapport in real-world care settings. 2.3 Robot–human interaction and social presence in elderly care While LLMs enable increasingly natural and adaptive dialogue, the social and emotional impact of such systems ultimately depends on how older adults perceive and experience embodied robotic agents. Research on human–robot interaction demonstrates that robots can elicit social and affective responses that, in some cases, resemble those triggered by living beings. A frequently cited comparison is animal-assisted therapy, which has been shown to reduce social isolation among older adults. Studies comparing animal-assisted interventions with robotic alternatives have found no significant differences in loneliness outcomes, suggesting that social bonding and emotional engagement can also emerge in interactions with artificial agents (Banks et al., 2008 ). Empirical work in long-term care settings further indicates that interactions with socially assistive robots may be associated with reduced loneliness and depressive symptoms for some older adults, although findings remain heterogeneous and context-dependent (Yen et al., 2024 ). These mixed out-comes highlight that conversational capability alone is insufficient; the social meaning of interaction is shaped by embodiment, interaction style, and user expectations. Earlier generations of socially assistive robots (SARs), such as PARO and ElliQ , illustrate both the promise and limitations of robotic companionship. While many users report positive affect, curiosity, and engagement, recurring challenges include high costs, technical reliability, and a perceived lack of conversational depth or authenticity (Hung et al., 2019 ; Broadbent et al., 2024; Coghlan & Waycott, 2022 ). These findings underscore that acceptance is influenced not only by what a robot can say, but also by how it looks, behaves, and responds over time. Appearance and embodied behavior play a critical role in shaping user responses. Human-oriented designs tend to elicit stronger feelings of social presence and sociability, whereas product-oriented designs are often evaluated as more efficient or predictable (Kwak, 2014 ). Psychological theories provide additional insight: the mere-exposure effect suggests that repeated interaction can increase comfort and familiarity, while the uncanny-valley hypothesis posits that near-human but imperfect realism may provoke discomfort or unease (Fiolka, Donnermann, & Lugrin, 2024 ; Mara, Appel, & Gnambs, 2022 ). Taken together, these findings suggest that the evaluation of LLM-enabled humanoid robots in elderly care must consider not only conversational intelligence, but also embodiment, familiarity, and perceived social presence. For expressive humanoids such as Ameca , conversational comfort, trust, and social acceptance emerge from the interplay between advanced dialogue capabilities and human-centered interaction design. 2.4 Synthesis and research questions Across residential-care contexts, persistent loneliness among older adults coexists with increasing work-load pressures and limited opportunities for sustained one-on-one interaction. While traditional psy-chosocial interventions can alleviate loneliness, they are often difficult to maintain at scale within institutional care environments. Recent advances in LLM-enabled social robotics introduce opportu-nities for more natural, adaptive dialogue, yet evidence remains limited on how such conversationally capable humanoid robots are experienced by older adults in real-world care settings. Against this backdrop, the present study examines interactions with Ameca , an expressive humanoid robot equipped with LLM-based conversational capabilities, deployed in a Flemish residential-care context. By combining standardized measures, UX ratings, conversational metrics, and qualitative feedback, the study aims to clarify how older adults experience short, structured conversations with a humanoid robot and how baseline psychosocial characteristics relate to perceived interaction quality. The study addresses the following research questions: RQ1 How do older adults experience short, structured conversations with the humanoid robot Ameca in terms of enjoyment, empathy, social comfort, trust, comprehensibility, and perceived appear-ance? RQ2 Among participants who completed both sessions ( N = 20), how do UX ratings differ between Session 1 (baseline) and Session 2 (optimized), and what is the magnitude and uncertainty of these differences? RQ3 What conversational characteristics (turn-taking balance, question–answer patterns, response length, and themes) are observed under the optimized configuration based on Phase 2 transcripts? RQ4 How do older adults qualitatively describe their interaction with Ameca , with particular attention to perceived conversational naturalness and embodied presence? 3 Methodological Framework 3.1 Ameca Humanoid Robot Platform As shown in Fig. 1 , Ameca is a humanoid robot developed by Engineered Arts Ltd. (UK), designed as a research platform for human-robot interaction. The system features a life-sized torso with 17 facial actuators, allowing for expressive gestures and natural gaze behaviour. Its embedded microphones, stereo speakers, and high-resolution RGB cameras enable real-time audiovisual sensing and speech processing. In this study, Ameca operated in conversational mode via the Tritium AI interface, in-tegrating an LLM-based dialogue model (GPT-4o), automatic speech recognition (Whisper large-v3), and neural text-to-speech synthesis. All inference ran through a secured cloud connection with sub-second response latency (typically 0.6-1.0 s end-to-end). Safety filters and topic restrictions prevented sensitive, medical, or personally identifying content, and no audio data were stored beyond temporary ASR buffering. The robot’s dialogue manager was configured to maintain a friendly, supportive tone, short sentence structure, and clear articulation tailored to older users. 3.2 Participants A total of 31 older adults participated in the study. The mean age was 83.2 years ( SD = 8.3), with 23 women and 8 men. Most participants resided in a group assisted living facility ( N = 19), while others lived in a care hotel ( N = 8), at home in the neighborhood ( N = 2), or in a nearby apartment ( N = 2). These characteristics reflect a relatively diverse sample of older adults living in both residential and community-based contexts. 3.3 Procedure and Measures Before the interaction, an informed consent and demographic information (age, gender, and living situation) was collected, followed by two standardized pre-interaction assessments. Loneliness was measured using the Three-Item UCLA Loneliness Scale (UCLA-3; (Hughes, Waite, Hawkley, & Ca-cioppo, 2004)), employing a four-point scale (1 = Never to 4 = Always), yielding a total score range of 3–12. Well-being was assessed using the WHO-5 Well-Being Index (Topp, Østergaard, Søndergaard, & Bech, 2015 ), consisting of five items rated on a six-point scale (0 = At no time to 5 = All of the time), converted to a percentage score (0-100) with higher values reflecting greater well-being. Following the pre-surveys, participants engaged in a 15–20 minute semi-structured conversation with Ameca . The robot was configured to speak Dutch and follow a consistent script emphasizing warmth, active listening, and clarity, covering everyday topics such as family, food, hobbies, and reminiscence. A facilitator introduced the robot using a standardized script and only intervened to ensure comfort or address technical issues. After each interaction, all participants completed a bespoke Ameca UX Questionnaire assessing five interaction domains—empathy, trust, social comfort, comprehensibility, and enjoyment—using five-point Likert scales (1 = Strongly disagree to 5 = Strongly agree), followed by several open-ended questions on perceived conversational naturalness and Ameca ’s physical appearance. Each session lasted approximately 30–35 minutes in total, including surveys, interaction, and debriefing. The UX questionnaire was developed for this study to support feasibility assessment and exploratory user experience analysis. The domains of Empathy, Trust, and Social Comfort were theoretically grounded in established human–robot interaction constructs, including social presence, perceived sup-port, and trust in automation (Hancock et al., 2011 ; Leite et al., 2014 ; Spatola, Belletier, Chausse, & Augustinova, 2021 ). Comprehensibility was informed by prior work on conversational usability and user experience in chatbot-based systems (Følstad & Brandtzæg, 2017 ). 3.4 Study Design This study followed a two-session, mixed design conducted over a two-week period. In Phase 1 , all participants ( N = 31) completed a baseline interaction with Ameca , See Fig. 3 Of these participants, 20 returned approximately one week later to complete Phase 2 , in which the same interaction protocol was administered after conversational refinements were implemented (see Section 3.4). Participants who completed Phase 2 did not differ substantially from the full Phase 1 sample in terms of age, gender distribution, or living situation and were therefore considered demographically comparable. Their mean age was 82.3 years ( SD = 8.3); 15 were female and 5 male. Most participants resided in a group assisted living facility ( N = 15), while others lived in a care hotel ( N = 3) or at home ( N = 2) or apartment ( N = 1) in the neighborhood. Although the initial aim was a full within-subjects design, practical constraints (limited testing slots, overlapping residential activities, and health-related drop-out) resulted in a partial repeated-measures design, with only a subset of participants ( N = 20) completing both phases. 3.5 Intervention In Phase 1 (baseline) , Ameca operated using the manufacturer’s default conversational prompt-ing configuration, without study-specific optimization. This configuration relied primarily on short, question-oriented exchanges and provided limited support for reflective listening, topic continuity, or socio-emotional signaling. Following these initial sessions, qualitative observations and participant feedback highlighted op-portunities to improve the robot’s listening behavior, turn-taking, topic continuity, and socio-emotional expressiveness. In Phase 2 (approximately one week later), no changes were made to Ameca ’s hardware, embod-iment, sensing modalities, speech recognition pipeline, or underlying language model. Instead, the intervention was refined exclusively at the level of the LLM-based conversational framework through systematic prompt engineering within Ameca’s native conversational platform. These refinements included: (a) adjusted turn-taking delays to reduce interruptions and allow nat-ural pacing; (b) increased use of reflective listening strategies, such as paraphrasing and empathic acknowledgments; (c) improved topic continuity and conversational closure when exchanges stalled; and (d) enhanced socio-emotional backchannels, including light humor and small compliments to estab-lish rapport. In addition, the conversational prompts gently encouraged familiar, reminiscence-based themes (e.g., family, hobbies, food, and everyday life) that are commonly accessible to older adults. A custom conversational persona was thus implemented, emphasizing a warm, attentive, and sup-portive interaction style suited to conversations with older adults. The conversational framework remained prompt-based rather than script-based, allowing generative flexibility while constraining in-teraction style toward more human-centered dialogue. No therapeutic content, coaching elements, or additional intervention components were introduced between phases. Together, these refinements aimed to enhance perceived empathy, social comfort, and conversational flow while maintaining consistency across sessions. 3.6 Data Analysis Quantitative analysis. Descriptive statistics (means, standard deviations, ranges) were computed for baseline psychosocial measures (UCLA-3, WHO-5) and for all UX questionnaire domains. Internal consistency was evaluated using Cronbach’s alpha (Cronbach, 1951 ). Reliability coefficients at the domain level ranged from acceptable to good for most constructs. In contrast, the Comprehensibility and Clarity domain demonstrated comparatively lower internal consistency (see Table 1 , N = 31), suggesting possible multidimensionality or limited item coherence. Consequently, findings associated with this domain should be interpreted with caution. For participants who completed both sessions ( N = 20), paired-sample t -tests were conducted to descriptively compare UX ratings between Phase 1 and Phase 2. Given the limited sample size, the sequential two-phase design, and the absence of correction for multiple comparisons, inferential claims were not emphasized. Effect sizes (Cohen’s d ) were reported to provide an indication of the magnitude of observed differences, which were interpreted descriptively rather than causally. Table 1 Questionnaire constructs, items, internal consistency, and descriptive statistics (1–5 Likert scale; Phase 1 baseline) Construct Item Item text (English) α M SD A. Overall Con- A1 I found the conversation with Ameca pleas- 3.7 0.7 versational En- ant. joyment A2 The conversation felt natural. A3 I felt good during the conversation with 0.79 Ameca. B. Empathy and B1 Ameca showed understanding for my situ- 3.8 0.6 Understanding ation. B2 I felt taken seriously by Ameca. B3 Ameca showed interest in what I said. B4 I felt that Ameca was really listening. 0.64 C. Social Com- C1 I felt comfortable during the conversation 3.6 0.7 fortandCom- with Ameca. panionship C2 I found Ameca pleasant company. C3 I would like to talk to Ameca more often. 0.70 D.Trustand D1 I trusted Ameca during the conversation. 3.8 0.6 Credibility D2 Ameca seemed honest and reliable. D3 I could be myself during the conversation. D4 I felt that Ameca understood me well. 0.83 E. Comprehensi- E1 Ameca’s answers were clear and under- 3.9 0.6 bility and Clarity standable. E2 Ameca used language that was easy for me to follow. E3 I knew when it was my turn to speak. E4 I knew when Ameca was speaking. 0.51 Conversational metrics. Due to technical constraints, conversational transcripts were available only for Phase 2. These transcripts were analyzed to characterize interaction patterns under the optimized conversational configuration. Using a structured coding protocol implemented in Excel , conversational metrics were extracted per participant, including total utterances, turn distribution between Ameca and the participant, number of questions and responses, mean response length, and proportion of short responses. All transcripts were reviewed independently by two researchers to verify consistency of segmentation and coding. Qualitative analysis. Qualitative data consisted of participants’ written responses to post-interaction open-ended questions addressing (1) their overall experience of the conversation with Ameca , (2) as-pects they liked most, (3) aspects they liked least, (4) suggested improvements for future interactions, and (5) their impressions of Ameca’s physical appearance. In total, 31 unique participants were en-rolled in the study. Of these, 28 provided qualitative responses in Phase 1, and 20 returned and completed the open-ended questions in Phase 2. Because the study followed a repeated-interaction design, some participants contributed responses in both phases. Qualitative responses were analysed separately per phase to explore changes in per-ceived strengths and limitations of the interaction. The analysis focused on identifying recurring themes and illustrative patterns within each phase rather than on formal within-subject comparisons. The qualitative analysis followed Braun and Clarke’s six-phase thematic analysis approach (Braun & Clarke, 2006 ). Researchers first familiarized themselves with the data through repeated reading. Ini-tial codes were generated inductively to capture salient experiential aspects such as conversational flow, perceived listening, emotional engagement, personalization, and embodiment. Codes were iteratively refined and clustered into candidate themes based on patterns of shared meaning across participants. Themes were subsequently reviewed, defined, and refined through comparison across participants and across sessions, with attention to both convergence and divergence in experiences. Although the analysis was primarily inductive, the final thematic structure was interpreted in relation to the study’s research questions (RQ3–RQ4). Two researchers were involved in coding and interpretation. As members of the research team also contributed to the optimization of the conversational configuration, reflexive attention was paid to potential interpretive bias. Discrepancies were discussed until consensus was reached. Trustwor-thiness was supported through investigator triangulation, triangulation across qualitative responses, quantitative UX ratings, and conversational metrics, and the use of illustrative participant quotations to ground interpretations in the data. 4 Results 4.1 Descriptive Statistics of Baseline Measures Overall, participants rated Ameca positively across all UX domains, with mean scores ranging between 3.6 and 3.9, corresponding to the “agree” range of the Likert scale. Comprehensibility ( M = 3.9) re-ceived the highest ratings, indicating that participants generally found Ameca ’s answers clear and easy to follow. Social comfort ( M = 3.6) was rated slightly lower, suggesting that although partic-ipants enjoyed the interaction, feelings of companionship were somewhat less pronounced. Empathy ( M = 3.8), trust ( M = 3.8), and enjoyment ( M = 3.7) fell in between, reflecting that Ameca was generally perceived as an attentive and trustworthy conversational partner. The relatively small stan-dard deviations (0.6–0.7) indicate moderate agreement among participants, although some individual variation remained. To contextualize these evaluations, baseline psychosocial measures were examined (see Table 2 ). The mean UCLA-3 score was 4.1 (range = 3–9), indicating relatively low loneliness scores on this 3–12 scale, albeit with noticeable variability across participants. As the UCLA-3 does not have a universally agreed clinical cut-off, these values are reported descriptively. In contrast, WHO-5 results suggested moderate levels of subjective well-being, with a mean percentage score of 48.6%. The wide range of WHO-5 scores (0–100%) indicates substantial inter-individual differences, and several participants scored below the established clinical cut-off of 52%, commonly used to indicate reduced well-being (Topp et al., 2015 ). Taken together, these baseline measures demonstrate considerable heterogeneity in psychosocial functioning, with a subset of participants reporting both lower well-being and higher loneliness. Table 2 summarizes the descriptive statistics for the UCLA-3 Loneliness Scale and the WHO-5 Well-Being Index. Table 2: Descriptive statistics for UCLA-3 and WHO-5 scores (Phase 1 baseline). 4.2 Correlations Between UX Constructs, Loneliness, and Well-Being As shown in Table 3 , correlations between baseline psychosocial measures and UX constructs revealed a differentiated pattern. For loneliness (UCLA-3), higher scores were negatively associated with all UX domains, although most correlations were small to moderate and did not reach statistical significance. A statistically significant and comparatively stronger association was observed for comprehensibility and clarity ( r = − 0.53, p = .004), indicating that participants reporting higher loneliness tended to perceive the interaction as less clear and predictable. This represents a moderate-to-strong negative correlation and was the only domain reaching statistical significance. For well-being (WHO-5), all associations were positive, suggesting that higher subjective well-being was consistently related to more favorable UX evaluations. However, none of these correlations reached conventional levels of statistical significance. Table 3 Pearson correlations between UX constructs and baseline measures of loneliness (UCLA-3) and well-being (WHO-5), N = 31. Construct UCLA-3 WHO-5 r p r p A. General enjoyment − 0.18 .36 0.33 .09 B. Empathy and understanding − 0.30 .12 0.27 .16 C. Social comfort − 0.10 .61 0.16 .40 D. Trust and credibility − 0.24 .22 0.34 .08 E. Comprehensibility and clarity − 0 . 53 .004 0.26 .17 4.3 Comparative Analyses of Two Phases In line with RQ2, the study investigates differences in participants’ evaluations of Ameca across the two interaction sessions. Results are presented in Table 4 and indicate that across both sessions, participants rated Ameca consistently in the “agree” range (means 3.6–4.0 on a 5-point Likert scale). None of the paired com-parisons reached statistical significance ( p > 0.05), and all effect sizes were small ( |d| < 0.40). Table 4 Comparison of participants’ evaluations between Phase 1 and Phase 2 ( N = 20). Factor Phase 1 mean Phase 2 mean t (19) p Cohen’s d A. General enjoyment 3.9 4.0 -1.28 0.218 -0.29 B. Empathy 3.9 3.9 -0.43 0.669 -0.10 C. Social comfort 3.7 3.8 -1.61 0.126 -0.36 D. Trust 3.9 3.8 0.13 0.901 0.03 E. Clarity 4.0 3.8 0.56 0.579 0.13 Although mean scores for enjoyment and social comfort were numerically slightly higher in Phase 2, these differences should be interpreted as descriptive only. Given the limited sample size ( N = 20), the absence of adjustment for multiple comparisons, and the sequential two-phase design, no inferential claims can be made. Instead, these results indicate overall stability of user experience ratings across sessions, supporting the feasibility and acceptability of the interaction framework. 4.4 Qualitative Analysis of Open-Ended Responses Qualitative responses provided deeper insight into how participants experienced the interaction with Ameca beyond the structured questionnaire measures. In Phase 1 , 28 participants generally described the conversation as pleasant and novel, often emphasizing that it was “nice to have someone to talk to” and that the exchange felt “interesting” or “enriching.” Nearly all participants provided at least one positive remark about the interaction. Frequently mentioned strengths included Ameca’s clarity of speech and attentive demeanor. At the same time, the majority of participants formulated at least one suggestion for improvement. Approximately 5 participants explicitly referred to conversational interruptions or turn-taking issues, noting that the robot “interrupted too often” or “talked over” them. Others described the interaction as somewhat limited or experimental and expressed a desire for greater personalization and contextual continuity, suggesting that Ameca should “remember what I said earlier.” These responses indicate that conversational flow and memory persistence were central to participants’ expectations of natural interaction. In Phase 2 , feedback of 20 participants remained largely positive and appeared slightly more favorable overall, with fewer explicit references to interruption-related issues. Participants again em-phasized clarity, engagement, and the pleasant experience of companionship. Nevertheless, suggestions for deeper conversational content and improved contextual awareness remained present across both sessions. Importantly, physical appearance emerged as an additional and distinct component of the user ex-perience. When explicitly invited to reflect on Ameca’s embodiment, 56% of participants described her appearance as pleasant or lifelike, 38% as neutral, and only 6% expressed negative impressions. Posi-tive comments frequently referred to her friendly facial expression, realistic movements, and feminine design, with several participants noting that she was “pleasant to look at” or “almost human.” Only a small minority perceived her appearance as somewhat artificial or lacking warmth, suggest-ing that “she could smile more” or that she “still looks like a robot.” 4.5 Conversational Metrics and Thematic Analysis Table 5 presents aggregated descriptive statistics across 20 participants in Phase 2 . Only the converation transcripts ( N = 20) from Phase 2 were analyzed, as the technical setup during the first round did not yet allow for reliable recording and data extraction. In average, Ameca produced 49 turns and participants 54 turns, with approximately 30 questions posed by Ameca and 46 responses provided by participants. The mean length of participant answers was about 11 words, but ranged between 6 and 20 words. The proportion of short answers averaged 23% but ranged from only 2% in highly engaged participants to more than half of all responses in low-engagement cases. Table 5 Descriptive statistics of conversational metrics across all participants of Phase 2 Metric Mean SD Min Max Total utterances 102.2 34.6 42 169 Turns by Ameca 48.6 16.4 19 80 Turns by participant 53.6 19.1 23 96 Questions by Ameca 29.6 9.2 12 50 Answers by participant 45.6 18.0 18 76 Repairs 0.0 0.0 0 0 Average answer length (words) 11.0 3.5 5.68 19.72 Short answers proportion 0.2 0.1 0.02 0.53 Table 5 : Descriptive statistics of conversational metrics across all participants of Phase 2 Thematic analysis was conducted in parallel with descriptive conversational metrics to contextualize observed interaction patterns (Table 5 ). While aggregated metrics indicated relatively balanced turn-taking across participants (mean turns by Ameca = 48.6; participants = 53.6), substantial inter-individual variability was observed in response length and elaboration. In particular, the proportion of short responses ranged from 0.02 to 0.53, indicating that in low-engagement cases, more than half of participant contributions consisted of minimal or non-elaborative answers. Against this quantitative backdrop, thematic analysis focused on identifying patterns of meaning that could explain variability in conversational engagement. Three analytically distinct themes were identified. The first theme concerned autobiographical engagement as a driver of conversational depth . Par-ticipants who engaged with autobiographical or emotionally salient topics—such as travel experiences, family relationships, artistic interests, or bereavement—tended to produce longer responses and fewer short answers. These interactions were characterized by narrative continuity and reciprocal elabora-tion, suggesting that conversational depth emerged when participants framed the interaction as an opportunity for personal storytelling rather than question answering. This pattern aligns with the higher mean response lengths observed in high-engagement cases. A second theme captured interactional asymmetry and question-driven structure . Across tran-scripts, Ameca posed an average of 29.6 questions per interaction, reflecting a predominantly inter-rogative conversational structure. In cases where participants provided brief or hesitant answers, this structure resulted in rapid topic cycling and limited conversational buildup. These low-engagement interactions corresponded to higher proportions of short responses (up to 53%), indicating that con-versational depth was constrained not by turn-taking balance per se, but by limited elaboration within participant turns. The third theme addressed turn-taking sensitivity and perceived listening . Despite balanced turn counts, several participants commented on interruptions or premature responses following pauses. These qualitative observations suggest that conversational smoothness cannot be inferred from turn distribution alone, but is sensitive to micro-level timing and pause interpretation. In several instances, participants explicitly requested additional silence or clarified that they were still speaking, highlighting the importance of temporal alignment for perceived listening and social comfort. Taken together, these themes indicate that conversational engagement in LLM-enabled humanoid interaction is co-constructed and heterogeneous. Quantitative metrics reveal overall feasibility and structural balance, while qualitative themes elucidate why some interactions developed into extended, meaningful exchanges whereas others remained superficial. Importantly, conversational depth was not uniformly produced by the optimized prompting configuration, but depended on participants’ willingness and capacity to elaborate, as well as on fine-grained interactional features such as timing, continuity, and responsiveness. 5 Discussion The present study examined how older adults experienced short, structured conversations with the humanoid robot Ameca in a residential-care setting, combining subjective user experience (RQ1–RQ2), transcript-based interaction characteristics available for Phase 2 (RQ3), and qualitative reflections (RQ4). 5.1 User Experience and Acceptance (RQ1–RQ2) Regarding RQ1 and RQ2 , participants’ perceptions of Ameca were generally positive across en-joyment, perceived empathy, social comfort, trust, and comprehensibility. Among participants who completed both sessions ( N = 20), UX ratings remained largely stable, with only small descriptive differences between Phase 1 and Phase 2. Given the sequential two-phase design and the concur-rent prompt-level refinements, these session-level differences cannot be causally attributed to either repeated exposure (familiarity) or to conversational optimization. Rather than implying improvement effects, the overall stability suggests that initial acceptability was maintained at re-contact, supporting feasibility in a residential-care setting. This interpretation aligns with prior work on socially assistive robots indicating that perceived empathy and trust tend to depend more on conversational coherence and affective cues than on familiarity alone (Broadbent, 2016). Importantly, the data and participants’ comments nonetheless underscore that perceived inter-action quality is sensitive to the configuration layer on top of the underlying LLM. Even without changes to hardware or the base model, interaction design decisions—including prompt and context engineering, explicit turn-taking rules, topic steering, and socio-emotional backchannels—appear to shape perceived naturalness, pacing, and social comfort. This design-relevant insight is consistent with recent feasibility work on LLM-augmented robots in healthcare contexts, which similarly empha-sizes interaction quality, clarity, and pacing as key drivers of user acceptance (van ’t Klooster et al., 2026 ). Likewise, user-centered LLM architectures for embodied robots highlight that conversational scaffolding and lightweight memory mechanisms can enhance perceived trust and social presence, while noting persistent challenges around robustness and latency (Pinto-Bernal et al., 2025 ). Finally, correlations between psychosocial baselines and UX evaluations point to meaningful hetero-geneity in how older adults experience humanoid interaction. Participants reporting higher loneliness tended to evaluate the interaction less favorably, most notably in comprehensibility. This suggests that baseline social vulnerability may be associated with lower perceived predictability or shared un-derstanding in robot dialogue; however, this result should be interpreted cautiously given the lower internal consistency of the comprehensibility/clarity domain in this sample. 5.2 Conversational Characteristics Under the Phase 2 Configuration (RQ3) In relation to RQ3 , transcript-based conversational metrics and thematic analyses provide insight into interaction patterns observed under the Phase 2 configuration. At an aggregate level, turn-taking was relatively balanced, suggesting reciprocal participation. At the same time, the interaction remained predominantly question-driven: Ameca posed many questions and participants primarily responded, reflecting a scaffolded, interview-like interaction structure rather than fully bidirectional free conversation. Crucially, metrics revealed substantial inter-individual variability in engagement (e.g., response length and short-answer proportion), which was echoed in the thematic interpretation. Participants who shifted toward autobiographical or emotionally salient topics (e.g., travel memories, family rela-tionships, bereavement) tended to provide longer responses and fewer minimal answers, indicating nar-rative continuity and reciprocal elaboration. In lower-engagement cases, brief answers combined with rapid topic cycling constrained conversational build-up. These patterns are consistent with human–robot interaction accounts emphasizing that conversational depth is co-constructed rather than gener-ated by the robot alone (Karaaziz, Can, & Keskinda˘g, 2017 ). The link between richer responses and personal themes further aligns with empathy and social presence perspectives suggesting that emotional salience can facilitate deeper conversational involvement (Morgante, Susinna, Culicetto, Quartarone, & Lo Buono, 2024 ). 5.3 Qualitative Reflections, Speech Robustness, and Embodiment (RQ4) RQ4 contextualizes these patterns through participants’ reflections. Across sessions, participants fre-quently described the interaction as “pleasant,” “interesting,” and “clear,” often highlighting compan-ionship and curiosity as motives for engagement. At the same time, three recurring improvement needs were consistently reported: (1) limited personalization and memory, (2) abrupt topic shifts, and (3) restricted emotional depth. These concerns map directly onto known constraints of many current LLM-driven social robots, particularly the absence of cross-session memory and adaptive context modeling. They are also consistent with prior work indicating that memory continuity and adaptive pacing are central to sustaining social presence, naturalness, and trust in embodied LLM systems (Pinto-Bernal et al., 2025 ; Verhelst & Belpaeme, 2025 ). Participants also occasionally requested repetition or noted non-response when they spoke softly or used dialect. These remarks align with known ASR challenges for older adults (e.g., reduced speech volume, slower articulation, dialectal variation). While the Whisper-based pipeline generally enabled fluent interaction, the findings suggest that future deployments could benefit from confidence-based clarification strategies, targeted robustness improvements for elderly speech, and dialect-aware handling in multilingual contexts such as Dutch (Flemish). Perceptions of Ameca ’s embodiment further shaped acceptance. Most participants described the robot as pleasant or lifelike, appreciating facial expressiveness and overall approachability. This aligns with accounts of moderate human-likeness within the uncanny-valley continuum (Mara et al., 2022 ). A minority still perceived the robot as insufficiently warm (e.g., “could smile more”), underscoring that perceived emotional authenticity depends not only on language generation but also on fine-grained embodied cues such as gaze timing, facial affect, and turn-taking sensitivity (Kwak, 2014 ). 5.4 Synthesis and Implications Taken together, these findings reinforce that humanoid robots such as Ameca can be feasible and generally well accepted for brief conversational encounters in residential care, while highlighting de-sign priorities for achieving more sustained, relationship-like interaction. Importantly, the present results should be interpreted as feasibility and interaction-quality evidence rather than evidence of improvement across sessions or clinical effectiveness, as the sequential design does not disentangle familiarity from prompt-level refinement. In line with recent feasibility studies of LLM-augmented robots in healthcare, the study suggests that acceptance depends strongly on interaction quality, clar-ity, and pacing (van ’t Klooster et al., 2026 ). At the same time, consistent with design-oriented work on embodied LLM systems, sustained naturalness remains contingent on advances in personal-ization, memory continuity, and robust turn-taking management (Pinto-Bernal et al., 2025 ). Future systems may therefore prioritize (i) improved pause detection and turn-taking sensitivity, (ii) robust clarification and repair strategies for older-adult speech and dialect, and (iii) retrieval-based memory mechanisms to support continuity and personalization across sessions. 6 Limitations and Future Directions Several limitations should be acknowledged. First, the sample size was modest, particularly for the repeated-measures component ( N = 20), limiting statistical power and generalizability. Second, ses-sions were brief (15–20 minutes), capturing short-term impressions rather than sustained relational dynamics. Third, the study lacked systematic logging of technical performance metrics (e.g., response latency, ASR confidence/accuracy), constraining deeper technical interpretation. A key limitation concerns the sequential two-phase design in which the baseline configuration was assessed in Phase 1 and the refined configuration in Phase 2. Because conditions were not counterbal-anced, session differences cannot be causally attributed to conversational refinement versus familiar-ity/repeated exposure; accordingly, session-level comparisons should be interpreted descriptively. Importantly, the present study should not be interpreted as a loneliness intervention. Although loneliness was assessed using the UCLA-3 scale, this measure was included to characterize baseline psychosocial context and to explore associations with perceived interaction quality, rather than to evaluate change or long-term effects. The brief two-session design and short interaction duration preclude causal conclusions regarding loneliness outcomes. Accordingly, references to long-term impact or intervention effects have been avoided, and findings are framed in terms of feasibility, acceptability, and short-term interaction experiences. Finally, the conversational model remained constrained by limited personalization and lack of cross-session memory. Although Phase 2 introduced prompt-level refinements, Ameca could not recall prior exchanges or adapt to individual preferences beyond the immediate context. Retrieval-augmented generation and lightweight memory continuity mechanisms represent promising avenues to address this constraint and to enable more coherent, relationship-like interaction over time. In parallel with these research directions, real-world deployment feedback can also inform iterative platform development. Where applicable, future work should document and evaluate interaction-level updates (e.g., voice quality, pause detection, facial synchronization) with systematic performance logging and controlled study designs. Longitudinal protocols and randomized/counterbalanced designs will be especially valuable to evaluate sustained engagement and to separate configuration effects from familiarity and learning effects. 7 Conclusion This two-session feasibility study indicates that short, structured conversations with an LLM-enabled humanoid robot ( Ameca ) are technically feasible and generally acceptable for older adults in a residential-care context. Participants reported overall positive UX ratings, and Phase 2 transcripts showed re- ciprocal participation with balanced turn-taking but a predominantly question–answer interaction structure. A central contribution is that interaction quality in embodied LLM systems is shaped not only by the underlying language model, but also by the conversational configuration layer (prompt and context engineering), which influences pacing, perceived listening, and socio-emotional backchannels without requiring changes to hardware or the base model. At the same time, qualitative feedback and interaction patterns point to persistent challenges, including limited personalization and memory continuity, occasional turn-taking timing issues, and the need for robust clarification strategies for older-adult speech and dialect. Together, these findings support further research using counterbalanced or randomized designs and longer-term deployments to assess how improvements in contingent responsiveness, personalization, and embodied warmth translate into sustained engagement and potential psychosocial benefit in care settings. Declarations Ethical approval This study was conducted in accordance with the principles of the Declaration of Helsinki. The research protocol was reviewed and approved in accordance with the institutional ethical procedures of Howest University of Applied Sciences. Given the non-invasive nature of the study and the absence of medical or therapeutic intervention, no formal medical ethics committee approval was required under applicable national regulations. Data Availability Statement Due to privacy considerations, the raw physiological and behavioral data from participants are not publicly available. Aggregated results and analysis scripts can be obtained from the corresponding author upon reasonable request. Consent to participate All participants received verbal and written information about the study prior to participation. Written informed consent was obtained from all participants before data collection. Participants were informed that participation was voluntary and that they could withdraw at any time without consequences. Consent to publish All participants provided explicit consent for the anonymized use of their data for scientific publica-tion. Any visual materials included in this manuscript were anonymized to prevent identification of individual participants. Declaration of interests The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Declaration of generative AI and AI-assisted technologies in the writing process During the preparation of this work the authors sometimes used ChatGPT 5 in order to rephrase smaller parts of the text for clarity or to correct language errors. The authors always reviewed and edited the content as needed and take full responsibility for the content of the published article. Collaboration Declaration This study was conducted in collaboration with Yin Oei (CEO at Living Tomorrow group) and Lisa Vanryckegem (i-mens). Living Tomorrow facilitated access to the humanoid robot platform and the experimental setting, while i-mens supported participant recruitment and on-site coordination within the residential care context. Technical platform support was provided by Engineered Arts. References Banks, M. R., Willoughby, L. M., & Banks, W. A. (2008). Animal-assisted therapy and loneliness in nursing homes: Use of robotic versus living dogs. Journal of the American Medical Directors Association , 9 (3), 173–177. doi: https://doi.org/10.1016/j.jamda.2007.11.007 Boamah, S. A., Weldrick, R., Lee, T.-S. J., & Taylor, N. (2021). Social isolation among older adults in long-term care: A scoping review. Journal of Aging and Health , 33 (7–8), 618–632. doi: https://doi.org/10.1177/08982643211004174 Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychology , 3 (2), 77–101. doi: 10.1191/1478088706qp063oa Broadbent, E. (2016). Interactions with robots: The truths we reveal about ourselves. Annual Review of Psychology , 68 (1), 627–652. doi: https://doi.org/10.1146/annurev-psych-010416-043958 Broadbent, E., Loveys, K., Ilan, G., Chen, G., Chilukuri, M. M., Boardman, S. G.,. .. Skuler, D. (2024). Elliq, an ai-driven social robot to alleviate loneliness: Progress and lessons learned. The Journal of Aging Research & Lifestyle , 13 , 22–28. doi: https://doi.org/10.14283/jarlife.2024.2 Cacioppo, J. T., & Cacioppo, S. (2014). Social relationships and health: The toxic effects of perceived social isolation. Social and Personality Psychology Compass , 8 (2), 58–72. doi: 10.1111/spc3 .12087 Campbell IV, W. A., Chick, J. F. B., Shin, D., & Makary, M. S. (2024). Understanding chatgpt for evidence-based utilization in interventional radiology. Clinical Imaging , 108 , 110098. doi: https://doi.org/10.1016/j.clinimag.2024.110098 Carson, D. B., Johansson, A. B., Schaumberg, M., & Hurtig, A.-K. (2024). Addressing the workforce crisis in rural social care: A scoping review. The International Journal of Health Planning and Management , 39 (3), 806–823. doi: https://doi.org/10.1002/hpm.3774 Coghlan, S., & Waycott, J. (2022). Can robots really be companions for older adults? Retrieved from https://research.unimelb.edu.au/strengths/updates/news/can-robots-really-be-companions-for-older-adults . Courtin, E., & Knapp, M. (2017). Social isolation, loneliness and health in old age: a scoping review. Health & Social Care in the Community , 25 (3), 799–812. doi: 10.1111/hsc.12311 Cronbach, L. J. (1951). Coefficient alpha and the internal structure of tests. Psychometrika , 16 (3), 297–334. De Jong-Gierveld, J., & Van Tilburg, T. G. (1999). Manual of the loneliness scale . Methoden en Technieken. Donovan, N. J., & Blazer, D. G. (2020). Social isolation and loneliness in older adults: Review and commentary of a national academies report. The American Journal of Geriatric Psychiatry , 28 (12), 1233–1244. doi: 10.1016/j.jagp.2020.08.005 Fiolka, A. K., Donnermann, M., & Lugrin, B. (2024). Investigating the mere exposure effect in relation to perceived eeriness and humaneness of a social robot., 453–457. Følstad, A., & Brandtzæg, P. B. (2017). Chatbots and the new world of hci. Interactions , 24 (4), 38–42. doi: https://doi.org/10.1145/3085558 Gardiner, C., Laud, P., Heaton, T., & Gott, M. (2020). What is the prevalence of loneliness amongst older people living in residential and nursing care homes? a systematic review and meta-analysis. Age and Ageing , 49 (5), 748–757. doi: 10.1093/ageing/afaa049 Hancock, P. A., Billings, D. R., Schaefer, K. E., Chen, J. Y., de Visser, E. J., & Parasuraman, R. (2011). A meta-analysis of factors affecting trust in human-robot interaction. Human Factors , 53 (5), 517–527. doi: https://doi.org/10.1177/0018720811417254 Hughes, M. E., Waite, L. J., Hawkley, L. C., & Cacioppo, J. T. (2004). A short scale for measuring loneliness in large surveys. Research on Aging , 26 (6), 655–672. doi: 10.1177/0164027504268574 Hung, L., Liu, C., Woldum, E., Au-Yeung, A., Berndt, A., Wallsworth, C.,.. . Chaudhury, H. (2019). The benefits of and barriers to using a social robot paro in care settings: A scoping review. BMC Geriatrics , 19 (1), 232. doi: https://doi.org/10.1186/s12877-019-1244-6 Karaaziz, M., Can, G., & Keskinda˘g, B. (2017). Examining empathy as a communication technique in the context of gender and cultural differences: A review study. Global Journal of Psychology Research: New Trends and Issues , 7 (1), 20–27. doi: https://doi.org/10.18844/gjpr.v7i1.401 Kwak, S. S. (2014). The impact of robot appearance types on social interaction with a robot and service evaluation of a robot. Archives of Design Research , 27 (2), 81–93. doi: https://doi.org/10.15187/adr.2014.05.110.2.81 Lal, S. (2025). Analyzing the impact of social isolation on aging: Health, mental state and cognitive functions. Critical Debates in Humanities, Science and Global Justice . Lapane, K. L., et al. (2022). Health effects of loneliness and social isolation in older adults living in congregate long-term care settings: A review. Journal of Aging Social Policy . doi: 10.1080/ 08959420.2022.2046767 Leite, I., Castellano, G., Pereira, A., Martinho, C., Paiva, A., & McOwan, P. W. (2014). Empathic robots for long-term interaction. International Journal of Social Robotics , 6 (3), 329–341. doi: https://doi.org/10.1007/s12369-014-0227-1 Mara, M., Appel, M., & Gnambs, T. (2022). Human-like robots and the uncanny valley. Zeitschrift fu¨r Psychologie . doi: 10.1027/2151-2604/a000486 Morgante, E., Susinna, C., Culicetto, L., Quartarone, A., & Lo Buono, V. (2024). Is it possible for people to develop a sense of empathy toward humanoid robots and establish meaningful relationships with them? Frontiers in Psychology , 15 , 1391832. doi: https://doi.org/10.3389/fpsyg.2024.1391832 Pinto-Bernal, M., Biondina, M., & Belpaeme, T. (2025). Designing social robots with large language models for engaging human interaction. Applied Sciences , 15 (11), 6377. doi: https://doi.org/10.3390/app15116377 Santini, Z. I., Jose, P. E., Cornwell, E. Y., Koyanagi, A., Nielsen, L., Hinrichsen, C.,. .. Koushede, V. (2020). Social disconnectedness, perceived isolation, and symptoms of depression and anxiety among older americans (nshap): A longitudinal mediation analysis. The Lancet Public Health , 5 (1), e62–e70. doi: https://doi.org/10.1016/S2468-2667(19)30230-0 Smith, R. W., Barnes, I., Green, J., Reeves, G. K., Beral, V., & Floud, S. (2021). Social isolation and risk of heart disease and stroke: Analysis of two large uk prospective studies. The Lancet Public Health , 6 (4), e232–e239. doi: https://doi.org/10.1016/S2468-2667(20)30291-7 Spatola, N., Belletier, C., Chausse, P., & Augustinova, M. (2021). Social perception of artificial agents: A meta-analysis. Psychological Bulletin , 147 (11), 1121–1143. doi: 10.1037/bul0000340 Topp, C. W., Østergaard, S. D., Søndergaard, S., & Bech, P. (2015). The who-5 well-being index: A systematic review of the literature. Psychotherapy and Psychosomatics , 84 (3), 167–176. doi: 10.1159/000376585 Tuneu, M. M., Roig, J. J., Molina, E. M., Planas, L. C., Salvans, A. E., Godayol, P. F.,.. . others (2023). Social and emotional loneliness among older people living in nursing homes in spain: A cross-sectional study. Anales de Psicolog´ıa / Annals of Psychology , 39 (3), 465–477. doi: https://doi.org/10.20944/preprints202201.0423.v1 van ’t Klooster, J., Capasso, M., van Gorssel, D., Vrolijk, E., Rettagliata, G., Gerritsen, D.,.. . Vonkeman, H. E. (2026). A gpt-reinforced social robot for patient communication: a pilot study. Frontiers in Digital Health , 7 , 1653168. doi: 10.3389/fdgth.2025.1653168 Verhelst, E., & Belpaeme, T. (2025). Large language models cover for speech recognition mistakes: Evaluating conversational ai for second language learners., 1705–1709. doi: https://doi.org/10.1109/HRI61500.2025.10974188 World Health Organization. (2025). Population ageing: Questions and answers. Retrieved from https://www.who.int/news-room/questions-and-answers/item/population-ageing . Yen, H.-Y., Huang, C. W., Chiu, H.-L., & Jin, G. (2024). The effect of social robots on depression and loneliness for older residents in long-term care facilities: A meta-analysis of randomized controlled trials. Journal of the American Medical Directors Association , 25 (6). doi: https://doi.org/10.1016/j.jamda.2024.02.017 Additional Declarations No competing interests reported. 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-9212133","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":628899103,"identity":"2c848a90-c501-4e4e-a87f-006dc673969c","order_by":0,"name":"Charlotte Larmuseau","email":"data:image/png;base64,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","orcid":"","institution":"University College West Flanders","correspondingAuthor":true,"prefix":"","firstName":"Charlotte","middleName":"","lastName":"Larmuseau","suffix":""},{"id":628899104,"identity":"47f43c07-9429-4ad0-bd76-57dd722ce74a","order_by":1,"name":"Mattice Rigole","email":"","orcid":"","institution":"Ghent University","correspondingAuthor":false,"prefix":"","firstName":"Mattice","middleName":"","lastName":"Rigole","suffix":""},{"id":628899105,"identity":"782c82c9-2a12-456e-82c6-d4e6b33dd339","order_by":2,"name":"Jens Krijgsman","email":"","orcid":"","institution":"University College West Flanders","correspondingAuthor":false,"prefix":"","firstName":"Jens","middleName":"","lastName":"Krijgsman","suffix":""}],"badges":[],"createdAt":"2026-03-24 12:53:55","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9212133/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9212133/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108006431,"identity":"47e5cb5e-d555-4bdd-8f6d-304e918fe054","added_by":"auto","created_at":"2026-04-28 12:55:33","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":15086,"visible":true,"origin":"","legend":"\u003cp\u003eAmeca humanoid robot during the conversational interaction in the residential care setting.\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9212133/v1/18ac20eddefd23bfc098d756.jpg"},{"id":107897612,"identity":"f4acf998-415a-4070-b87b-819d5cd7c316","added_by":"auto","created_at":"2026-04-27 10:59:32","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":771434,"visible":true,"origin":"","legend":"\u003cp\u003eExample interactions with the humanoid robot \u003cem\u003eAmeca\u003c/em\u003e. Participant faces have been blurred to preserve anonymity.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9212133/v1/21da850c2b650ded65d46375.png"},{"id":108006109,"identity":"7b95f150-7c5e-4b63-a1aa-1d38086897ba","added_by":"auto","created_at":"2026-04-28 12:53:22","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":21789,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of the two-session mixed design (baseline and follow-up after prompt-level revi-sion).\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9212133/v1/5b74fb6170ca746807ac5270.jpg"},{"id":108490828,"identity":"4d44abbe-d21f-4eee-b191-00b2195ab03d","added_by":"auto","created_at":"2026-05-05 09:49:01","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1619708,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9212133/v1/97e823c3-018f-4e00-9550-6bfba23bd5ed.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"LLM-Enabled Humanoid Interaction in Care Settings: A Two-Phase Study with Ameca","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eLoneliness among older adults is a growing public health concern, particularly in residential care settings where reduced mobility, loss of social networks, and limited family contact can contribute to social isolation (Courtin \u0026amp; Knapp, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Persistent loneliness has been linked to adverse health outcomes, including cognitive decline, depression, and increased mortality risk (Cacioppo \u0026amp; Cacioppo, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Simultaneously, healthcare systems and long-term care facilities face persistent staff shortages, limiting the time available for meaningful one-on-one conversations and emotional support.\u003c/p\u003e \u003cp\u003eAdvances in social robotics offer new opportunities to address these challenges. Humanoid robots, such as \u003cem\u003eAmeca\u003c/em\u003e, are capable of naturalistic speech, facial expressions, and responsive interaction, making them a potential tool to supplement human contact and provide moments of social engage-ment (Broadbent, 2016). The present work evaluates an LLM-enabled humanoid interaction frame-work deployed in a real residential-care context, focusing on feasibility and UX outcomes and how these relate to baseline psychosocial differences.\u003c/p\u003e \u003cp\u003eSocially assistive robots may influence loneliness-related experiences through proximal interaction mechanisms rather than direct clinical effects. First, humanoid embodiment can increase \u003cem\u003eperceived so-cial presence\u003c/em\u003e\u0026mdash;the sense of \u0026ldquo;being with\u0026rdquo; a responsive other\u0026mdash;through gaze behavior, facial expressive-ness, and responsive backchannels. Second, \u003cem\u003econtingent responsiveness\u003c/em\u003e (timely, contextually appropriate\u003c/p\u003e \u003cp\u003ereplies) supports reciprocity and can reduce the experience of one-sided communication. Third, \u003cem\u003ecom-panionship cues\u003c/em\u003e (affiliative language, warmth, humor, and empathic acknowledgments) can increase approachability and momentary feelings of companionship. Fourth, structured prompts and remi-niscence topics can enhance \u003cem\u003eengagement\u003c/em\u003e, encouraging conversational participation and self-disclosure. In this feasibility study, these mechanisms are operationalized through UX ratings (enjoyment, per-ceived empathy, social comfort, trust, and clarity) and conversational interaction metrics as proximal indicators that may precede longer-term psychosocial change.\u003c/p\u003e"},{"header":"2 Theoretical Framework","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Loneliness in Residential Care as a Socio-Technical Challenge: Impli-cations for Scalable Care Technologies\u003c/h2\u003e \u003cp\u003eLoneliness in later life refers to a perceived mismatch between the quantity or quality of social re-lationships people have and those they desire, reflecting not merely social contact but unmet social needs (De Jong-Gierveld \u0026amp; Van Tilburg, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e1999\u003c/span\u003e). In 2024, global life expectancy has risen to 73.3 years, and the population aged 60 and over is expected to grow from 1.1\u0026nbsp;billion in 2023 to 1.4\u0026nbsp;billion by 2030 (World Health Organization, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Importantly, older adults living in residential care and long-term care facilities experience particularly high levels of loneliness compared with community-dwelling peers, with systematic reviews estimating that approximately 61% of residents experience moderate loneliness and around 35% severe loneliness in care homes (Gardiner, Laud, Heaton, \u0026amp; Gott, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Loneliness in congregate care settings is associated with adverse mental and physical health outcomes, emphasizing the unique social and structural challenges faced within institutional care con-texts (Lapane et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Donovan \u0026amp; Blazer, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Within residential care environments, these risks are further amplified by practical constraints: staff shortages, time pressure, and increasing care demands often limit opportunities for extended one-on-one conversations and emotionally meaning-ful interactions (Boamah, Weldrick, Lee, \u0026amp; Taylor, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Carson, Johansson, Schaumberg, \u0026amp; Hurtig, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). While traditional interventions such as animal-assisted therapy or organized group activities can temporarily reduce loneliness, they are typically resource-intensive, difficult to scale, and challenging to sustain over time (Banks, Willoughby, \u0026amp; Banks, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2008\u003c/span\u003e; Yen, Huang, Chiu, \u0026amp; Jin, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eSocial isolation has diverse consequences. People with limited social contact show higher rates of physical health conditions (e.g., hypertension, arthritis, and heart disease) and increased risk of cognitive decline and dementia, alongside reduced emotional resilience (Lal, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Lower social contact can delay diagnosis of health problems and is associated with increased mortality risk (Smith et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Furthermore, social disconnection is linked to poorer perceived health and increased symptoms of depression and anxiety (Santini et al., 2020; Tuneu et al., \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Given these wide-ranging effects, there is a growing need for innovative, scalable approaches that can complement human care and support social and emotional well-being in later life.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Large Language Models, care applications, and social interaction\u003c/h2\u003e \u003cp\u003eLarge Language Models (LLMs) are software systems designed to generate text based on prompts, pro-ducing contextually appropriate responses through probabilistic next-word prediction (Campbell IV, Chick, Shin, \u0026amp; Makary, 2024). In mental health and older-adult care, LLM-based systems are increas-ingly explored for conversational support, including psycho-educational explanations, motivational coaching, guided reminiscence, companionship-style dialogue, and practical assistance (e.g., reminders or informational support), typically under safety constraints that prohibit medical diagnosis or sensi-tive content. In these settings, the primary promise is not to replace caregivers, but to provide scalable conversational availability and adaptive dialogue that may support engagement and perceived social presence when human time for extended one-on-one interaction is limited.\u003c/p\u003e \u003cp\u003eRelated feasibility work on LLM-augmented social robots further illustrates this development. Van \u0026rsquo;t Klooster et al. (\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2026\u003c/span\u003e) report on the design and pilot evaluation of a GPT-reinforced social robot intended to support patient\u0026ndash;provider communication in clinical practice. The robot was connected to whitelisted, validated clinical sources and engaged users in natural dialogue to convey disease-and treatment-related information. User experience (UX) was assessed using the UX Questionnaire (UEQ) and semi-structured interviews in both a laboratory pilot and a hospital deployment with\u003c/p\u003e \u003cp\u003eosteoarthritis patients and healthcare professionals. While the study focused on informational and clinical communication rather than emotionally oriented dialogue, it demonstrated the feasibility of LLM-enhanced social robots in applied healthcare contexts and highlighted the importance of inter-action quality, clarity, and pacing for user acceptance (van \u0026rsquo;t Klooster et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2026\u003c/span\u003e). Building on such feasibility work, Pinto-Bernal, Biondina, and Belpaeme (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) specifically examine the integration of LLMs into a physically embodied social robot Pepper with the aim of supporting engaging, socially oriented interaction. Their work describes a user-centered LLM architecture that combines real-time speech processing, persona conditioning, adaptive turn-taking, and lightweight memory mechanisms to sustain conversational coherence over time. Through iterative evaluations, including exploratory interactions with older adults, the authors show that features such as memory continuity and adap-tive dialogue pacing contribute to higher perceived trust, naturalness, and social presence, while also noting persistent challenges related to latency and robustness. This study underscores the importance of interaction design and UX evaluation when deploying LLM-enabled humanoid robots for socially meaningful conversations with older adults (Pinto-Bernal, Biondina, \u0026amp; Belpaeme, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWithin social robotics research more broadly, LLMs have been studied for enabling adaptive turn-taking, culturally attuned conversational style, and more natural dialogue flow, which can shape user perceptions of trust, naturalness, and social presence (Pinto-Bernal et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Verhelst \u0026amp; Belpaeme, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). However, the effectiveness and acceptability of such systems depend not only on language generation capabilities, but also on how users respond to embodied agents and how interaction design supports comfortable pacing, clarity, and rapport in real-world care settings.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Robot\u0026ndash;human interaction and social presence in elderly care\u003c/h2\u003e \u003cp\u003eWhile LLMs enable increasingly natural and adaptive dialogue, the social and emotional impact of such systems ultimately depends on how older adults perceive and experience embodied robotic agents. Research on human\u0026ndash;robot interaction demonstrates that robots can elicit social and affective responses that, in some cases, resemble those triggered by living beings. A frequently cited comparison is animal-assisted therapy, which has been shown to reduce social isolation among older adults. Studies comparing animal-assisted interventions with robotic alternatives have found no significant differences in loneliness outcomes, suggesting that social bonding and emotional engagement can also emerge in interactions with artificial agents (Banks et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2008\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eEmpirical work in long-term care settings further indicates that interactions with socially assistive robots may be associated with reduced loneliness and depressive symptoms for some older adults, although findings remain heterogeneous and context-dependent (Yen et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). These mixed out-comes highlight that conversational capability alone is insufficient; the social meaning of interaction is shaped by embodiment, interaction style, and user expectations.\u003c/p\u003e \u003cp\u003eEarlier generations of socially assistive robots (SARs), such as \u003cem\u003ePARO\u003c/em\u003e and \u003cem\u003eElliQ\u003c/em\u003e, illustrate both the promise and limitations of robotic companionship. While many users report positive affect, curiosity, and engagement, recurring challenges include high costs, technical reliability, and a perceived lack of conversational depth or authenticity (Hung et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Broadbent et al., 2024; Coghlan \u0026amp; Waycott, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). These findings underscore that acceptance is influenced not only by what a robot can say, but also by how it looks, behaves, and responds over time.\u003c/p\u003e \u003cp\u003eAppearance and embodied behavior play a critical role in shaping user responses. Human-oriented designs tend to elicit stronger feelings of social presence and sociability, whereas product-oriented designs are often evaluated as more efficient or predictable (Kwak, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Psychological theories provide additional insight: the mere-exposure effect suggests that repeated interaction can increase comfort and familiarity, while the uncanny-valley hypothesis posits that near-human but imperfect realism may provoke discomfort or unease (Fiolka, Donnermann, \u0026amp; Lugrin, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mara, Appel, \u0026amp; Gnambs, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTaken together, these findings suggest that the evaluation of LLM-enabled humanoid robots in elderly care must consider not only conversational intelligence, but also embodiment, familiarity, and perceived social presence. For expressive humanoids such as \u003cem\u003eAmeca\u003c/em\u003e, conversational comfort, trust, and social acceptance emerge from the interplay between advanced dialogue capabilities and human-centered interaction design.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Synthesis and research questions\u003c/h2\u003e \u003cp\u003eAcross residential-care contexts, persistent loneliness among older adults coexists with increasing work-load pressures and limited opportunities for sustained one-on-one interaction. While traditional psy-chosocial interventions can alleviate loneliness, they are often difficult to maintain at scale within institutional care environments. Recent advances in LLM-enabled social robotics introduce opportu-nities for more natural, adaptive dialogue, yet evidence remains limited on how such conversationally capable humanoid robots are experienced by older adults in real-world care settings.\u003c/p\u003e \u003cp\u003eAgainst this backdrop, the present study examines interactions with \u003cem\u003eAmeca\u003c/em\u003e, an expressive humanoid robot equipped with LLM-based conversational capabilities, deployed in a Flemish residential-care context. By combining standardized measures, UX ratings, conversational metrics, and qualitative feedback, the study aims to clarify how older adults experience short, structured conversations with a humanoid robot and how baseline psychosocial characteristics relate to perceived interaction quality.\u003c/p\u003e \u003cp\u003eThe study addresses the following research questions:\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRQ1\u003c/strong\u003e \u003cp\u003eHow do older adults experience short, structured conversations with the humanoid robot \u003cem\u003eAmeca\u003c/em\u003e in terms of enjoyment, empathy, social comfort, trust, comprehensibility, and perceived appear-ance?\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRQ2\u003c/strong\u003e \u003cp\u003eAmong participants who completed both sessions (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;20), how do UX ratings differ between Session 1 (baseline) and Session 2 (optimized), and what is the magnitude and uncertainty of these differences?\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRQ3\u003c/strong\u003e \u003cp\u003eWhat conversational characteristics (turn-taking balance, question\u0026ndash;answer patterns, response length, and themes) are observed under the optimized configuration based on Phase 2 transcripts?\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRQ4\u003c/strong\u003e \u003cp\u003eHow do older adults qualitatively describe their interaction with \u003cem\u003eAmeca\u003c/em\u003e, with particular attention to perceived conversational naturalness and embodied presence?\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"3 Methodological Framework","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Ameca Humanoid Robot Platform\u003c/h2\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, \u003cem\u003eAmeca\u003c/em\u003e is a humanoid robot developed by Engineered Arts Ltd. (UK), designed as a research platform for human-robot interaction. The system features a life-sized torso with 17 facial actuators, allowing for expressive gestures and natural gaze behaviour. Its embedded microphones, stereo speakers, and high-resolution RGB cameras enable real-time audiovisual sensing and speech processing. In this study, \u003cem\u003eAmeca\u003c/em\u003e operated in conversational mode via the Tritium AI interface, in-tegrating an LLM-based dialogue model (GPT-4o), automatic speech recognition (Whisper large-v3), and neural text-to-speech synthesis. All inference ran through a secured cloud connection with sub-second response latency (typically 0.6-1.0 s end-to-end). Safety filters and topic restrictions prevented sensitive, medical, or personally identifying content, and no audio data were stored beyond temporary ASR buffering. The robot\u0026rsquo;s dialogue manager was configured to maintain a friendly, supportive tone, short sentence structure, and clear articulation tailored to older users.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Participants\u003c/h2\u003e \u003cp\u003eA total of 31 older adults participated in the study. The mean age was 83.2 years (\u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;8.3), with 23 women and 8 men. Most participants resided in a group assisted living facility (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;19), while others lived in a care hotel (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;8), at home in the neighborhood (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2), or in a nearby apartment (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2). These characteristics reflect a relatively diverse sample of older adults living in both residential and community-based contexts.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Procedure and Measures\u003c/h2\u003e \u003cp\u003eBefore the interaction, an informed consent and demographic information (age, gender, and living situation) was collected, followed by two standardized pre-interaction assessments. Loneliness was measured using the Three-Item UCLA Loneliness Scale (UCLA-3; (Hughes, Waite, Hawkley, \u0026amp; Ca-cioppo, 2004)), employing a four-point scale (1\u0026thinsp;=\u0026thinsp;Never to 4\u0026thinsp;=\u0026thinsp;Always), yielding a total score range\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eof 3\u0026ndash;12. Well-being was assessed using the WHO-5 Well-Being Index (Topp, \u0026Oslash;stergaard, S\u0026oslash;ndergaard, \u0026amp; Bech, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2015\u003c/span\u003e), consisting of five items rated on a six-point scale (0\u0026thinsp;=\u0026thinsp;At no time to 5\u0026thinsp;=\u0026thinsp;All of the time), converted to a percentage score (0-100) with higher values reflecting greater well-being.\u003c/p\u003e \u003cp\u003eFollowing the pre-surveys, participants engaged in a 15\u0026ndash;20 minute semi-structured conversation with \u003cem\u003eAmeca\u003c/em\u003e. The robot was configured to speak Dutch and follow a consistent script emphasizing warmth, active listening, and clarity, covering everyday topics such as family, food, hobbies, and reminiscence. A facilitator introduced the robot using a standardized script and only intervened to ensure comfort or address technical issues.\u003c/p\u003e \u003cp\u003eAfter each interaction, all participants completed a bespoke Ameca UX Questionnaire assessing five interaction domains\u0026mdash;empathy, trust, social comfort, comprehensibility, and enjoyment\u0026mdash;using five-point Likert scales (1\u0026thinsp;=\u0026thinsp;Strongly disagree to 5\u0026thinsp;=\u0026thinsp;Strongly agree), followed by several open-ended questions on perceived conversational naturalness and \u003cem\u003eAmeca\u003c/em\u003e\u0026rsquo;s physical appearance. Each session lasted approximately 30\u0026ndash;35 minutes in total, including surveys, interaction, and debriefing.\u003c/p\u003e \u003cp\u003eThe UX questionnaire was developed for this study to support feasibility assessment and exploratory user experience analysis. The domains of Empathy, Trust, and Social Comfort were theoretically grounded in established human\u0026ndash;robot interaction constructs, including social presence, perceived sup-port, and trust in automation (Hancock et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2011\u003c/span\u003e; Leite et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Spatola, Belletier, Chausse, \u0026amp; Augustinova, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Comprehensibility was informed by prior work on conversational usability and user experience in chatbot-based systems (F\u0026oslash;lstad \u0026amp; Brandtz\u0026aelig;g, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Study Design\u003c/h2\u003e \u003cp\u003eThis study followed a two-session, mixed design conducted over a two-week period. In \u003cb\u003ePhase 1\u003c/b\u003e, all participants (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;31) completed a baseline interaction with \u003cem\u003eAmeca\u003c/em\u003e, See Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e\u003c/p\u003e \u003cp\u003eOf these participants, 20 returned approximately one week later to complete \u003cb\u003ePhase 2\u003c/b\u003e, in which the same interaction protocol was administered after conversational refinements were implemented (see Section 3.4).\u003c/p\u003e \u003cp\u003eParticipants who completed Phase 2 did not differ substantially from the full Phase 1 sample in terms of age, gender distribution, or living situation and were therefore considered demographically comparable. Their mean age was 82.3 years (\u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;8.3); 15 were female and 5 male. Most participants resided in a group assisted living facility (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;15), while others lived in a care hotel (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;3) or at home (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2) or apartment (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1) in the neighborhood. Although the initial aim was a full within-subjects design, practical constraints (limited testing slots, overlapping residential activities, and health-related drop-out) resulted in a partial repeated-measures design, with only a subset of participants (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;20) completing both phases.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Intervention\u003c/h2\u003e \u003cp\u003eIn \u003cb\u003ePhase 1 (baseline)\u003c/b\u003e, \u003cem\u003eAmeca\u003c/em\u003e operated using the manufacturer\u0026rsquo;s default conversational prompt-ing configuration, without study-specific optimization. This configuration relied primarily on short, question-oriented exchanges and provided limited support for reflective listening, topic continuity, or socio-emotional signaling.\u003c/p\u003e \u003cp\u003eFollowing these initial sessions, qualitative observations and participant feedback highlighted op-portunities to improve the robot\u0026rsquo;s listening behavior, turn-taking, topic continuity, and socio-emotional expressiveness.\u003c/p\u003e \u003cp\u003eIn \u003cb\u003ePhase 2\u003c/b\u003e (approximately one week later), no changes were made to \u003cem\u003eAmeca\u003c/em\u003e\u0026rsquo;s hardware, embod-iment, sensing modalities, speech recognition pipeline, or underlying language model. Instead, the intervention was refined exclusively at the level of the LLM-based conversational framework through systematic prompt engineering within Ameca\u0026rsquo;s native conversational platform.\u003c/p\u003e \u003cp\u003eThese refinements included: (a) adjusted turn-taking delays to reduce interruptions and allow nat-ural pacing; (b) increased use of reflective listening strategies, such as paraphrasing and empathic acknowledgments; (c) improved topic continuity and conversational closure when exchanges stalled; and (d) enhanced socio-emotional backchannels, including light humor and small compliments to estab-lish rapport. In addition, the conversational prompts gently encouraged familiar, reminiscence-based themes (e.g., family, hobbies, food, and everyday life) that are commonly accessible to older adults.\u003c/p\u003e \u003cp\u003eA custom conversational persona was thus implemented, emphasizing a warm, attentive, and sup-portive interaction style suited to conversations with older adults. The conversational framework remained prompt-based rather than script-based, allowing generative flexibility while constraining in-teraction style toward more human-centered dialogue. No therapeutic content, coaching elements, or additional intervention components were introduced between phases.\u003c/p\u003e \u003cp\u003eTogether, these refinements aimed to enhance perceived empathy, social comfort, and conversational flow while maintaining consistency across sessions.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.6 Data Analysis\u003c/h2\u003e \u003cp\u003e \u003cb\u003eQuantitative analysis.\u003c/b\u003e Descriptive statistics (means, standard deviations, ranges) were computed for baseline psychosocial measures (UCLA-3, WHO-5) and for all UX questionnaire domains. Internal consistency was evaluated using Cronbach\u0026rsquo;s alpha (Cronbach, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e1951\u003c/span\u003e). Reliability coefficients at the domain level ranged from acceptable to good for most constructs. In contrast, the Comprehensibility and Clarity domain demonstrated comparatively lower internal consistency (see Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, N\u0026thinsp;=\u0026thinsp;31), suggesting possible multidimensionality or limited item coherence. Consequently, findings associated with this domain should be interpreted with caution.\u003c/p\u003e \u003cp\u003eFor participants who completed both sessions (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;20), paired-sample \u003cem\u003et\u003c/em\u003e-tests were conducted to descriptively compare UX ratings between Phase 1 and Phase 2. Given the limited sample size, the sequential two-phase design, and the absence of correction for multiple comparisons, inferential claims\u003c/p\u003e \u003cp\u003ewere not emphasized. Effect sizes (Cohen\u0026rsquo;s \u003cem\u003ed\u003c/em\u003e) were reported to provide an indication of the magnitude of observed differences, which were interpreted descriptively rather than causally.\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\u003eQuestionnaire constructs, items, internal consistency, and descriptive statistics (1\u0026ndash;5 Likert scale; Phase 1 baseline)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstruct\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eItem\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eItem text (English)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003eα\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSD\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eA. Overall\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eCon-\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eA1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eI found the conversation with Ameca pleas-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eversational\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eEn-\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eant.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ejoyment\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eA2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThe conversation felt natural.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eA3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eI felt good during the conversation with\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.79\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAmeca.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eB. Empathy and\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eB1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAmeca showed understanding for my situ-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eUnderstanding\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eation.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eB2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eI felt taken seriously by Ameca.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eB3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAmeca showed interest in what I said.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eB4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eI felt that Ameca was really listening.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.64\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eC. Social Com-\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eC1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eI felt comfortable during the conversation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003efortandCom-\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ewith Ameca.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003epanionship\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eC2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eI found Ameca pleasant company.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eC3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eI would like to talk to Ameca more often.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.70\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eD.Trustand\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eD1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eI trusted Ameca during the conversation.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCredibility\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eD2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAmeca seemed honest and reliable.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eD3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eI could be myself during the conversation.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eD4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eI felt that Ameca understood me well.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.83\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eE. Comprehensi-\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eE1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAmeca\u0026rsquo;s answers were clear and under-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003ebility and Clarity\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003estandable.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eE2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAmeca used language that was easy for me\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eto follow.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eE3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eI knew when it was my turn to speak.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eE4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eI knew when Ameca was speaking.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.51\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e\u003cb\u003eConversational metrics.\u003c/b\u003e Due to technical constraints, conversational transcripts were available only for Phase 2. These transcripts were analyzed to characterize interaction patterns under the optimized conversational configuration. Using a structured coding protocol implemented in \u003cem\u003eExcel\u003c/em\u003e, conversational metrics were extracted per participant, including total utterances, turn distribution between Ameca and the participant, number of questions and responses, mean response length, and proportion of short responses. All transcripts were reviewed independently by two researchers to verify consistency of segmentation and coding.\u003c/p\u003e \u003cp\u003e\u003cb\u003eQualitative analysis.\u003c/b\u003e Qualitative data consisted of participants\u0026rsquo; written responses to post-interaction open-ended questions addressing (1) their overall experience of the conversation with \u003cem\u003eAmeca\u003c/em\u003e, (2) as-pects they liked most, (3) aspects they liked least, (4) suggested improvements for future interactions, and (5) their impressions of \u003cem\u003eAmeca\u0026rsquo;s\u003c/em\u003e physical appearance. In total, 31 unique participants were en-rolled in the study. Of these, 28 provided qualitative responses in Phase 1, and 20 returned and completed the open-ended questions in Phase 2.\u003c/p\u003e \u003cp\u003eBecause the study followed a repeated-interaction design, some participants contributed responses in both phases. Qualitative responses were analysed separately per phase to explore changes in per-ceived strengths and limitations of the interaction. The analysis focused on identifying recurring themes and illustrative patterns within each phase rather than on formal within-subject comparisons.\u003c/p\u003e \u003cp\u003eThe qualitative analysis followed Braun and Clarke\u0026rsquo;s six-phase thematic analysis approach (Braun \u0026amp; Clarke, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2006\u003c/span\u003e). Researchers first familiarized themselves with the data through repeated reading. Ini-tial codes were generated inductively to capture salient experiential aspects such as conversational flow,\u003c/p\u003e \u003cp\u003eperceived listening, emotional engagement, personalization, and embodiment. Codes were iteratively refined and clustered into candidate themes based on patterns of shared meaning across participants. Themes were subsequently reviewed, defined, and refined through comparison across participants and across sessions, with attention to both convergence and divergence in experiences. Although the analysis was primarily inductive, the final thematic structure was interpreted in relation to the study\u0026rsquo;s\u003c/p\u003e \u003cp\u003eresearch questions (RQ3\u0026ndash;RQ4).\u003c/p\u003e \u003cp\u003eTwo researchers were involved in coding and interpretation. As members of the research team also contributed to the optimization of the conversational configuration, reflexive attention was paid to potential interpretive bias. Discrepancies were discussed until consensus was reached. Trustwor-thiness was supported through investigator triangulation, triangulation across qualitative responses, quantitative UX ratings, and conversational metrics, and the use of illustrative participant quotations to ground interpretations in the data.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4 Results","content":"\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\n \u003ch2\u003e4.1 Descriptive Statistics of Baseline Measures\u003c/h2\u003e\n \u003cp\u003eOverall, participants rated \u003cem\u003eAmeca\u003c/em\u003e positively across all UX domains, with mean scores ranging between\u003c/p\u003e\n \u003cp\u003e3.6 and 3.9, corresponding to the \u0026ldquo;agree\u0026rdquo; range of the Likert scale. Comprehensibility (\u003cem\u003eM\u003c/em\u003e\u0026thinsp;=\u0026thinsp;3.9) re-ceived the highest ratings, indicating that participants generally found \u003cem\u003eAmeca\u003c/em\u003e\u0026rsquo;s answers clear and easy to follow. Social comfort (\u003cem\u003eM\u003c/em\u003e\u0026thinsp;=\u0026thinsp;3.6) was rated slightly lower, suggesting that although partic-ipants enjoyed the interaction, feelings of companionship were somewhat less pronounced. Empathy (\u003cem\u003eM\u003c/em\u003e\u0026thinsp;=\u0026thinsp;3.8), trust (\u003cem\u003eM\u003c/em\u003e\u0026thinsp;=\u0026thinsp;3.8), and enjoyment (\u003cem\u003eM\u003c/em\u003e\u0026thinsp;=\u0026thinsp;3.7) fell in between, reflecting that \u003cem\u003eAmeca\u003c/em\u003e was generally perceived as an attentive and trustworthy conversational partner. The relatively small stan-dard deviations (0.6\u0026ndash;0.7) indicate moderate agreement among participants, although some individual variation remained.\u003c/p\u003e\n \u003cp\u003eTo contextualize these evaluations, baseline psychosocial measures were examined (see Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e). The mean UCLA-3 score was 4.1 (range\u0026thinsp;=\u0026thinsp;3\u0026ndash;9), indicating relatively low loneliness scores on this 3\u0026ndash;12 scale, albeit with noticeable variability across participants. As the UCLA-3 does not have a universally agreed clinical cut-off, these values are reported descriptively. In contrast, WHO-5 results suggested moderate levels of subjective well-being, with a mean percentage score of 48.6%. The wide range of WHO-5 scores (0\u0026ndash;100%) indicates substantial inter-individual differences, and several participants scored below the established clinical cut-off of 52%, commonly used to indicate reduced well-being (Topp et al., \u003cspan class=\"CitationRef\"\u003e2015\u003c/span\u003e).\u003c/p\u003e\n \u003cp\u003eTaken together, these baseline measures demonstrate considerable heterogeneity in psychosocial functioning, with a subset of participants reporting both lower well-being and higher loneliness.\u003c/p\u003e\n \u003cp\u003eTable 2 summarizes the descriptive statistics for the UCLA-3 Loneliness Scale and the WHO-5 Well-Being Index.\u003c/p\u003eTable 2: Descriptive statistics for UCLA-3 and WHO-5 scores (Phase 1 baseline).\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\n \u003ch2\u003e4.2 Correlations Between UX Constructs, Loneliness, and Well-Being\u003c/h2\u003e\n \u003cp\u003eAs shown in Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e, correlations between baseline psychosocial measures and UX constructs revealed a differentiated pattern.\u003c/p\u003e\n \u003cp\u003eFor loneliness (UCLA-3), higher scores were negatively associated with all UX domains, although most correlations were small to moderate and did not reach statistical significance. A statistically significant and comparatively stronger association was observed for comprehensibility and clarity (\u003cem\u003er\u003c/em\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.53, \u003cem\u003ep\u003c/em\u003e = .004), indicating that participants reporting higher loneliness tended to perceive the interaction as less clear and predictable. This represents a moderate-to-strong negative correlation and was the only domain reaching statistical significance.\u003c/p\u003e\n \u003cp\u003eFor well-being (WHO-5), all associations were positive, suggesting that higher subjective well-being was consistently related to more favorable UX evaluations. However, none of these correlations reached conventional levels of statistical significance.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u003cbr\u003e\u003c/div\u003e\n \u003ctable id=\"Tab4\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003ePearson correlations between UX constructs and baseline measures of loneliness (UCLA-3) and well-being (WHO-5), \u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;31.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003eConstruct\u003c/p\u003e\n \u003c/th\u003e\n \u003cth colspan=\"2\" align=\"left\"\u003e\n \u003cp\u003eUCLA-3\u003c/p\u003e\n \u003c/th\u003e\n \u003cth colspan=\"2\" align=\"left\"\u003e\n \u003cp\u003eWHO-5\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003er\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003er\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eA. General enjoyment\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026minus;\u003c/em\u003e0.18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e.36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e.09\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eB. Empathy and understanding\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026minus;\u003c/em\u003e0.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e.16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eC. Social comfort\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026minus;\u003c/em\u003e0.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e.40\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eD. Trust and credibility\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026minus;\u003c/em\u003e0.24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.34\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e.08\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eE. Comprehensibility and clarity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026minus;\u003c/em\u003e\u003cstrong\u003e0\u003c/strong\u003e.\u003cstrong\u003e53\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e.004\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e.17\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\n \u003ch2\u003e4.3 Comparative Analyses of Two Phases\u003c/h2\u003e\n \u003cp\u003eIn line with RQ2, the study investigates differences in participants\u0026rsquo; evaluations of \u003cem\u003eAmeca\u003c/em\u003e across the two interaction sessions.\u003c/p\u003e\n \u003cp\u003eResults are presented in Table \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e and indicate that across both sessions, participants rated \u003cem\u003eAmeca\u003c/em\u003e consistently in the \u0026ldquo;agree\u0026rdquo; range (means 3.6\u0026ndash;4.0 on a 5-point Likert scale). None of the paired com-parisons reached statistical significance (\u003cem\u003ep\u0026thinsp;\u0026gt;\u003c/em\u003e\u0026thinsp;0.05), and all effect sizes were small (\u003cem\u003e|d| \u0026lt;\u003c/em\u003e 0.40).\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u003cbr\u003e\u003c/div\u003e\n \u003ctable id=\"Tab5\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eComparison of participants\u0026rsquo; evaluations between Phase 1 and Phase 2 (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;20).\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eFactor\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePhase 1 mean\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePhase 2 mean\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003et\u003c/em\u003e(19)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eCohen\u0026rsquo;s \u003cem\u003ed\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eA. General enjoyment\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-1.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.218\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-0.29\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eB. Empathy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-0.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.669\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-0.10\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eC. Social comfort\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-1.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.126\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-0.36\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eD. Trust\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.901\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eE. Clarity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.579\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.13\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eAlthough mean scores for enjoyment and social comfort were numerically slightly higher in Phase 2, these differences should be interpreted as descriptive only. Given the limited sample size (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;20), the absence of adjustment for multiple comparisons, and the sequential two-phase design, no inferential claims can be made. Instead, these results indicate overall stability of user experience ratings across sessions, supporting the feasibility and acceptability of the interaction framework.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\n \u003ch2\u003e4.4 Qualitative Analysis of Open-Ended Responses\u003c/h2\u003e\n \u003cp\u003eQualitative responses provided deeper insight into how participants experienced the interaction with\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eAmeca\u003c/em\u003e beyond the structured questionnaire measures.\u003c/p\u003e\n \u003cp\u003eIn \u003cstrong\u003ePhase 1\u003c/strong\u003e, 28 participants generally described the conversation as pleasant and novel, often emphasizing that it was \u0026ldquo;nice to have someone to talk to\u0026rdquo; and that the exchange felt \u0026ldquo;interesting\u0026rdquo; or \u0026ldquo;enriching.\u0026rdquo; Nearly all participants provided at least one positive remark about the interaction. Frequently mentioned strengths included \u003cem\u003eAmeca\u0026rsquo;s\u003c/em\u003e clarity of speech and attentive demeanor.\u003c/p\u003e\n \u003cp\u003eAt the same time, the majority of participants formulated at least one suggestion for improvement. Approximately 5 participants explicitly referred to conversational interruptions or turn-taking issues, noting that the robot \u0026ldquo;interrupted too often\u0026rdquo; or \u0026ldquo;talked over\u0026rdquo; them. Others described the interaction as somewhat limited or experimental and expressed a desire for greater personalization and contextual continuity, suggesting that \u003cem\u003eAmeca\u003c/em\u003e should \u0026ldquo;remember what I said earlier.\u0026rdquo; These responses indicate that conversational flow and memory persistence were central to participants\u0026rsquo; expectations of natural interaction.\u003c/p\u003e\n \u003cp\u003eIn \u003cstrong\u003ePhase 2\u003c/strong\u003e, feedback of 20 participants remained largely positive and appeared slightly more favorable overall, with fewer explicit references to interruption-related issues. Participants again em-phasized clarity, engagement, and the pleasant experience of companionship. Nevertheless, suggestions for deeper conversational content and improved contextual awareness remained present across both sessions.\u003c/p\u003e\n \u003cp\u003eImportantly, physical appearance emerged as an additional and distinct component of the user ex-perience. When explicitly invited to reflect on \u003cem\u003eAmeca\u0026rsquo;s\u003c/em\u003e embodiment, 56% of participants described her appearance as pleasant or lifelike, 38% as neutral, and only 6% expressed negative impressions. Posi-tive comments frequently referred to her friendly facial expression, realistic movements, and feminine design, with several participants noting that she was \u0026ldquo;pleasant to look at\u0026rdquo; or \u0026ldquo;almost human.\u0026rdquo;\u003c/p\u003e\n \u003cp\u003eOnly a small minority perceived her appearance as somewhat artificial or lacking warmth, suggest-ing that \u0026ldquo;she could smile more\u0026rdquo; or that she \u0026ldquo;still looks like a robot.\u0026rdquo;\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e\n \u003ch2\u003e4.5 Conversational Metrics and Thematic Analysis\u003c/h2\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003eTable 5 presents aggregated descriptive statistics across 20 participants in \u003cstrong\u003ePhase 2\u003c/strong\u003e. Only the converation transcripts (\u003cem\u003eN\u0026nbsp;\u003c/em\u003e= 20) from \u003cstrong\u003ePhase 2\u0026nbsp;\u003c/strong\u003ewere analyzed, as the technical setup during the first round did not yet allow for reliable recording and data extraction. In average, \u003cem\u003eAmeca\u0026nbsp;\u003c/em\u003eproduced 49 turns and participants 54 turns, with approximately 30 questions posed by \u003cem\u003eAmeca\u0026nbsp;\u003c/em\u003eand 46 responses provided by participants. The mean length of participant answers was about 11 words, but ranged between 6 and 20 words. The proportion of short answers averaged 23% but ranged from only 2% in highly engaged participants to more than half of all responses in low-engagement cases.\u0026nbsp;\u003c/div\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u003cbr\u003e\u003c/div\u003e\n \u003ctable id=\"Tab6\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eDescriptive statistics of conversational metrics across all participants of \u003cstrong\u003ePhase 2\u003c/strong\u003e\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMetric\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMean\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSD\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMin\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMax\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTotal utterances\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e102.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e34.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e169\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTurns by Ameca\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e48.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e16.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e80\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTurns by participant\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e53.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e19.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e23\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e96\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eQuestions by Ameca\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e29.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAnswers by participant\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e45.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e18.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e76\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRepairs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAverage answer length (words)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e11.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e19.72\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eShort answers proportion\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.53\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e: Descriptive statistics of conversational metrics across all participants of \u003cstrong\u003ePhase 2\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eThematic analysis was conducted in parallel with descriptive conversational metrics to contextualize observed interaction patterns (Table \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e). While aggregated metrics indicated relatively balanced turn-taking across participants (mean turns by \u003cem\u003eAmeca\u003c/em\u003e\u0026thinsp;=\u0026thinsp;48.6; participants\u0026thinsp;=\u0026thinsp;53.6), substantial inter-individual variability was observed in response length and elaboration. In particular, the proportion of short responses ranged from 0.02 to 0.53, indicating that in low-engagement cases, more than half of participant contributions consisted of minimal or non-elaborative answers.\u003c/p\u003e\n \u003cp\u003eAgainst this quantitative backdrop, thematic analysis focused on identifying patterns of meaning that could explain variability in conversational engagement. Three analytically distinct themes were identified.\u003c/p\u003e\n \u003cp\u003eThe first theme concerned \u003cem\u003eautobiographical engagement as a driver of conversational depth\u003c/em\u003e. Par-ticipants who engaged with autobiographical or emotionally salient topics\u0026mdash;such as travel experiences, family relationships, artistic interests, or bereavement\u0026mdash;tended to produce longer responses and fewer short answers. These interactions were characterized by narrative continuity and reciprocal elabora-tion, suggesting that conversational depth emerged when participants framed the interaction as an opportunity for personal storytelling rather than question answering. This pattern aligns with the higher mean response lengths observed in high-engagement cases.\u003c/p\u003e\n \u003cp\u003eA second theme captured \u003cem\u003einteractional asymmetry and question-driven structure\u003c/em\u003e. Across tran-scripts, \u003cem\u003eAmeca\u003c/em\u003e posed an average of 29.6 questions per interaction, reflecting a predominantly inter-rogative conversational structure. In cases where participants provided brief or hesitant answers, this structure resulted in rapid topic cycling and limited conversational buildup. These low-engagement interactions corresponded to higher proportions of short responses (up to 53%), indicating that con-versational depth was constrained not by turn-taking balance per se, but by limited elaboration within participant turns.\u003c/p\u003e\n \u003cp\u003eThe third theme addressed \u003cem\u003eturn-taking sensitivity and perceived listening\u003c/em\u003e. Despite balanced turn counts, several participants commented on interruptions or premature responses following pauses. These qualitative observations suggest that conversational smoothness cannot be inferred from turn distribution alone, but is sensitive to micro-level timing and pause interpretation. In several instances, participants explicitly requested additional silence or clarified that they were still speaking, highlighting the importance of temporal alignment for perceived listening and social comfort.\u003c/p\u003e\n \u003cp\u003eTaken together, these themes indicate that conversational engagement in LLM-enabled humanoid interaction is co-constructed and heterogeneous. Quantitative metrics reveal overall feasibility and structural balance, while qualitative themes elucidate why some interactions developed into extended, meaningful exchanges whereas others remained superficial. Importantly, conversational depth was not uniformly produced by the optimized prompting configuration, but depended on participants\u0026rsquo; willingness and capacity to elaborate, as well as on fine-grained interactional features such as timing, continuity, and responsiveness.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"5 Discussion","content":"\u003cp\u003eThe present study examined how older adults experienced short, structured conversations with the humanoid robot \u003cem\u003eAmeca\u003c/em\u003e in a residential-care setting, combining subjective user experience (RQ1\u0026ndash;RQ2), transcript-based interaction characteristics available for Phase 2 (RQ3), and qualitative reflections (RQ4).\u003c/p\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e5.1 User Experience and Acceptance (RQ1\u0026ndash;RQ2)\u003c/h2\u003e \u003cp\u003e\u003cb\u003eRegarding RQ1 and RQ2\u003c/b\u003e, participants\u0026rsquo; perceptions of \u003cem\u003eAmeca\u003c/em\u003e were generally positive across en-joyment, perceived empathy, social comfort, trust, and comprehensibility. Among participants who completed both sessions (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;20), UX ratings remained largely stable, with only small descriptive differences between Phase 1 and Phase 2. Given the sequential two-phase design and the concur-rent prompt-level refinements, these session-level differences cannot be causally attributed to either repeated exposure (familiarity) or to conversational optimization. Rather than implying improvement effects, the overall stability suggests that initial acceptability was maintained at re-contact, supporting feasibility in a residential-care setting. This interpretation aligns with prior work on socially assistive robots indicating that perceived empathy and trust tend to depend more on conversational coherence and affective cues than on familiarity alone (Broadbent, 2016).\u003c/p\u003e \u003cp\u003eImportantly, the data and participants\u0026rsquo; comments nonetheless underscore that perceived inter-action quality is sensitive to the \u003cem\u003econfiguration layer\u003c/em\u003e on top of the underlying LLM. Even without changes to hardware or the base model, interaction design decisions\u0026mdash;including prompt and context engineering, explicit turn-taking rules, topic steering, and socio-emotional backchannels\u0026mdash;appear to shape perceived naturalness, pacing, and social comfort. This design-relevant insight is consistent with recent feasibility work on LLM-augmented robots in healthcare contexts, which similarly empha-sizes interaction quality, clarity, and pacing as key drivers of user acceptance (van \u0026rsquo;t Klooster et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2026\u003c/span\u003e). Likewise, user-centered LLM architectures for embodied robots highlight that conversational scaffolding and lightweight memory mechanisms can enhance perceived trust and social presence, while noting persistent challenges around robustness and latency (Pinto-Bernal et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFinally, correlations between psychosocial baselines and UX evaluations point to meaningful hetero-geneity in how older adults experience humanoid interaction. Participants reporting higher loneliness tended to evaluate the interaction less favorably, most notably in comprehensibility. This suggests that baseline social vulnerability may be associated with lower perceived predictability or shared un-derstanding in robot dialogue; however, this result should be interpreted cautiously given the lower internal consistency of the comprehensibility/clarity domain in this sample.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Conversational Characteristics Under the Phase 2 Configuration (RQ3)\u003c/h2\u003e \u003cp\u003e\u003cb\u003eIn relation to RQ3\u003c/b\u003e, transcript-based conversational metrics and thematic analyses provide insight into interaction patterns observed under the Phase 2 configuration. At an aggregate level, turn-taking was relatively balanced, suggesting reciprocal participation. At the same time, the interaction remained predominantly question-driven: \u003cem\u003eAmeca\u003c/em\u003e posed many questions and participants primarily responded, reflecting a scaffolded, interview-like interaction structure rather than fully bidirectional free conversation.\u003c/p\u003e \u003cp\u003eCrucially, metrics revealed substantial inter-individual variability in engagement (e.g., response length and short-answer proportion), which was echoed in the thematic interpretation. Participants who shifted toward autobiographical or emotionally salient topics (e.g., travel memories, family rela-tionships, bereavement) tended to provide longer responses and fewer minimal answers, indicating nar-rative continuity and reciprocal elaboration. In lower-engagement cases, brief answers combined with rapid topic cycling constrained conversational build-up. These patterns are consistent with human\u0026ndash;robot interaction accounts emphasizing that conversational depth is co-constructed rather than gener-ated by the robot alone (Karaaziz, Can, \u0026amp; Keskinda˘g, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). The link between richer responses and personal themes further aligns with empathy and social presence perspectives suggesting that emotional salience can facilitate deeper conversational involvement (Morgante, Susinna, Culicetto, Quartarone, \u0026amp; Lo Buono, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Qualitative Reflections, Speech Robustness, and Embodiment (RQ4)\u003c/h2\u003e \u003cp\u003e\u003cb\u003eRQ4\u003c/b\u003e contextualizes these patterns through participants\u0026rsquo; reflections. Across sessions, participants fre-quently described the interaction as \u0026ldquo;pleasant,\u0026rdquo; \u0026ldquo;interesting,\u0026rdquo; and \u0026ldquo;clear,\u0026rdquo; often highlighting compan-ionship and curiosity as motives for engagement. At the same time, three recurring improvement needs were consistently reported: (1) limited personalization and memory, (2) abrupt topic shifts, and (3) restricted emotional depth. These concerns map directly onto known constraints of many current LLM-driven social robots, particularly the absence of cross-session memory and adaptive context modeling. They are also consistent with prior work indicating that memory continuity and adaptive pacing are central to sustaining social presence, naturalness, and trust in embodied LLM systems (Pinto-Bernal et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Verhelst \u0026amp; Belpaeme, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e Participants also occasionally requested repetition or noted non-response when they spoke softly or used dialect. These remarks align with known ASR challenges for older adults (e.g., reduced speech volume, slower articulation, dialectal variation). While the Whisper-based pipeline generally enabled fluent interaction, the findings suggest that future deployments could benefit from confidence-based clarification strategies, targeted robustness improvements for elderly speech, and dialect-aware handling in multilingual contexts such as Dutch (Flemish).\u003c/p\u003e \u003cp\u003ePerceptions of \u003cem\u003eAmeca\u003c/em\u003e\u0026rsquo;s embodiment further shaped acceptance. Most participants described the robot as pleasant or lifelike, appreciating facial expressiveness and overall approachability. This aligns with accounts of moderate human-likeness within the uncanny-valley continuum (Mara et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). A\u003c/p\u003e \u003cp\u003eminority still perceived the robot as insufficiently warm (e.g., \u0026ldquo;could smile more\u0026rdquo;), underscoring that perceived emotional authenticity depends not only on language generation but also on fine-grained embodied cues such as gaze timing, facial affect, and turn-taking sensitivity (Kwak, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2014\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003e5.4 Synthesis and Implications\u003c/h2\u003e \u003cp\u003eTaken together, these findings reinforce that humanoid robots such as \u003cem\u003eAmeca\u003c/em\u003e can be feasible and generally well accepted for brief conversational encounters in residential care, while highlighting de-sign priorities for achieving more sustained, relationship-like interaction. Importantly, the present results should be interpreted as feasibility and interaction-quality evidence rather than evidence of improvement across sessions or clinical effectiveness, as the sequential design does not disentangle familiarity from prompt-level refinement. In line with recent feasibility studies of LLM-augmented robots in healthcare, the study suggests that acceptance depends strongly on interaction quality, clar-ity, and pacing (van \u0026rsquo;t Klooster et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2026\u003c/span\u003e). At the same time, consistent with design-oriented work on embodied LLM systems, sustained naturalness remains contingent on advances in personal-ization, memory continuity, and robust turn-taking management (Pinto-Bernal et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Future systems may therefore prioritize (i) improved pause detection and turn-taking sensitivity, (ii) robust clarification and repair strategies for older-adult speech and dialect, and (iii) retrieval-based memory mechanisms to support continuity and personalization across sessions.\u003c/p\u003e \u003c/div\u003e"},{"header":"6 Limitations and Future Directions","content":"\u003cp\u003eSeveral limitations should be acknowledged. First, the sample size was modest, particularly for the repeated-measures component (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;20), limiting statistical power and generalizability. Second, ses-sions were brief (15\u0026ndash;20 minutes), capturing short-term impressions rather than sustained relational dynamics. Third, the study lacked systematic logging of technical performance metrics (e.g., response latency, ASR confidence/accuracy), constraining deeper technical interpretation.\u003c/p\u003e \u003cp\u003eA key limitation concerns the sequential two-phase design in which the baseline configuration was assessed in Phase 1 and the refined configuration in Phase 2. Because conditions were not counterbal-anced, session differences cannot be causally attributed to conversational refinement versus familiar-ity/repeated exposure; accordingly, session-level comparisons should be interpreted descriptively.\u003c/p\u003e \u003cp\u003eImportantly, the present study should not be interpreted as a loneliness intervention. Although loneliness was assessed using the UCLA-3 scale, this measure was included to characterize baseline psychosocial context and to explore associations with perceived interaction quality, rather than to evaluate change or long-term effects. The brief two-session design and short interaction duration preclude causal conclusions regarding loneliness outcomes. Accordingly, references to long-term impact or intervention effects have been avoided, and findings are framed in terms of feasibility, acceptability, and short-term interaction experiences.\u003c/p\u003e \u003cp\u003eFinally, the conversational model remained constrained by limited personalization and lack of cross-session memory. Although Phase 2 introduced prompt-level refinements, \u003cem\u003eAmeca\u003c/em\u003e could not recall prior exchanges or adapt to individual preferences beyond the immediate context. Retrieval-augmented generation and lightweight memory continuity mechanisms represent promising avenues to address this constraint and to enable more coherent, relationship-like interaction over time.\u003c/p\u003e \u003cp\u003eIn parallel with these research directions, real-world deployment feedback can also inform iterative platform development. Where applicable, future work should document and evaluate interaction-level updates (e.g., voice quality, pause detection, facial synchronization) with systematic performance logging and controlled study designs. Longitudinal protocols and randomized/counterbalanced designs will be especially valuable to evaluate sustained engagement and to separate configuration effects from familiarity and learning effects.\u003c/p\u003e"},{"header":"7 Conclusion","content":"\u003cp\u003eThis two-session feasibility study indicates that short, structured conversations with an LLM-enabled humanoid robot (\u003cem\u003eAmeca\u003c/em\u003e) are technically feasible and generally acceptable for older adults in a residential-care context. Participants reported overall positive UX ratings, and Phase 2 transcripts showed re-\u003c/p\u003e \u003cp\u003e ciprocal participation with balanced turn-taking but a predominantly question\u0026ndash;answer interaction structure.\u003c/p\u003e \u003cp\u003eA central contribution is that interaction quality in embodied LLM systems is shaped not only by the underlying language model, but also by the conversational configuration layer (prompt and context engineering), which influences pacing, perceived listening, and socio-emotional backchannels without requiring changes to hardware or the base model. At the same time, qualitative feedback and interaction patterns point to persistent challenges, including limited personalization and memory continuity, occasional turn-taking timing issues, and the need for robust clarification strategies for older-adult speech and dialect.\u003c/p\u003e \u003cp\u003eTogether, these findings support further research using counterbalanced or randomized designs and longer-term deployments to assess how improvements in contingent responsiveness, personalization, and embodied warmth translate into sustained engagement and potential psychosocial benefit in care settings.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eEthical approval\u003c/h2\u003e\n\u003cp\u003eThis\u0026nbsp;study\u0026nbsp;was\u0026nbsp;conducted\u0026nbsp;in\u0026nbsp;accordance\u0026nbsp;with\u0026nbsp;the\u0026nbsp;principles\u0026nbsp;of\u0026nbsp;the\u0026nbsp;Declaration\u0026nbsp;of\u0026nbsp;Helsinki.\u0026nbsp;The\u0026nbsp;research protocol\u0026nbsp;was\u0026nbsp;reviewed\u0026nbsp;and approved\u0026nbsp;in accordance with\u0026nbsp;the institutional\u0026nbsp;ethical procedures\u0026nbsp;of Howest\u0026nbsp;University\u0026nbsp;of\u0026nbsp;Applied\u0026nbsp;Sciences.\u0026nbsp;Given\u0026nbsp;the\u0026nbsp;non-invasive\u0026nbsp;nature\u0026nbsp;of\u0026nbsp;the\u0026nbsp;study\u0026nbsp;and\u0026nbsp;the\u0026nbsp;absence\u0026nbsp;of\u0026nbsp;medical or therapeutic intervention, no formal medical ethics committee approval was required under applicable\u0026nbsp;national regulations.\u003c/p\u003e\n\u003ch2\u003eData\u0026nbsp;Availability\u0026nbsp;Statement\u003c/h2\u003e\n\u003cp\u003eDue to privacy considerations, the raw physiological and behavioral data from participants are not publicly available.\u0026nbsp;Aggregated results and analysis scripts can be obtained from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003ch2\u003eConsent to participate\u003c/h2\u003e\n\u003cp\u003eAll\u0026nbsp;participants\u0026nbsp;received\u0026nbsp;verbal\u0026nbsp;and\u0026nbsp;written\u0026nbsp;information\u0026nbsp;about\u0026nbsp;the\u0026nbsp;study\u0026nbsp;prior\u0026nbsp;to\u0026nbsp;participation.\u0026nbsp;Written informed\u0026nbsp;consent\u0026nbsp;was\u0026nbsp;obtained\u0026nbsp;from\u0026nbsp;all\u0026nbsp;participants\u0026nbsp;before\u0026nbsp;data\u0026nbsp;collection.\u0026nbsp;Participants\u0026nbsp;were\u0026nbsp;informed\u0026nbsp;that participation was voluntary and that they could withdraw at any time without consequences.\u003c/p\u003e\n\u003ch2\u003eConsent to publish\u003c/h2\u003e\n\u003cp\u003eAll participants provided explicit consent for the anonymized use of their data for scientific publica-tion.\u0026nbsp;Any visual materials included in this manuscript were anonymized to prevent identification of individual participants.\u003c/p\u003e\n\u003ch2\u003eDeclaration\u0026nbsp;of interests\u003c/h2\u003e\n\u003cp\u003eThe\u0026nbsp;authors\u0026nbsp;declare\u0026nbsp;that\u0026nbsp;they\u0026nbsp;have\u0026nbsp;no\u0026nbsp;known\u0026nbsp;competing\u0026nbsp;financial\u0026nbsp;interests\u0026nbsp;or\u0026nbsp;personal\u0026nbsp;relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\n\u003ch2\u003eFunding\u003c/h2\u003e\n\u003cp\u003eThis research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n\u003ch2\u003eDeclaration\u0026nbsp;of\u0026nbsp;generative\u0026nbsp;AI\u0026nbsp;and\u0026nbsp;AI-assisted\u0026nbsp;technologies\u0026nbsp;in\u0026nbsp;the\u0026nbsp;writing\u0026nbsp;process\u003c/h2\u003e\n\u003cp\u003eDuring the preparation of this work the authors sometimes used ChatGPT 5 in order to rephrase smaller parts of the text for clarity or to correct language errors.\u0026nbsp;The authors always reviewed and edited the content as needed and take full responsibility for the content of the published article.\u003c/p\u003e\n\u003ch2\u003eCollaboration\u0026nbsp;Declaration\u003c/h2\u003e\n\u003cp\u003eThis study was conducted in collaboration with Yin Oei (CEO at Living Tomorrow group) and Lisa Vanryckegem (i-mens).\u0026nbsp;Living Tomorrow facilitated access to the humanoid robot platform and the experimental setting, while i-mens supported participant recruitment and on-site coordination within\u0026nbsp;the residential care context.\u0026nbsp;Technical platform support was provided by Engineered Arts.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBanks, M. R., Willoughby, L. M., \u0026amp; Banks, W. A. (2008). Animal-assisted therapy and loneliness in nursing homes: Use of robotic versus living dogs. \u003cem\u003eJournal of the American Medical Directors Association\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e (3), 173\u0026ndash;177. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jamda.2007.11.007\u003c/span\u003e\u003cspan address=\"10.1016/j.jamda.2007.11.007\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoamah, S. A., Weldrick, R., Lee, T.-S. J., \u0026amp; Taylor, N. (2021). Social isolation among older adults in long-term care: A scoping review. \u003cem\u003eJournal of Aging and Health\u003c/em\u003e, \u003cem\u003e33\u003c/em\u003e (7\u0026ndash;8), 618\u0026ndash;632. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/08982643211004174\u003c/span\u003e\u003cspan address=\"10.1177/08982643211004174\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBraun, V., \u0026amp; Clarke, V. (2006). Using thematic analysis in psychology. \u003cem\u003eQualitative Research in Psychology\u003c/em\u003e, \u003cem\u003e3\u003c/em\u003e (2), 77\u0026ndash;101. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1191/1478088706qp063oa\u003c/span\u003e\u003cspan address=\"10.1191/1478088706qp063oa\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBroadbent, E. (2016). Interactions with robots: The truths we reveal about ourselves. \u003cem\u003eAnnual Review of Psychology\u003c/em\u003e, \u003cem\u003e68\u003c/em\u003e (1), 627\u0026ndash;652. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1146/annurev-psych-010416-043958\u003c/span\u003e\u003cspan address=\"10.1146/annurev-psych-010416-043958\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e Broadbent, E., Loveys, K., Ilan, G., Chen, G., Chilukuri, M. M., Boardman, S. G.,. .. Skuler, D. (2024). Elliq, an ai-driven social robot to alleviate loneliness: Progress and lessons learned. \u003cem\u003eThe Journal of Aging Research \u0026amp; Lifestyle\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e, 22\u0026ndash;28. doi: https://doi.org/10.14283/jarlife.2024.2\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCacioppo, J. T., \u0026amp; Cacioppo, S. (2014). Social relationships and health: The toxic effects of perceived social isolation. \u003cem\u003eSocial and Personality Psychology Compass\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e (2), 58\u0026ndash;72. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/spc3\u003c/span\u003e\u003cspan address=\"10.1111/spc3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e.12087\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCampbell IV, W. A., Chick, J. F. B., Shin, D., \u0026amp; Makary, M. S. (2024). Understanding chatgpt for evidence-based utilization in interventional radiology. \u003cem\u003eClinical Imaging\u003c/em\u003e, \u003cem\u003e108\u003c/em\u003e, 110098. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.clinimag.2024.110098\u003c/span\u003e\u003cspan address=\"10.1016/j.clinimag.2024.110098\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCarson, D. B., Johansson, A. B., Schaumberg, M., \u0026amp; Hurtig, A.-K. (2024). Addressing the workforce crisis in rural social care: A scoping review. \u003cem\u003eThe International Journal of Health Planning and Management\u003c/em\u003e, \u003cem\u003e39\u003c/em\u003e (3), 806\u0026ndash;823. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/hpm.3774\u003c/span\u003e\u003cspan address=\"10.1002/hpm.3774\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCoghlan, S., \u0026amp; Waycott, J. (2022). \u003cem\u003eCan robots really be companions for older adults?\u003c/em\u003e Retrieved from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://research.unimelb.edu.au/strengths/updates/news/can-robots-really-be-companions-for-older-adults\u003c/span\u003e\u003cspan address=\"https://research.unimelb.edu.au/strengths/updates/news/can-robots-really-be-companions-for-older-adults\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCourtin, E., \u0026amp; Knapp, M. (2017). Social isolation, loneliness and health in old age: a scoping review.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e\u003cem\u003eHealth \u0026amp; Social Care in the Community\u003c/em\u003e, \u003cem\u003e25\u003c/em\u003e (3), 799\u0026ndash;812. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/hsc.12311\u003c/span\u003e\u003cspan address=\"10.1111/hsc.12311\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCronbach, L. J. (1951). Coefficient alpha and the internal structure of tests. \u003cem\u003ePsychometrika\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e (3), 297\u0026ndash;334.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDe Jong-Gierveld, J., \u0026amp; Van Tilburg, T. G. (1999). \u003cem\u003eManual of the loneliness scale\u003c/em\u003e. Methoden en Technieken.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDonovan, N. J., \u0026amp; Blazer, D. G. (2020). Social isolation and loneliness in older adults: Review and commentary of a national academies report. \u003cem\u003eThe American Journal of Geriatric Psychiatry\u003c/em\u003e, \u003cem\u003e28\u003c/em\u003e (12), 1233\u0026ndash;1244. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jagp.2020.08.005\u003c/span\u003e\u003cspan address=\"10.1016/j.jagp.2020.08.005\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFiolka, A. K., Donnermann, M., \u0026amp; Lugrin, B. (2024). Investigating the mere exposure effect in relation to perceived eeriness and humaneness of a social robot., 453\u0026ndash;457.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eF\u0026oslash;lstad, A., \u0026amp; Brandtz\u0026aelig;g, P. B. (2017). Chatbots and the new world of hci. \u003cem\u003eInteractions\u003c/em\u003e, \u003cem\u003e24\u003c/em\u003e (4), 38\u0026ndash;42. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1145/3085558\u003c/span\u003e\u003cspan address=\"10.1145/3085558\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGardiner, C., Laud, P., Heaton, T., \u0026amp; Gott, M. (2020). What is the prevalence of loneliness amongst older people living in residential and nursing care homes? a systematic review and meta-analysis. \u003cem\u003eAge and Ageing\u003c/em\u003e, \u003cem\u003e49\u003c/em\u003e (5), 748\u0026ndash;757. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/ageing/afaa049\u003c/span\u003e\u003cspan address=\"10.1093/ageing/afaa049\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHancock, P. A., Billings, D. R., Schaefer, K. E., Chen, J. Y., de Visser, E. J., \u0026amp; Parasuraman, R. (2011). A meta-analysis of factors affecting trust in human-robot interaction. \u003cem\u003eHuman Factors\u003c/em\u003e, \u003cem\u003e53\u003c/em\u003e (5), 517\u0026ndash;527. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/0018720811417254\u003c/span\u003e\u003cspan address=\"10.1177/0018720811417254\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHughes, M. E., Waite, L. J., Hawkley, L. C., \u0026amp; Cacioppo, J. T. (2004). A short scale for measuring loneliness in large surveys. \u003cem\u003eResearch on Aging\u003c/em\u003e, \u003cem\u003e26\u003c/em\u003e (6), 655\u0026ndash;672. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/0164027504268574\u003c/span\u003e\u003cspan address=\"10.1177/0164027504268574\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHung, L., Liu, C., Woldum, E., Au-Yeung, A., Berndt, A., Wallsworth, C.,.. . Chaudhury, H. (2019). The benefits of and barriers to using a social robot paro in care settings: A scoping review. \u003cem\u003eBMC Geriatrics\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e (1), 232. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s12877-019-1244-6\u003c/span\u003e\u003cspan address=\"10.1186/s12877-019-1244-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKaraaziz, M., Can, G., \u0026amp; Keskinda˘g, B. (2017). Examining empathy as a communication technique in the context of gender and cultural differences: A review study. \u003cem\u003eGlobal Journal of Psychology Research: New Trends and Issues\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e (1), 20\u0026ndash;27. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.18844/gjpr.v7i1.401\u003c/span\u003e\u003cspan address=\"10.18844/gjpr.v7i1.401\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKwak, S. S. (2014). The impact of robot appearance types on social interaction with a robot and service evaluation of a robot. \u003cem\u003eArchives of Design Research\u003c/em\u003e, \u003cem\u003e27\u003c/em\u003e (2), 81\u0026ndash;93. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.15187/adr.2014.05.110.2.81\u003c/span\u003e\u003cspan address=\"10.15187/adr.2014.05.110.2.81\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLal, S. (2025). Analyzing the impact of social isolation on aging: Health, mental state and cognitive functions. \u003cem\u003eCritical Debates in Humanities, Science and Global Justice\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLapane, K. L., et al. (2022). Health effects of loneliness and social isolation in older adults living in congregate long-term care settings: A review. \u003cem\u003eJournal of Aging Social Policy\u003c/em\u003e. doi: 10.1080/ 08959420.2022.2046767\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeite, I., Castellano, G., Pereira, A., Martinho, C., Paiva, A., \u0026amp; McOwan, P. W. (2014). Empathic robots for long-term interaction. \u003cem\u003eInternational Journal of Social Robotics\u003c/em\u003e, \u003cem\u003e6\u003c/em\u003e (3), 329\u0026ndash;341. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s12369-014-0227-1\u003c/span\u003e\u003cspan address=\"10.1007/s12369-014-0227-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMara, M., Appel, M., \u0026amp; Gnambs, T. (2022). Human-like robots and the uncanny valley. \u003cem\u003eZeitschrift fu\u0026uml;r Psychologie\u003c/em\u003e. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1027/2151-2604/a000486\u003c/span\u003e\u003cspan address=\"10.1027/2151-2604/a000486\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMorgante, E., Susinna, C., Culicetto, L., Quartarone, A., \u0026amp; Lo Buono, V. (2024). Is it possible for people to develop a sense of empathy toward humanoid robots and establish meaningful relationships with them? \u003cem\u003eFrontiers in Psychology\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e, 1391832. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fpsyg.2024.1391832\u003c/span\u003e\u003cspan address=\"10.3389/fpsyg.2024.1391832\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePinto-Bernal, M., Biondina, M., \u0026amp; Belpaeme, T. (2025). Designing social robots with large language models for engaging human interaction. \u003cem\u003eApplied Sciences\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e (11), 6377. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/app15116377\u003c/span\u003e\u003cspan address=\"10.3390/app15116377\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSantini, Z. I., Jose, P. E., Cornwell, E. Y., Koyanagi, A., Nielsen, L., Hinrichsen, C.,. .. Koushede,\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eV. (2020). Social disconnectedness, perceived isolation, and symptoms of depression and anxiety among older americans (nshap): A longitudinal mediation analysis. \u003cem\u003eThe Lancet Public Health\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e (1), e62\u0026ndash;e70. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/S2468-2667(19)30230-0\u003c/span\u003e\u003cspan address=\"10.1016/S2468-2667(19)30230-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSmith, R. W., Barnes, I., Green, J., Reeves, G. K., Beral, V., \u0026amp; Floud, S. (2021). Social isolation and risk of heart disease and stroke: Analysis of two large uk prospective studies. \u003cem\u003eThe Lancet Public Health\u003c/em\u003e, \u003cem\u003e6\u003c/em\u003e (4), e232\u0026ndash;e239. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/S2468-2667(20)30291-7\u003c/span\u003e\u003cspan address=\"10.1016/S2468-2667(20)30291-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSpatola, N., Belletier, C., Chausse, P., \u0026amp; Augustinova, M. (2021). Social perception of artificial agents: A meta-analysis. \u003cem\u003ePsychological Bulletin\u003c/em\u003e, \u003cem\u003e147\u003c/em\u003e (11), 1121\u0026ndash;1143. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1037/bul0000340\u003c/span\u003e\u003cspan address=\"10.1037/bul0000340\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTopp, C. W., \u0026Oslash;stergaard, S. D., S\u0026oslash;ndergaard, S., \u0026amp; Bech, P. (2015). The who-5 well-being index: A systematic review of the literature. \u003cem\u003ePsychotherapy and Psychosomatics\u003c/em\u003e, \u003cem\u003e84\u003c/em\u003e (3), 167\u0026ndash;176. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1159/000376585\u003c/span\u003e\u003cspan address=\"10.1159/000376585\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTuneu, M. M., Roig, J. J., Molina, E. M., Planas, L. C., Salvans, A. E., Godayol, P. F.,.. . others (2023). Social and emotional loneliness among older people living in nursing homes in spain: A cross-sectional study. \u003cem\u003eAnales de Psicolog\u0026acute;ıa / Annals of Psychology\u003c/em\u003e, \u003cem\u003e39\u003c/em\u003e (3), 465\u0026ndash;477. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.20944/preprints202201.0423.v1\u003c/span\u003e\u003cspan address=\"10.20944/preprints202201.0423.v1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan \u0026rsquo;t Klooster, J., Capasso, M., van Gorssel, D., Vrolijk, E., Rettagliata, G., Gerritsen, D.,.. . Vonkeman, H. E. (2026). A gpt-reinforced social robot for patient communication: a pilot study. \u003cem\u003eFrontiers in Digital Health\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e, 1653168. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fdgth.2025.1653168\u003c/span\u003e\u003cspan address=\"10.3389/fdgth.2025.1653168\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVerhelst, E., \u0026amp; Belpaeme, T. (2025). Large language models cover for speech recognition mistakes: Evaluating conversational ai for second language learners., 1705\u0026ndash;1709. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/HRI61500.2025.10974188\u003c/span\u003e\u003cspan address=\"10.1109/HRI61500.2025.10974188\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWorld Health Organization. (2025). \u003cem\u003ePopulation ageing: Questions and answers.\u003c/em\u003e Retrieved from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.who.int/news-room/questions-and-answers/item/population-ageing\u003c/span\u003e\u003cspan address=\"https://www.who.int/news-room/questions-and-answers/item/population-ageing\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYen, H.-Y., Huang, C. W., Chiu, H.-L., \u0026amp; Jin, G. (2024). The effect of social robots on depression and loneliness for older residents in long-term care facilities: A meta-analysis of randomized controlled trials. \u003cem\u003eJournal of the American Medical Directors Association\u003c/em\u003e, \u003cem\u003e25\u003c/em\u003e (6). doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jamda.2024.02.017\u003c/span\u003e\u003cspan address=\"10.1016/j.jamda.2024.02.017\" targettype=\"DOI\" 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":"","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":"humanoid robot, large language models, conversational AI, user experience, feasibility study, older-adult care","lastPublishedDoi":"10.21203/rs.3.rs-9212133/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9212133/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLoneliness among older adults is a growing public health concern, particularly in residential care settings where opportunities for sustained one-on-one interaction are limited. Recent advances in large language models (LLMs) enable socially assistive humanoid robots to engage in more natural dialogue, potentially supporting interaction-level experiences related to social presence and engagement, which are conceptually relevant to loneliness. However, evidence from real-world care deployments remains limited. This study evaluated the feasibility and user experience (UX) of interactions with \u003cem\u003eAmeca\u003c/em\u003e (Engineered Arts, UK) in a Flemish residential-care setting and examined how baseline psychosocial factors relate to perceived interaction quality.\u003c/p\u003e \u003cp\u003eWe conducted a two-session, mixed-methods feasibility study. Thirty-one older adults (63\u0026ndash;93 years) completed a 15\u0026ndash;20 minute semi-structured Dutch conversation with \u003cem\u003eAmeca\u003c/em\u003e (Session 1); twenty participants returned approximately one week later for Session 2 after a prompt-level re-finement of the conversational configuration. Pre-interaction measures assessed loneliness and well-being. Post-interaction measures captured user experience (UX): perceived enjoyment, empathy, social comfort, trust, and comprehensibility complemented by open-ended feedback. Conversa-tion transcripts were available for Session 2 and were analyzed for turn-taking balance, response patterns, and thematic content.\u003c/p\u003e \u003cp\u003eUX ratings were generally positive and stable across sessions. Higher baseline loneliness was associated with less favorable UX evaluations, most notably reduced comprehensibility. Session 2 transcript analyses indicated balanced turn-taking but predominantly question\u0026ndash;answer interaction patterns, with conversational depth varying across participants and themes centering on everyday life. Overall, findings support the feasibility and acceptability of LLM-enabled humanoid inter-action in residential care and highlight design priorities for future systems, including improved pacing, personalization, and context handling to better support social presence and engagement.\u003c/p\u003e","manuscriptTitle":"LLM-Enabled Humanoid Interaction in Care Settings: A Two-Phase Study with Ameca","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-27 10:59:22","doi":"10.21203/rs.3.rs-9212133/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":"93137079-4019-483f-a093-69f62211a996","owner":[],"postedDate":"April 27th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":66938347,"name":"Health sciences/Health care"},{"id":66938348,"name":"Biological sciences/Psychology"},{"id":66938349,"name":"Social science/Psychology"}],"tags":[],"updatedAt":"2026-04-27T10:59:22+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-27 10:59:22","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9212133","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9212133","identity":"rs-9212133","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}