Large language model-supported companion robots for loneliness in older people: A UK–Japan qualitative study integrating focus groups and in-home deployment

Large language model-supported companion robots for loneliness in older people: A UK–Japan qualitative study integrating focus groups and in-home deployment | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Large language model-supported companion robots for loneliness in older people: A UK–Japan qualitative study integrating focus groups and in-home deployment Yuto Satake, Minna Nygren, Clare Yu, Kunihiko Katsuki, Ippei Endo, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9093780/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 Background: Loneliness is a critical social issue in older adults, with clinical implications. Conversational companion robots have been studied as one approach to ameliorating this issue. Large language models (LLMs) can enable flexible conversational ability in companion robots, but acceptability and suitability remain uncertain. We explored older people’s expectations and concerns regarding an LLM-supported companion robot for loneliness support. Methods: We conducted a UK–Japan qualitative study comprising hands-on focus groups for community-dwelling older adults in London (n=17) and a one-week in-home use with follow-up interviews in Osaka among outpatients with mild cognitive impairment (MCI; n=8). Transcripts were analysed using reflexive thematic analysis; for cross-site reporting, Japan themes/codes were mapped onto the thematic structure generated from the larger UK dataset. Descriptive questionnaire measures and at-home conversational log metrics were collected to contextualise qualitative findings. Results: Participants saw value of the companion robot as a support for older people with loneliness but emphasised that acceptability depends on interaction mechanics and user agency. Three cross-context themes were identified: (1) Practical use and functionality (response latency, turn-taking, desired features, and controllability in home use); (2) Emotional connection and engagement (social presence alongside perceived limits in conversational fit and depth); and (3) Ethical and societal reflections (privacy/data governance, access, and concerns about substituting for human contact). Conclusions: LLM-supported companion robots may provide acceptable low-intensity support for some older people, including those with MCI, provided that usability, user-adjustable control and ethical governance are prioritised. Longer deployments are needed to evaluate potential sustained benefit and burden. loneliness robots large language models AI mild cognitive impairment community-dwelling qualitative research Figures Figure 1 Introduction Loneliness in later life is increasingly recognised as a clinically relevant and societally important issue, associated with poorer mental health, reduced quality of life, and adverse health outcomes( 1 – 3 ). Older people may face heightened vulnerability to loneliness due to bereavement, reduced mobility, and shrinking social networks( 3 , 4 ). In mental health contexts, these challenges are often compounded by cognitive impairment and neuropsychiatric symptoms, which further complicate both the lived experience of loneliness and the feasibility of conventional social interventions. A range of approaches have been proposed to mitigate loneliness, including group-based programmes, befriending interventions, and social prescribing; however, implementation challenges, resource demands, and the need for tailoring and adaptability mean that effectiveness and fit may vary across settings and individuals( 5 – 7 ). For some individuals—particularly those living alone, with functional limitations, or with reduced confidence in leaving the home—available options may be difficult to access or sustain, and may not align with the practical realities of everyday life. Digital and assistive technologies have consequently attracted increasing attention as potentially scalable forms of low-intensity psychosocial support. Prior reviews and meta-analyses suggest that social robots and conversational agents may offer modest benefits for loneliness- and mood-related outcomes in older people, although the evidence base remains heterogeneous and is frequently characterised by methodological limitations ( 8 , 9 ). In our recent systematic review and meta-analysis of autonomous conversational agents in older people, interventions were associated with small-to-moderate improvements in loneliness and depressive symptoms, albeit with considerable heterogeneity and limited comparative evidence ( 10 ). Importantly, much of the existing literature focuses on rule-based or domain-limited systems, and relatively few studies have examined implementation in realistic home settings—particularly among individuals living with cognitive impairment or mental health conditions. Our groups’ feasibility work—including home-based studies of template-based conversational robots—suggests that some cognitively impaired older people may experience comfort or companionship from an embodied agent, while also encountering practical barriers such as limited conversational depth, response delays, and operational challenges ( 11 , 12 ). Acceptability depends not only on perceived psychosocial benefit but also on usability, conversational quality, perceived social appropriateness, and trust-related concerns such as privacy, costs, and fears of substituting technology for human contact ( 13 , 14 ) Recent advances in generative artificial intelligence—especially large language models (LLMs)—have further shifted expectations for quality of conversational support. Compared with scripted dialogue, LLM-enabled systems can produce more flexible and contextually varied responses, potentially supporting open-ended conversation where continuity and responsiveness may matter as much as task completion. However, risks associated with foundation models such as hallucinations, opacity, bias, and challenges in reliably aligning responses with user intent raise important concerns regarding safety, governance, and psychological appropriateness, particularly for vulnerable users ( 15 , 16 ). In embodied robotic settings, these risks may be further amplified by turn‑taking constraints, response latency, and heightened social expectations elicited by physical presence. Participatory work with older people has highlighted that LLM-supported companion robots can produce slow, repetitive, superficial, incoherent, or emotionally mismatched responses, and can prompt user worry ( 17 , 18 ). In an initial report from the broader Japan installation programme, we previously described a week-long home use case of an LLM-supported tabletop robot in an individual with mild cognitive impairment (MCI), illustrating both enjoyment and concerns about interaction burden when questioning felt too intelligent ( 19 ). To design AI companion robots that are acceptable, safe, and clinically appropriate, it is essential to understand users’ expectations and concerns in realistic contexts. In this study, we examine older people’s expectations and concerns regarding the use of an AI-enabled companion robot for loneliness support across two complementary contexts (UK focus groups and a Japan in-home installation study). The present manuscript incorporates and substantially extends the initial report by analysing a broader set of interviews and integrating these data with a complementary UK dataset. Our aim was to generate user-informed insights into the design tensions that shape acceptability and clinical appropriateness of generative AI companion robots for loneliness in later life. Methods Study Design We conducted a UK–Japan multi-site qualitative study comprising two complementary components: (1) focus group discussions, including interactions with a tabletop LLM-supported conversational robot (Sota) with community-dwelling older people in London, UK, and (2) a one-week in-home use of the robot with follow-up semi-structured interviews among Japanese outpatients with MCI in Osaka, Japan. We predetermined a target of at least four focus groups (totalling 17 participants) and eight individual interviews. This sample size aligns with established recommendations for qualitative research to achieve data saturation. Questionnaire measures were also collected in both components and were presented descriptively to contextualise the qualitative findings. This combined, mixed-methods approach—integrating group-based discussion with extended in-home use—enabled a more comprehensive examination of user expectations, experiences, and concerns in both short-term and everyday contexts. Ethical approval was obtained from the UCL Interaction Centre Local Research Ethics Committee (ID: 0416) and Ethical Review Board of Osaka University Hospital (approval ID: 24073(T1)). All participants provided written informed consent prior to participation. Robot/system overview The robot used in this study was Sota (Vstone, Osaka, Japan), a compact tabletop humanoid communication robot. Sota supports spoken interaction via onboard microphones and a speaker, provides basic arm gestures, and uses eye LEDs to communicate interaction state. The dialogue system followed a cloud-based pipeline consisting of automatic speech recognition (ASR), response generation using a large language model (GPT-4o via Azure OpenAI), and text-to-speech (TTS) synthesis. Dialogue mode could be initiated either by a physical control on the robot or by camera-based motion detection (used only as a trigger; no video was recorded or stored). Eye LEDs provided turn-taking cues. Further technical details are provided in Supplementary Note . Participants UK (London): Focus groups Participants were recruited between January and March 2025 via direct outreach to individuals who had previously taken part in ageing or dementia-prevention studies, and through referrals from existing participants. Inclusion criteria were: (1) age 60 years or older, (2) living independently in the community, (3) able to provide written informed consent, and (4) sufficient English proficiency to participate in a focus group. Exclusion criteria were: (1) diagnosed with dementia, and (2) uncontrolled physical illness. One participant aged 49 attended a focus group due to a recruiting error and was excluded from the analytic sample. Japan (Osaka): In-home installation and interviews Participants were recruited between April 2024 and March 2025 from the outpatient psychiatry clinic at Osaka University Hospital. Inclusion criteria were: (1) diagnosis of mild cognitive impairment according to the 2011 National Institute on Aging and Alzheimer’s Association (NIA-AA) criteria; and (2) age 60 years or older. Exclusion criteria were: (1) uncontrolled physical illness that may impede interaction with the robot, and (2) insufficient Japanese proficiency to participate in interviews. Data collection and procedures UK: Four focus groups were conducted at University College London (UCL) between February and March 2025. Sessions included 4–6 participants and lasted approximately two hours (including a short break). Sessions were facilitated by YS with at least one other team member as co-moderator. A discovery-oriented topic guide was used to elicit expectations and concerns regarding an AI companion robot for loneliness support, along with hands-on interaction with the robot. Sessions began with a demonstration of the robot, followed by both structured and unstructured interactions between participants and the robot and group discussion. Sessions were audio-recorded, and brief field notes taken during or immediately after each session. At the end of each session, participants completed questionnaires assessing user’s subjective impression and usability of robots (described later) and demographic questions. Japan: Following enrolment, a researcher visited each participant’s home to install the robot. The installation visit included an approximately 30-minute orientation, and participants received a brief written guide describing the significance of eye-colour indicators and basic turn-taking cues. Participants were encouraged to use the robot freely, with no minimum usage requirement; technical support was available on request. The robot was placed in a location suitable for conversation (e.g., on a living-room table) and positioned to minimise unintended audio capture from television sound where feasible. The intended duration was one week. When retrieval exactly one week later was not feasible, the robot was collected on the earliest practicable date thereafter (installations occurred between July 2024 and May 2025). A individual semi-structured interview (approximately 15 minutes) was conducted by KK or YS after the home use period and included questionnaire administration. Figure 1 shows photographs of the data collection settings for the UK focus groups and the Japan in-home deployment. Transcription, anonymisation, and translation Interviews and focus groups were audio-recorded using digital recorders and/or Microsoft Teams under institutional licences. The recordings were transcribed using Microsoft Teams automated transcription, stored within a university-managed cloud environment, and subsequently reviewed manually to correct errors and anonymise identifiers. Excerpts selected as illustrative quotations were translated from Japanese into English by the research team. System logs and in-home engagement metrics (Japan) We collected and analysed system logs to derive two engagement metrics (i) the proportion of full deployment days meeting the criterion for use (use-days) and (ii) mean daily dialogue-mode time (minutes/day). Use-days were defined as days with ≥20 minutes of recorded dialogue-mode interaction. Dialogue-mode time was defined as the duration from activation of dialogue mode to its termination, calculated from the corresponding timestamps. Engagement metrics were calculated across full deployment days, excluding the installation and removal days (partial days), and are presented descriptively. Questionnaire assessing subjective impressions and usability of the robot Subjective impressions on the robot were assessed using the Godspeed Questionnaire Series (GQS), reported as five subscales (Anthropomorphism, Animacy, Likeability, Perceived Intelligence, and Perceived Safety). Items were rated on a 5-point scale, and we report mean subscale scores (range 1–5), with higher scores indicating more favourable perceptions(20). Usability was assessed using the 10-item System Usability Scale (SUS; with ‘system’ replaced by ‘robot’), scored using the standard 0–100 metric (higher scores indicate better usability)(21). To aid interpretation, SUS scores were contextualised descriptively using published guidance (e.g., ≥70 generally acceptable; 70–89 good; >90 excellent; <70 marginal; <50 potentially unacceptable), while recognising that SUS should not be used in isolation for absolute judgements(22). Participants completed these questionnaires within both UK and Japan studies. Questionnaire results are presented descriptively to contextualise the qualitative findings. Qualitative data analysis UK focus group transcripts were managed and coded in NVivo (versions 14 and 15) and analysed using reflexive thematic analysis(23). Coding was led by YS, with MN acting as a second coder to support iterative refinement of codes and themes. Analysis involved repeated familiarisation, line-by-line coding, and development, review, and naming of themes through reflexive discussion within the research team. Japan post-installation interview transcripts were managed in NVivo (version 15) and analysed using reflexive thematic analysis(23). We used the thematic structure generated from the UK dataset as an organising framework. Coding was led by YS, with KK contributing to the iterative development and refinement of a shared coding framework to support team-based analysis. Initial coding was sensitised by domains covered in the interview guide (usability/operational challenges, emotional responses during interaction, and perceived limitations), while allowing inductive refinement and theme development through ongoing analytic discussion. To support cross-site integration, the thematic structure generated from the larger UK dataset was used as an organising framework. Codes and themes from the Japan dataset were subsequently mapped onto this structure through discussion among the UK and Japan research team. Instances of divergence (e.g., between UK and Japan participants; focus groups versus in-home deployment; cognitively unimpaired participants versus those with MCI) were explored through data triangulation and team-based discussion and were preserved as part of the contextual interpretation within themes. . Results Participant characteristics In the UK, 18 participants attended first-round focus groups; one participant aged 49 did not meet the age eligibility criteria and was excluded from the analytic sample. The UK sample comprised of 17 participants (mean age 72.5 years, SD 7.1; 61–85), including 10 women and 7 men, with mean years of education 14.6 (SD 3.4; 10–21). In Japan, eight participants were included (all female; mean age 81.1 years, SD 6.1; 73–93). Diagnoses included probable MCI with Lewy Bodies (MCI-LB (n = 5), and MCI due to Alzheimer’s Disease (AD) (n = 3); mean Mini-Mental Status Examination (MMSE) was 25.4 (SD 0.9; 24–27), all participants had Clinical Dementia Rating = 0.5. Participant characteristics are summarised in Table 1 . Table 1 Participant characteristics Characteristic UK focus groups (analytic sample) Japan home installation N 17 8 Age, mean (SD); range 72.5 (7.1); 61–85 81.1 (6.1); 73–93 Female, n (%) 10 (58.8%) 8 (100%) Education (years), mean (SD); range 14.6 (3.4); 10–21 12.8 (3.2); 9–18 Living alone, n (%) 10 (58.8%) 8 (100%) Digital Device Usage Smartphone (n = 17, 100%); Tablet (n = 13, 76.5%); PC (n = 14, 82.4%); Conversational agents (n = 6, 35.3%) Smartphone (n = 7, 87.5%); Tablet (n = 1, 12.5%); PC (n = 2, 25%); None (n = 1, 12.5%) Diagnosis Community-dwelling; no dementia diagnosis MCI-LB (n = 5); MCI due to AD (n = 3) MMSE, mean (SD); range — 25.4 (0.9); 24–27 CDR — 0.5 (all) CDR-SB, mean (SD); range — 2.1 (1.2); 0.5–4.0 UK focus groups represent the analytic sample. Abbreviations: UK, United Kingdom; MMSE, Mini-Mental State Examination; CDR, Clinical Dementia Rating; CDR-SB, Clinical Dementia Rating–Sum of Boxes; MCI, mild cognitive impairment; MCI-LB, mild cognitive impairment with Lewy bodies; AD, Alzheimer’s disease. CDR was 0.5 for all participants in the Japan home installation cohort. ‘—’ indicates not collected/available. Questionnaire measures and in-home engagement GQS and SUS scores are reported descriptively for both cohorts (Table 2 ) to contextualise the qualitative findings. In the UK focus groups (n = 17), mean (SD) GQS subscale scores were 2.9 (1.1) for Anthropomorphism, 3.2 (1.0) for Animacy, 3.8 (0.9) for Likeability, 3.4 (0.8) for Perceived Intelligence, and 3.9 (0.6) for Perceived Safety; mean SUS was 56.2 (11.4). In the Japan home installation cohort (n = 8), corresponding GQS scores were 3.3 (0.9), 3.7 (0.6), 4.4 (0.8), 4.0 (0.6), and 3.6 (0.5), and mean SUS was 62.8 (18.8). In-home engagement in the Japan cohort indicated use on 78.6% of deployment days (use-day defined as ≥ 20 minutes of recorded dialogue-mode interaction), and mean daily dialogue-mode time was 68.2 minutes/day (SD 41.7). Table 2 Questionnaire measures and in-home engagement Measure UK focus groups (n = 17) Japan home installation (n = 8) GQS Anthropomorphism, mean (SD) 2.9 (1.1) 3.3 (0.9) GQS Animacy, mean (SD) 3.2 (1.0) 3.7 (0.6) GQS Likeability, mean (SD) 3.8 (0.9) 4.4 (0.8) GQS Perceived Intelligence, mean (SD) 3.4 (0.8) 4.0 (0.6) GQS Perceived Safety, mean (SD) 3.9 (0.6) 3.6 (0.5) SUS score (0–100), mean (SD) 56.2 (11.4) 62.8 (18.8) In-home engagement (use-days, % of deployment days) — 72.4% In-home engagement (mean daily dialogue-mode time, minutes/day; SD) — 68.2 (41.7) Mean daily dialogue-mode time was calculated across full deployment days, excluding the installation and removal days (partial days). Days with no recorded dialogue were included and counted as 0 minutes. Use-days were defined as days with ≥ 20 minutes of recorded dialogue-mode interaction. Abbreviations: GQS, Godspeed Questionnaire Series; SUS, System Usability Scale. GQS subscales are scored on 5-point Likert-type items (higher scores indicate more favourable perceptions). SUS is reported on a 0–100 scale (higher scores indicate better usability). ‘—’ indicates not applicable/not assessed. Qualitative findings In the UK focus groups, we identified three overarching themes (with six sub-themes) capturing users’ expectations and concerns about using an AI companion robot for loneliness support in later life: ( 1 ) Practical Use and Functionality, ( 2 ) Emotional Connection and Engagement, and ( 3 ) Ethical and Societal Reflections ( Supplementary Table S1 ). Japan post-installation interviews were analysed independently using reflexive thematic analysis, yielding four themes : Becoming fluent in robot-mediated conversation; Maintaining agency over engagement and interruption; Constructing social presence through conversation and embodiment; and Negotiating boundaries and psychological load. For cross-site reporting, Japan themes/codes were subsequently mapped onto the UK-derived thematic framework ( Supplementary Table S2 ). Illustrative participant quotations are included to support each theme; UK focus group participants are labelled UK1–UK18, and Japan participants are labelled JP1–JP8. Theme 1: Practical Use and Functionality Subtheme 1.1. Operational Challenges and User Adaptation Across both contexts, participants considered that smooth interaction depended on basic usability features and learnable turn-taking. In the UK, brief hands-on exposure revealed that interaction latency and cue interpretation could disrupt conversational flow, with one participant noting: “You had to sit and wait for 10 seconds (for a response)—it was 9 1/2 seconds too long. (UK2)” In addition to latency, participants sometimes struggled to understand when to speak or how to recover from breakdowns or interruptions to conversational flow. In the Japan home setting, participants similarly described uncertainty around timing and whether the robot had heard and understood their speech. Several reported adapting their own communication style over time (e.g., simplifying phrasing or choosing words more deliberately). One participant noted that a one‑week deployment limited this process: “This time it was only a week… If it had been here all the time, I think I’d start to figure out.” (JP6). These accounts align with the descriptive SUS scores, indicating variable usability in both cohorts. Subtheme 1.2. Suggestions and Expectations for Improved Design and Functionality Across both contexts, participants articulated design expectations oriented toward everyday usefulness, particularly features that would reduce effort and increase practical value beyond conversation. In the UK focus groups, participants discussed features that could make the robot more practically useful and better integrated with existing routines and technologies, rather than remaining a standalone novelty. Music playing was repeatedly framed as meaningful for loneliness support (e.g., “There are lonely people—they like to listen to music. That’s important,” UK15). In the Japan study, improvement suggestions were particularly shaped by maintaining agency—being able to decide when to engage and when to suspend interaction in the flow of household life (“When my daughter or the grandkids come over, I switch it off… because it keeps talking the whole time”, JP7). Theme 2: Emotional Connection and Engagement Subtheme 2.1. Conversational Quality and Limitations In the UK focus groups, many participants reported that the robot’s conversational naturalness could be convincing at first (e.g., “It is definitely… it makes you feel like you’re talking to somebody,” UK1). However, this was frequently tempered by perceived limitations in depth, specificity, and continuity, with responses sometimes experienced as generic or non-committal (e.g., “It sounds like a politician—it won’t answer the question,” UK12). In Japan, conversational limitations were often described in terms of mismatch and interaction burden during everyday use. One participant noted intermittent misalignment: “There were times when the conversation didn’t quite line up… maybe I didn’t explain well enough” (JP7). Another highlighted that the robot’s frequent, unsolicited talk could become excessive: “It doesn’t need to react every time I pass by… It talks too much—too much chatter” (JP1). Subtheme 2.2. Expectations for the Robot as a Companion Across both contexts, participants’ expectations of “companionship” were strongly tied to embodiment and social presence—the sense that the robot is there as a responsive other, not merely a voice interface. In the UK, several participants suggested that physical movement and attentional signalling could make the robot feel more present than disembodied assistants (“The movement… would be better than Alexa. Because it’s more sort of… present,” UK11). Some also expected that companionability would depend on longer-term adaptation, including memory of prior conversations and preferences, enabling a relationship that could develop over time. In Japan, companion framing was often expressed through everyday relational talk directed toward the robot, akin to speaking to a pet or housemate. One participant described: “I talked to it like I would with a normal person… Even when I went to day care, I’d say, ‘I’m heading to day care now,’ or ‘Please watch the house’” (JP2). A small number also reported emerging attachment and reluctance to part after the one-week installation. Theme 3: Ethical and Societal Reflections Subtheme 3.1: Concerns about Access and Ethical Use In the UK focus groups, ethical acceptability was frequently discussed through the topics of privacy, profiling, and downstream misuse. Participants expressed unease that an AI companion could accumulate more intimate information than might be shared with clinicians and that such data could be repurposed or commercialised: “It’s going to have more accurate profiles on us… would you want your psychologist to be selling your data?” (UK14). Several extended this to crime-related vulnerability, noting that older people with pensions and assets could become “perfect targets” if profiling enabled exploitation. Some extended this to heightened vulnerability to scams or exploitation, and a broader dislike or scepticism toward “unnecessary” technology. In the Japan home-installation interviews, comparatively few participants raised explicit access or ethical-use concerns. Aside from a single privacy-boundary comment noted in the codebook, discussion in this subtheme was limited; the present dataset does not allow firm attribution to sample, setting, or cultural factors. Subtheme 3.2. Views on the Psychological Supportive Role of Robots Across both contexts, participants generally recognised potential benefit, but framed the robot’s role as conditional and situational, particularly for people who are isolated or have limited opportunities for social contact (e.g., “Someone can’t get out—it could be a godsend,” UK13). In the UK focus groups, this conditional framing was paired with concern about how the technology might be positioned within services: “If this is a partial solution, I would hate to see it being the only solution” (UK2). In the Japan home-installation interviews, participants likewise described context-dependent usefulness (e.g., for someone living alone), while a small number also noted psychological load (e.g., post-interaction rumination affecting sleep, JP8). Overall, these accounts suggest that supportive value depends on aligning conversational depth and initiative with users’ circumstances and capacity, rather than assuming uniform benefit. Discussion In this UK–Japan qualitative study, participants in both component sub-studies broadly recognised the potential value of an LLM-enabled conversational robot as a conditional form of support for loneliness in later life, particularly for people who are isolated or have limited opportunities for social contact. The UK focus groups, conducted after brief hands-on interaction, generated active discussion around desired functions and usability, the social acceptability of using a robot for companionship, and ethical concerns such as privacy and potential misuse. In contrast, the Japan home-installation interviews (analysed independently into four themes: becoming fluent in robot-mediated conversation; maintaining agency over engagement and interruption; constructing social presence through conversation and embodiment; and negotiating boundaries and psychological load) foregrounded the practical realities of use at home—especially the time needed to learn interaction routines and the importance of maintaining agency in everyday contexts. Together, the two components provide complementary insights for designing and governing LLM-enabled companion robots for older people. Engagement logs in the Japan cohort indicated use on 78.6% of full deployment days and a mean daily dialogue-mode time of 68.2 minutes/day (SD 41.7), suggesting that meaningful day-to-day use was feasible for most participants during the one-week deployment. Practical usability and the burden of adaptation For a conversational companion robot to be sustainable in people’s homes, basic dialogue mechanics must come first: turn-taking stability, low end-to-end latency, and robust speech recognition. Our iterative refinement work found these are not merely desirable but essential prerequisites for perceived usability, consistent with LLM-robot reports that highlight both timing and ASR breakdowns as prominent barriers ( 17 ). In the present study, participants likewise framed usability as something that requires effort and learning, indicating that shortcomings in timing and recognition translate into a burden of adaptation rather than a minor inconvenience ( 14 ). Second, participants also expressed a wish for additional functions (e.g., music for comfort/enjoyment, multilingual support, and safety-oriented features such as contacting family after a fall), echoing the findings of our earlier home work ( 12 ). However, adding functions can increase configuration demands and discoverability burdens; both our field experience and general usability evidence suggest that introducing new, untested components can reduce usability unless the interface and workflow are stabilised ( 22 ). A staged strategy therefore seems safer: establish a feasible “core companion” with reliable conversational fluency first, then expanding functionality incrementally with user testing—especially when designing for users with cognitive impairment. Third, sustained domestic use depends on context-sensitive control and agency: users need to be able to initiate, pause, silence, or limit interaction in ways that fit daily routines and social context. In our Japan home installation, “maintaining agency” emerged as a salient requirement (e.g., the need to suppress interaction when family members visit), suggesting that mode control and interruptibility are core usability features rather than optional extras. This complements acceptance models in older people, where perceived controllability and ease of use are central determinants of intention to use and continued use ( 14 ). Emotional engagement, companionship, and the importance of control A key advance in the present work is that introducing an LLM enabled a level of conversational naturalness that is difficult to achieve with scenario-based dialogue alone. For example, in our earlier home study with RoBoHoN, conversation was explicitly scenario-based and could not respond to more complex speech, highlighting inherent constraints of scripted pathways ( 12 ). In contrast, participants in both the UK and Japan components described the LLM-enabled interaction as feeling close to human conversation (e.g., “it makes you feel like you’re talking to somebody,” and “it really felt like talking with a person”). This improvement is important because conversational fluency is a prerequisite for companionship, but it was also clear that fluency alone is not sufficient for this. Companionship was repeatedly linked to presence, namely whether the robot is experienced as a socially present counterpart, which embodiment plausibly supports (e.g., movement being “more… present” than a disembodied assistant). Prior acceptance work similarly positions social presence as a pathway to more positive experience and higher acceptance of companion robots ( 24 ). The Japan home installation further suggested that repeated everyday contact may foster attachment (including reluctance to part after only one week), echoing longitudinal HRI evidence that attachment-related responses can emerge with ongoing exposure, while remaining context-dependent rather than inevitable ( 25 – 27 ). Despite high ratings of naturalness, UK participants frequently pointed to insufficient depth (generic or non-committal answers and limited contextual continuity). This raises a design question: whether depth is primarily a limitation of open-domain LLM chat, or can be improved through prompt design and system-level scaffolding. Dialogue research has argued that unconstrained “chat about anything” often produces shallow interaction unless systems establish common ground and structured purposes ( 28 ). For LLM-enabled companion robots, recent work likewise emphasises that improving perceived depth likely requires architecture-level strategies (e.g., grounding, retrieval, structured interaction, and memory) rather than generation alone ( 17 , 29 ). At the same time, increasing “personality” or opinionatedness to deepen companionship introduces ethical tension: stronger social presence and more assertive dialogue may increase persuasive influence, and sustained emotionally intense engagement with conversational AI has been discussed as a potential risk factor for delusion-like experiences in vulnerable individuals ( 30 ), underscoring the need for safety-by-design and clear governance in socially assistive AI systems( 31 ). Finally, Japan participants’ concerns about the robot talking too much highlight a complementary design challenge: unsolicited talk can feel intrusive in everyday home contexts, particularly when users are moving around or when other people are present and may undermine long-term acceptability. This reinforces recommendations that talkativeness and initiative should be adjustable to context and user preference, with clear controls for pausing, silencing, and interrupting interaction rather than treated as uniformly desirable ( 18 ). In-home engagement also showed substantial interaction time, suggesting the importance of calibrating interaction intensity so that it remains comfortable rather than burdensome. Ethical and psychological appropriateness in psychogeriatric contexts Broader ethical concerns were salient, particularly around privacy, data governance, equitable access, and the fear that technology might be used to substitute for human contact. Such concerns are well recognised in the sociology and ethics of care robotics, including anxieties about deception, dependence, and the reconfiguration of care obligations ( 13 ). These issues matter acutely in psychogeriatrics because loneliness interventions often operate within care ecosystems characterised by resource constraints, service gaps, and vulnerability to inappropriate substitution. Participants’ accounts suggest that acceptability depends on clear framing of companion robots as adjunctive supports—potentially helpful in specific contexts (e.g., living alone, limited day-to-day contact)—rather than an “only solution.” A particularly clinically relevant contribution from the Osaka in-home interviews was the articulation of psychological intensity and interaction burden. Some participants described pressure to respond “seriously” or experiencing discomfort when the robot’s questioning felt as though it was “seeing through” them, with effects that could persist beyond the interaction (e.g., rumination). Together, these findings suggest that generative AI systems, if not carefully calibrated, may drift toward a quasi-therapeutic or probing interaction style that some users experience as intrusive—particularly those with cognitive impairment or heightened anxiety. In practice, this supports several safeguards: transparent role framing (what the robot is and is not), user-selectable conversation modes (light chat versus reflective talk), explicit consent cues before sensitive topics, and de-escalation pathways when distress is detected. These design priorities also align with broader discussions of foundation-model risks—where hallucination, opacity, and miscalibrated social persuasion are salient—and underscore the importance of risk-calibrated oversight and conservative defaults in clinical-adjacent deployments ( 15 , 16 ). It is also possible that differences in study format and participant characteristics (cognitively unimpaired focus groups vs brief in-home interviews with outpatients with MCI) shaped what was foregrounded, with broader ethical considerations potentially less salient or harder to articulate in the latter context. This further supports conservative defaults and clear governance in psychogeriatric deployments, where downstream privacy, profiling, or substitution risks may be consequential even when not spontaneously raised. Strengths and limitations This study integrates perspectives from two complementary contexts—hands-on group interaction in the UK and situated in-home use in Japan—includes a clinical outpatient sample (MCI), and combines qualitative accounts with brief subjective measures and planned in-home log metrics. Key limitations should be noted. First, recruitment occurred through specific channels, limiting generalisability and increasing the likelihood of selection bias. Second, the UK and Japan components differed substantially in clinical characteristics, cultural context, and study format (hands-on focus groups vs in-home deployment; cognitively unimpaired participants vs outpatients with MCI); while informative, these differences constrain direct comparison and mean that observed differences should not be attributed to any single factor. Third, the Osaka cohort was small and clinically heterogeneous (MCI due to Alzheimer’s disease and Lewy body disease), and the one-week deployment cannot establish longer-term trajectories such as habituation, sustained attachment, disengagement, or adverse psychological effects. Fourth, qualitative findings are sensitive to interview context and phrasing; cross-language reporting adds additional interpretive layers because only selected excerpts were translated. Although this was mitigated through careful anonymisation and team review, future work would benefit from more formalised translation procedures and bilingual audit trails. Conclusion Generative AI-supported companion robots may provide older people with interactions relevant to loneliness support, but acceptability depends on practical usability, user control, and calibration of conversational depth and initiative. Cross-context user perspectives—particularly from in-home use among individuals with MCI—highlight design tensions and safeguards needed for clinically responsible implementation. Clinically, these findings support cautious optimism for conditional use where social contact is limited, provided privacy governance and psychological appropriateness are addressed. Future studies should test longer deployments and evaluate clinically meaningful outcomes, while examining how design choices influence both benefit and burden. Declarations Conflict of interest Authors AT, IE and KU are employees of NTT West, which provided in-kind support for this study, including the loan of the Sota robot(s), server infrastructure, and payment of Microsoft Azure OpenAI service usage fees. The remaining authors declare no competing interests. Description of author’s role YS designed the studies, led participant recruitment, and conducted data collection and analysis. YS also wrote the first draft of the manuscript. MN, CY, KK and IE supported data collection and data curation, with CY and KK additionally assisting with recruitment. IE, AT and KU developed the robot’s dialogue system and contributed to improvements in user experience. PR provided oversight of qualitative data collection and analysis. NB contributed to the ethics approval process and advised on refinements to the robot’s UX design. MI obtained funding and supervised the study design. RH provided overall project leadership and supervision and guided manuscript development. All authors critically reviewed the manuscript and approved the final version. Acknowledgements This work was supported by the SENSHIN Medical Research Foundation (Overseas Study grant), Daiichi-Sankyo ”Habataku” Support Program for the Next Generation of Researchers and by the EPSRC-funded Sensing to Collaboration Programme Grant (EP/V000748/1). Declaration of generative AI and AI-assisted technologies in the manuscript preparation process During the preparation of this work, the authors used ChatGPT (OpenAI) to support language editing and manuscript organisation. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article. References Holt-Lunstad J (2024) Social connection as a critical factor for mental and physical health: evidence, trends, challenges, and future implications. World Psychiatry 23(3):312–332. 10.1002/wps.21224 Lee SL, Pearce E, Ajnakina O, Johnson S, Lewis G, Mann F et al (2021) The association between loneliness and depressive symptoms among adults aged 50 years and older: a 12-year population-based cohort study. Lancet Psychiatry 8(1):48–57. 10.1016/S2215-0366(20)30383-7 Salari N, Najafi H, Rasoulpoor S, Canbary Z, Heidarian P, Mohammadi M (2025) The global prevalence and associated factors of loneliness in older adults: a systematic review and meta-analysis. Humanit Soc Sci Commun 12(1):985. 10.1057/s41599-025-05304-x Luhmann M, Hawkley LC (2016) Age Differences in Loneliness from Late Adolescence to Oldest Old Age. Dev Psychol 52(6):943–959. 10.1037/dev0000117 Gardiner C, Geldenhuys G, Gott M (2018) Interventions to reduce social isolation and loneliness among older people: an integrative review. Health Soc Care Community 26(2):147–157. 10.1111/hsc.12367 Gilbody S, Littlewood E, McMillan D, Atha L, Bailey D, Baird K et al (2024) Behavioural activation to mitigate the psychological impacts of COVID-19 restrictions on older people in England and Wales (BASIL+): a pragmatic randomised controlled trial. Lancet Healthy Longev 5(2):e97–107. 10.1016/S2666-7568(23)00238-6 Hoang P, King JA, Moore S, Moore K, Reich K, Sidhu H et al (2022) Interventions Associated With Reduced Loneliness and Social Isolation in Older Adults: A Systematic Review and Meta-analysis. JAMA Netw Open 5(10):e2236676. 10.1001/jamanetworkopen.2022.36676 Gasteiger N, Loveys K, Law M, Broadbent E (2021) Friends from the Future: A Scoping Review of Research into Robots and Computer Agents to Combat Loneliness in Older People. Clin Interv Aging 16:941–971. 10.2147/CIA.S282709 Pu L, Moyle W, Jones C, Todorovic M (2019) The Effectiveness of Social Robots for Older Adults: A Systematic Review and Meta-Analysis of Randomized Controlled Studies. Gerontologist 59(1):e37–51. 10.1093/geront/gny046 Satake Y, Costello H, Naran N, Ishimaru D, Ikeda M, Howard R (2026) Autonomous conversational agents for loneliness, social isolation, depression, and anxiety in older people without cognitive impairment: Systematic review and meta-analysis. Psychol Med 56:e27. 10.1017/S0033291725103073 Nagata Y, Satake Y, Yamazaki R, Nishio S, Ishiguro H, Ikeda M (2026) Case report: A five–case series of 18 – month home use of a conversational companion robot for psychological support in older people with mild cognitive impairment or late–onset psychosis. Front Psychiatry 16:1700340. 10.3389/fpsyt.2025.1700340 Nagata Y, Satake Y, Yamazaki R, Nishio S, Suzuki M, Kanemoto H et al (2025) A Conversational Robot for Cognitively Impaired Older People Who Live Alone: An Exploratory Feasibility Study. Psychogeriatrics 25(5):e70076. 10.1111/psyg.70076 Berridge C, Zhou Y, Robillard JM, Kaye J (2023) Companion robots to mitigate loneliness among older adults: Perceptions of benefit and possible deception. Front Psychol 14:1106633. 10.3389/fpsyg.2023.1106633 Heerink M, Kröse B, Evers V, Wielinga B (2010) Assessing Acceptance of Assistive Social Agent Technology by Older Adults: the Almere Model. Int J Soc Robot 2(4):361–375. 10.1007/s12369-010-0068-5 Bender EM, Gebru T, McMillan-Major A, Shmitchell S (2021) On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 列. In: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency [Internet]. Virtual Event Canada: ACM; pp. 610–23. 10.1145/3442188.3445922 Bommasani R, Hudson DA, Adeli E, Altman R, Arora S, von Arx S et al (2022) On the Opportunities and Risks of Foundation Models [Internet]. 10.48550/arXiv.2108.07258 . arXiv Irfan B, Kuoppamäki S, Hosseini A, Skantze G (2025) Between reality and delusion: challenges of applying large language models to companion robots for open-domain dialogues with older adults. Auton Robots 49(1):9. 10.1007/s10514-025-10190-y Irfan B, Kuoppamäki S, Skantze G (2024) Recommendations for designing conversational companion robots with older adults through foundation models. Front Robot AI 11:1363713. 10.3389/frobt.2024.1363713 Satake Y, Ishimaru D, Katsuki K, Rapaport P, Berthouze N, Howard R et al (2025) A Week With a Conversational Large Language Model Companion Robot. Am J Geriatr Psychiatry 33(7):799–800. 10.1016/j.jagp.2025.03.010 Bartneck C, Kulić D, Croft E, Zoghbi S (2009) Measurement Instruments for the Anthropomorphism, Animacy, Likeability, Perceived Intelligence, and Perceived Safety of Robots. Int J Soc Robot 1(1):71–81. 10.1007/s12369-008-0001-3 Brooke J (1996) SUS - A quick and dirty usability scale. Usability Evaluation in Industry. Taylor & Francis, pp 189–194 Bangor A, Kortum PT, Miller JT (2008) An Empirical Evaluation of the System Usability Scale. Int J Hum-Comput Interact 24(6):574–594. 10.1080/10447310802205776 Braun V, Clarke V (2021) Thematic analysis: a practical guide. SAGE, London; Thousand Oaks, California, p 376 Heerink M, Kröse B, Wielinga B, Evers V (2008) Enjoyment intention to use and actual use of a conversational robot by elderly people. In: Proceedings of the 3rd ACM/IEEE international conference on Human robot interaction [Internet]. New York, NY, USA: Association for Computing Machinery; pp. 113–20. (HRI ’08). 10.1145/1349822.1349838 Dziergwa M, Kaczmarek M, Kaczmarek P, Kędzierski J, Wadas-Szydłowska K (2018) Long-Term Cohabitation with a Social Robot: A Case Study of the Influence of Human Attachment Patterns. Int J Soc Robot 10(1):163–176. 10.1007/s12369-017-0439-2 van Maris A, Zook N, Caleb-Solly P, Studley M, Winfield A, Dogramadzi S (2020) Designing Ethical Social Robots—A Longitudinal Field Study With Older Adults. Front Robot AI 7:1. 10.3389/frobt.2020.00001 Yamazaki R, Nishio S, Nagata Y, Satake Y, Suzuki M, Kanemoto H et al (2023) Long-term effect of the absence of a companion robot on older adults: A preliminary pilot study. Front Comput Sci 5:1129506. 10.3389/fcomp.2023.1129506 Skantze G, Doğruöz AS (2023) The Open-domain Paradox for Chatbots: Common Ground as the Basis for Human-like Dialogue. In: Stoyanchev S, Joty S, Schlangen D, Dusek O, Kennington C, Alikhani M, editors. Proceedings of the 24th Annual Meeting of the Special Interest Group on Discourse and Dialogue [Internet]. Prague, Czechia: Association for Computational Linguistics; pp. 605–14. 10.18653/v1/2023.sigdial-1.57 Pan Z, Wu Q, Jiang H, Luo X, Cheng H, Li D et al (2025) On Memory Construction and Retrieval for Personalized Conversational Agents [Internet]. 10.48550/ARXIV.2502.05589 . arXiv Hudon A, Stip E (2025) Delusional Experiences Emerging From AI Chatbot Interactions or AI Psychosis. JMIR Ment Health 12(1):e85799. 10.2196/85799 Ranisch R, Haltaufderheide J (2025) Rapid Integration of LLMs in Healthcare Raises Ethical Concerns: An Investigation into Deceptive Patterns in Social Robots. Digit Soc 4(1):7. 10.1007/s44206-025-00161-2 Additional Declarations The authors declare potential competing interests as follows: Authors AT, IE and KU are employees of NTT West, which provided in-kind support for this study, including the loan of the Sota robot(s), server infrastructure, and payment of Microsoft Azure OpenAI service usage fees. The remaining authors declare no competing interests. Supplementary Files SupplementaryNote.docx SupplementaryTableS1.docx SupplementaryTableS2.docx 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-9093780","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":604391774,"identity":"a8411dbd-bbed-4aff-b9e3-a17e9afcf967","order_by":0,"name":"Yuto Satake","email":"data:image/png;base64,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","orcid":"","institution":"Division of Psychiatry, University College London","correspondingAuthor":true,"prefix":"","firstName":"Yuto","middleName":"","lastName":"Satake","suffix":""},{"id":604391775,"identity":"bfc783ec-49b4-4376-a3a8-13f08c092a1e","order_by":1,"name":"Minna Nygren","email":"","orcid":"","institution":"UCL Interaction Centre, University College London","correspondingAuthor":false,"prefix":"","firstName":"Minna","middleName":"","lastName":"Nygren","suffix":""},{"id":604391776,"identity":"12115ece-23c9-4226-b3d1-453361a4f14c","order_by":2,"name":"Clare Yu","email":"","orcid":"","institution":"Division of Psychiatry, University College London","correspondingAuthor":false,"prefix":"","firstName":"Clare","middleName":"","lastName":"Yu","suffix":""},{"id":604391777,"identity":"1f2ea243-952d-4511-8e3c-9a51e57e8168","order_by":3,"name":"Kunihiko Katsuki","email":"","orcid":"","institution":"Department of Psychiatry, The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Kunihiko","middleName":"","lastName":"Katsuki","suffix":""},{"id":604391778,"identity":"189a44f3-e694-4052-af5f-8691308b3bdb","order_by":4,"name":"Ippei Endo","email":"","orcid":"","institution":"NTT WEST, Inc.","correspondingAuthor":false,"prefix":"","firstName":"Ippei","middleName":"","lastName":"Endo","suffix":""},{"id":604391779,"identity":"828dc001-9add-49e0-98a3-61909c8079c6","order_by":5,"name":"Akira Tsuboi","email":"","orcid":"","institution":"NTT WEST, Inc.","correspondingAuthor":false,"prefix":"","firstName":"Akira","middleName":"","lastName":"Tsuboi","suffix":""},{"id":604391780,"identity":"58229c73-f9d4-4d6e-a8fe-ca49aae95ef0","order_by":6,"name":"Daiki Ishimaru","email":"","orcid":"","institution":"Department of Psychiatry, The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Daiki","middleName":"","lastName":"Ishimaru","suffix":""},{"id":604391781,"identity":"8408ee9d-9ab7-4dbe-9a49-ab83ee6c1184","order_by":7,"name":"Kazuhiro Umemura","email":"","orcid":"","institution":"NTT WEST, Inc.","correspondingAuthor":false,"prefix":"","firstName":"Kazuhiro","middleName":"","lastName":"Umemura","suffix":""},{"id":604391782,"identity":"b81389c9-c85c-4a79-a6e8-b069209ba13d","order_by":8,"name":"Penny Rapaport","email":"","orcid":"","institution":"Division of Psychiatry, University College London","correspondingAuthor":false,"prefix":"","firstName":"Penny","middleName":"","lastName":"Rapaport","suffix":""},{"id":604391783,"identity":"ebfb5edc-2e3a-4b59-8205-e3b35e7a3f71","order_by":9,"name":"Nadia Bianchi-Berthouze","email":"","orcid":"","institution":"UCL Interaction Centre, University College London","correspondingAuthor":false,"prefix":"","firstName":"Nadia","middleName":"","lastName":"Bianchi-Berthouze","suffix":""},{"id":604391784,"identity":"5c13d61a-9a80-4740-99cb-09754f8fc614","order_by":10,"name":"Manabu Ikeda","email":"","orcid":"","institution":"Department of Psychiatry, The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Manabu","middleName":"","lastName":"Ikeda","suffix":""},{"id":604391785,"identity":"726caed4-e2a5-4021-b5dd-c64cf4d2e5cc","order_by":11,"name":"Robert Howard","email":"","orcid":"","institution":"Division of Psychiatry, University College London","correspondingAuthor":false,"prefix":"","firstName":"Robert","middleName":"","lastName":"Howard","suffix":""}],"badges":[],"createdAt":"2026-03-11 11:47:30","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":true,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9093780/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9093780/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104583040,"identity":"8cbdb254-c94d-4d4e-b1f9-2f5aa69e8bb8","added_by":"auto","created_at":"2026-03-13 15:17:40","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":6752,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eData collection settings in the UK and Japan components\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(A) UK focus group setting. Participants took turns interacting one-to-one with the companion robot positioned at the centre of the table, followed by a facilitated group discussion. The facilitator (YS) is seated at the far end of the room.\u003c/p\u003e\n\u003cp\u003e(B) Japan in-home deployment setting. The robot was installed in the home of a participant living alone with mild cognitive impairment. A brief user guide provided by the research team was placed on the table to support everyday use during the deployment period.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9093780/v1/1cee6e589c4e51f47a9951b5.png"},{"id":104781067,"identity":"3fc41aed-d136-4105-bae7-96170a07de11","added_by":"auto","created_at":"2026-03-17 07:54:38","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":963338,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9093780/v1/42043c49-3366-4497-a61f-d198c73198c3.pdf"},{"id":104583041,"identity":"4d4b77fe-f82d-4865-9ef4-c8b4d1953e90","added_by":"auto","created_at":"2026-03-13 15:17:40","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":16213,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryNote.docx","url":"https://assets-eu.researchsquare.com/files/rs-9093780/v1/e8a75cbb986ba07af078282b.docx"},{"id":104583043,"identity":"bb5fd5f6-b69d-4340-8a6a-f5f0d540e407","added_by":"auto","created_at":"2026-03-13 15:17:41","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":35152,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS1.docx","url":"https://assets-eu.researchsquare.com/files/rs-9093780/v1/baf11aca2883cddfe043a7ee.docx"},{"id":104583042,"identity":"55cd8e38-9ff6-47d5-966a-e8c9f1d6dcec","added_by":"auto","created_at":"2026-03-13 15:17:41","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":27779,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS2.docx","url":"https://assets-eu.researchsquare.com/files/rs-9093780/v1/c986afbeaf441e7f8c565006.docx"}],"financialInterests":"The authors declare potential competing interests as follows: Authors AT, IE and KU are employees of NTT West, which provided in-kind support for this study, including the loan of the Sota robot(s), server infrastructure, and payment of Microsoft Azure OpenAI service usage fees. The remaining authors declare no competing interests.","formattedTitle":"\u003cp\u003eLarge language model-supported companion robots for loneliness in older people: A UK–Japan qualitative study integrating focus groups and in-home deployment\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eLoneliness in later life is increasingly recognised as a clinically relevant and societally important issue, associated with poorer mental health, reduced quality of life, and adverse health outcomes(\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). Older people may face heightened vulnerability to loneliness due to bereavement, reduced mobility, and shrinking social networks(\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e). In mental health contexts, these challenges are often compounded by cognitive impairment and neuropsychiatric symptoms, which further complicate both the lived experience of loneliness and the feasibility of conventional social interventions. A range of approaches have been proposed to mitigate loneliness, including group-based programmes, befriending interventions, and social prescribing; however, implementation challenges, resource demands, and the need for tailoring and adaptability mean that effectiveness and fit may vary across settings and individuals(\u003cspan additionalcitationids=\"CR6\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e). For some individuals\u0026mdash;particularly those living alone, with functional limitations, or with reduced confidence in leaving the home\u0026mdash;available options may be difficult to access or sustain, and may not align with the practical realities of everyday life.\u003c/p\u003e \u003cp\u003eDigital and assistive technologies have consequently attracted increasing attention as potentially scalable forms of low-intensity psychosocial support. Prior reviews and meta-analyses suggest that social robots and conversational agents may offer modest benefits for loneliness- and mood-related outcomes in older people, although the evidence base remains heterogeneous and is frequently characterised by methodological limitations (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e). In our recent systematic review and meta-analysis of autonomous conversational agents in older people, interventions were associated with small-to-moderate improvements in loneliness and depressive symptoms, albeit with considerable heterogeneity and limited comparative evidence (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e). Importantly, much of the existing literature focuses on rule-based or domain-limited systems, and relatively few studies have examined implementation in realistic home settings\u0026mdash;particularly among individuals living with cognitive impairment or mental health conditions.\u003c/p\u003e \u003cp\u003eOur groups\u0026rsquo; feasibility work\u0026mdash;including home-based studies of template-based conversational robots\u0026mdash;suggests that some cognitively impaired older people may experience comfort or companionship from an embodied agent, while also encountering practical barriers such as limited conversational depth, response delays, and operational challenges (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e). Acceptability depends not only on perceived psychosocial benefit but also on usability, conversational quality, perceived social appropriateness, and trust-related concerns such as privacy, costs, and fears of substituting technology for human contact (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e)\u003c/p\u003e \u003cp\u003eRecent advances in generative artificial intelligence\u0026mdash;especially large language models (LLMs)\u0026mdash;have further shifted expectations for quality of conversational support. Compared with scripted dialogue, LLM-enabled systems can produce more flexible and contextually varied responses, potentially supporting open-ended conversation where continuity and responsiveness may matter as much as task completion. However, risks associated with foundation models such as hallucinations, opacity, bias, and challenges in reliably aligning responses with user intent raise important concerns regarding safety, governance, and psychological appropriateness, particularly for vulnerable users (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e). In embodied robotic settings, these risks may be further amplified by turn‑taking constraints, response latency, and heightened social expectations elicited by physical presence. Participatory work with older people has highlighted that LLM-supported companion robots can produce slow, repetitive, superficial, incoherent, or emotionally mismatched responses, and can prompt user worry (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e). In an initial report from the broader Japan installation programme, we previously described a week-long home use case of an LLM-supported tabletop robot in an individual with mild cognitive impairment (MCI), illustrating both enjoyment and concerns about interaction burden when questioning felt too intelligent (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTo design AI companion robots that are acceptable, safe, and clinically appropriate, it is essential to understand users\u0026rsquo; expectations and concerns in realistic contexts. In this study, we examine older people\u0026rsquo;s expectations and concerns regarding the use of an AI-enabled companion robot for loneliness support across two complementary contexts (UK focus groups and a Japan in-home installation study). The present manuscript incorporates and substantially extends the initial report by analysing a broader set of interviews and integrating these data with a complementary UK dataset. Our aim was to generate user-informed insights into the design tensions that shape acceptability and clinical appropriateness of generative AI companion robots for loneliness in later life.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eStudy Design\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe conducted a UK–Japan multi-site qualitative study comprising two complementary components: (1) focus group discussions, including interactions with a tabletop LLM-supported conversational robot (Sota) with community-dwelling older people in London, UK, and (2) a one-week in-home use of the robot with follow-up semi-structured interviews among Japanese outpatients with MCI in Osaka, Japan. We predetermined a target of at least four focus groups (totalling 17 participants) and eight individual interviews. This sample size aligns with established recommendations for qualitative research to achieve data saturation. Questionnaire measures were also collected in both components and were presented descriptively to contextualise the qualitative findings. This combined, mixed-methods approach—integrating group-based discussion with extended in-home use—enabled a more comprehensive examination of user expectations, experiences, and concerns in both short-term and everyday contexts.\u003c/p\u003e\n\u003cp\u003eEthical approval was obtained from the UCL Interaction Centre Local Research Ethics Committee (ID: 0416) and Ethical Review Board of Osaka University Hospital (approval ID: 24073(T1)). All participants provided written informed consent prior to participation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRobot/system overview\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe robot used in this study was Sota (Vstone, Osaka, Japan), a compact tabletop humanoid communication robot. Sota supports spoken interaction via onboard microphones and a speaker, provides basic arm gestures, and uses eye LEDs to communicate interaction state. The dialogue system followed a cloud-based pipeline consisting of automatic speech recognition (ASR), response generation using a large language model (GPT-4o via Azure OpenAI), and text-to-speech (TTS) synthesis. Dialogue mode could be initiated either by a physical control on the robot or by camera-based motion detection (used only as a trigger; no video was recorded or stored). Eye LEDs provided turn-taking cues. Further technical details are provided in \u003cstrong\u003eSupplementary Note\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eParticipants\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eUK (London): Focus groups\u003c/p\u003e\n\u003cp\u003eParticipants were recruited between January and March 2025 via direct outreach to individuals who had previously taken part in ageing or dementia-prevention studies, and through referrals from existing participants. Inclusion criteria were: (1) age 60 years or older, (2) living independently in the community, (3) able to provide written informed consent, and (4) sufficient English proficiency to participate in a focus group. Exclusion criteria were: (1) diagnosed with dementia, and (2) uncontrolled physical illness.\u0026nbsp;One participant aged 49 attended a focus group due to a recruiting error and was excluded from the analytic sample.\u003c/p\u003e\n\u003cp\u003eJapan (Osaka): In-home installation and interviews\u003c/p\u003e\n\u003cp\u003eParticipants were recruited between April 2024 and March 2025 from the outpatient psychiatry clinic at Osaka University Hospital. Inclusion criteria were: (1) diagnosis of mild cognitive impairment according to the 2011 National Institute on Aging and Alzheimer’s Association (NIA-AA) criteria; and (2) age 60 years or older. Exclusion criteria were: (1) uncontrolled physical illness that may impede interaction with the robot, and (2) insufficient Japanese proficiency to participate in interviews.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData collection and procedures\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eUK:\u003c/p\u003e\n\u003cp\u003eFour focus groups were conducted at University College London (UCL) between February and March 2025. Sessions included 4–6 participants and lasted approximately two hours (including a short break). Sessions were facilitated by YS with at least one other team member as co-moderator. A discovery-oriented topic guide was used to elicit expectations and concerns regarding an AI companion robot for loneliness support, along with hands-on interaction with the robot. Sessions began with a demonstration of the robot, followed by both structured and unstructured interactions between participants and the robot and group discussion. Sessions were audio-recorded, and brief field notes taken during or immediately after each session. At the end of each session, participants completed questionnaires assessing user’s subjective impression and usability of robots (described later) and demographic questions.\u003c/p\u003e\n\u003cp\u003eJapan:\u003c/p\u003e\n\u003cp\u003eFollowing enrolment, a researcher visited each participant’s home to install the robot. The installation visit included an approximately 30-minute orientation, and participants received a brief written guide describing the significance of eye-colour indicators and basic turn-taking cues. Participants were encouraged to use the robot freely, with no minimum usage requirement; technical support was available on request. The robot was placed in a location suitable for conversation (e.g., on a living-room table) and positioned to minimise unintended audio capture from television sound where feasible. The intended duration was one week. When retrieval exactly one week later was not feasible, the robot was collected on the earliest practicable date thereafter (installations occurred between July 2024 and May 2025). A individual semi-structured interview (approximately 15 minutes) was conducted by KK or YS after the home use period and included questionnaire administration.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFigure 1\u003c/strong\u003e shows photographs of the data collection settings for the UK focus groups and the Japan in-home deployment.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTranscription, anonymisation, and translation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eInterviews and focus groups were audio-recorded using digital recorders and/or Microsoft Teams under institutional licences. The recordings were transcribed using Microsoft Teams automated transcription, stored within a university-managed cloud environment, and subsequently reviewed manually to correct errors and anonymise identifiers. Excerpts selected as illustrative quotations were translated from Japanese into English by the research team.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSystem logs and in-home engagement metrics (Japan)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe collected and analysed system logs to derive two engagement metrics (i) the proportion of full deployment days meeting the criterion for use (use-days) and (ii) mean daily dialogue-mode time (minutes/day). Use-days were defined as days with ≥20 minutes of recorded dialogue-mode interaction. Dialogue-mode time was defined as the duration from activation of dialogue mode to its termination, calculated from the corresponding timestamps. Engagement metrics were calculated across full deployment days, excluding the installation and removal days (partial days), and are presented descriptively.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eQuestionnaire assessing subjective impressions and usability of the robot\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSubjective impressions on the robot were assessed using the Godspeed Questionnaire Series (GQS), reported as five subscales (Anthropomorphism, Animacy, Likeability, Perceived Intelligence, and Perceived Safety). Items were rated on a 5-point scale, and we report mean subscale scores (range 1–5), with higher scores indicating more favourable perceptions(20). Usability was assessed using the 10-item System Usability Scale (SUS; with ‘system’ replaced by ‘robot’), scored using the standard 0–100 metric (higher scores indicate better usability)(21). To aid interpretation, SUS scores were contextualised descriptively using published guidance (e.g., ≥70 generally acceptable; 70–89 good; \u0026gt;90 excellent; \u0026lt;70 marginal; \u0026lt;50 potentially unacceptable), while recognising that SUS should not be used in isolation for absolute judgements(22). Participants completed these questionnaires within both UK and Japan studies. Questionnaire results are presented descriptively to contextualise the qualitative findings.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eQualitative data analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eUK focus group transcripts were managed and coded in NVivo (versions 14 and 15) and analysed using reflexive thematic analysis(23). Coding was led by YS, with MN acting as a second coder to support iterative refinement of codes and themes. Analysis involved repeated familiarisation, line-by-line coding, and development, review, and naming of themes through reflexive discussion within the research team.\u003c/p\u003e\n\u003cp\u003eJapan post-installation interview transcripts were managed in NVivo (version 15) and analysed using reflexive thematic analysis(23). We used the thematic structure generated from the UK dataset as an organising framework. Coding was led by YS, with KK contributing to the iterative development and refinement of a shared coding framework to support team-based analysis. Initial coding was sensitised by domains covered in the interview guide (usability/operational challenges, emotional responses during interaction, and perceived limitations), while allowing inductive refinement and theme development through ongoing analytic discussion.\u003c/p\u003e\n\u003cp\u003eTo support cross-site integration, the thematic structure generated from the larger UK dataset was used as an organising framework. Codes and themes from the Japan dataset were subsequently mapped onto this structure through discussion among the UK and Japan research team. Instances of divergence (e.g., between UK and Japan participants; focus groups versus in-home deployment; cognitively unimpaired participants versus those with MCI) were explored through data triangulation and team-based discussion and were preserved as part of the contextual interpretation within themes. .\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eParticipant characteristics\u003c/h2\u003e \u003cp\u003eIn the UK, 18 participants attended first-round focus groups; one participant aged 49 did not meet the age eligibility criteria and was excluded from the analytic sample. The UK sample comprised of 17 participants (mean age 72.5 years, SD 7.1; 61\u0026ndash;85), including 10 women and 7 men, with mean years of education 14.6 (SD 3.4; 10\u0026ndash;21). In Japan, eight participants were included (all female; mean age 81.1 years, SD 6.1; 73\u0026ndash;93). Diagnoses included probable MCI with Lewy Bodies (MCI-LB (n\u0026thinsp;=\u0026thinsp;5), and MCI due to Alzheimer\u0026rsquo;s Disease (AD) (n\u0026thinsp;=\u0026thinsp;3); mean Mini-Mental Status Examination (MMSE) was 25.4 (SD 0.9; 24\u0026ndash;27), all participants had Clinical Dementia Rating\u0026thinsp;=\u0026thinsp;0.5. Participant characteristics are summarised in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\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\u003eParticipant characteristics\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCharacteristic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUK focus groups (analytic sample)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eJapan home installation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, mean (SD); range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e72.5 (7.1); 61\u0026ndash;85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e81.1 (6.1); 73\u0026ndash;93\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10 (58.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8 (100%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEducation (years), mean (SD); range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14.6 (3.4); 10\u0026ndash;21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.8 (3.2); 9\u0026ndash;18\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLiving alone, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10 (58.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8 (100%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDigital Device Usage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSmartphone (n\u0026thinsp;=\u0026thinsp;17, 100%); Tablet (n\u0026thinsp;=\u0026thinsp;13, 76.5%); PC (n\u0026thinsp;=\u0026thinsp;14, 82.4%); Conversational agents (n\u0026thinsp;=\u0026thinsp;6, 35.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSmartphone (n\u0026thinsp;=\u0026thinsp;7, 87.5%); Tablet (n\u0026thinsp;=\u0026thinsp;1, 12.5%); PC (n\u0026thinsp;=\u0026thinsp;2, 25%); None (n\u0026thinsp;=\u0026thinsp;1, 12.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiagnosis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCommunity-dwelling; no dementia diagnosis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMCI-LB (n\u0026thinsp;=\u0026thinsp;5); MCI due to AD (n\u0026thinsp;=\u0026thinsp;3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMMSE, mean (SD); range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e25.4 (0.9); 24\u0026ndash;27\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCDR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.5 (all)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCDR-SB, mean (SD); range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.1 (1.2); 0.5\u0026ndash;4.0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"3\"\u003eUK focus groups represent the analytic sample. Abbreviations: UK, United Kingdom; MMSE, Mini-Mental State Examination; CDR, Clinical Dementia Rating; CDR-SB, Clinical Dementia Rating\u0026ndash;Sum of Boxes; MCI, mild cognitive impairment; MCI-LB, mild cognitive impairment with Lewy bodies; AD, Alzheimer\u0026rsquo;s disease. CDR was 0.5 for all participants in the Japan home installation cohort. \u0026lsquo;\u0026mdash;\u0026rsquo; indicates not collected/available.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eQuestionnaire measures and in-home engagement\u003c/h2\u003e \u003cp\u003eGQS and SUS scores are reported descriptively for both cohorts (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) to contextualise the qualitative findings. In the UK focus groups (n\u0026thinsp;=\u0026thinsp;17), mean (SD) GQS subscale scores were 2.9 (1.1) for Anthropomorphism, 3.2 (1.0) for Animacy, 3.8 (0.9) for Likeability, 3.4 (0.8) for Perceived Intelligence, and 3.9 (0.6) for Perceived Safety; mean SUS was 56.2 (11.4). In the Japan home installation cohort (n\u0026thinsp;=\u0026thinsp;8), corresponding GQS scores were 3.3 (0.9), 3.7 (0.6), 4.4 (0.8), 4.0 (0.6), and 3.6 (0.5), and mean SUS was 62.8 (18.8). In-home engagement in the Japan cohort indicated use on 78.6% of deployment days (use-day defined as \u0026ge;\u0026thinsp;20 minutes of recorded dialogue-mode interaction), and mean daily dialogue-mode time was 68.2 minutes/day (SD 41.7).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eQuestionnaire measures and in-home engagement\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\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=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMeasure\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUK focus groups (n\u0026thinsp;=\u0026thinsp;17)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eJapan home installation (n\u0026thinsp;=\u0026thinsp;8)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGQS Anthropomorphism, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.9 (1.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.3 (0.9)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGQS Animacy, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.2 (1.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.7 (0.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGQS Likeability, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.8 (0.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4.4 (0.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGQS Perceived Intelligence, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.4 (0.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4.0 (0.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGQS Perceived Safety, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.9 (0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.6 (0.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSUS score (0\u0026ndash;100), mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e56.2 (11.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e62.8 (18.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIn-home engagement (use-days, % of deployment days)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e72.4%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIn-home engagement (mean daily dialogue-mode time, minutes/day; SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e68.2 (41.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"3\"\u003eMean daily dialogue-mode time was calculated across full deployment days, excluding the installation and removal days (partial days). Days with no recorded dialogue were included and counted as 0 minutes. Use-days were defined as days with \u0026ge;\u0026thinsp;20 minutes of recorded dialogue-mode interaction. Abbreviations: GQS, Godspeed Questionnaire Series; SUS, System Usability Scale. GQS subscales are scored on 5-point Likert-type items (higher scores indicate more favourable perceptions). SUS is reported on a 0\u0026ndash;100 scale (higher scores indicate better usability). \u0026lsquo;\u0026mdash;\u0026rsquo; indicates not applicable/not assessed.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eQualitative findings\u003c/h2\u003e \u003cp\u003eIn the UK focus groups, we identified three overarching themes (with six sub-themes) capturing users\u0026rsquo; expectations and concerns about using an AI companion robot for loneliness support in later life: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) Practical Use and Functionality, (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) Emotional Connection and Engagement, and (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) Ethical and Societal Reflections (\u003cb\u003eSupplementary Table S1\u003c/b\u003e). Japan post-installation interviews were analysed independently using reflexive thematic analysis, yielding four themes : Becoming fluent in robot-mediated conversation; Maintaining agency over engagement and interruption; Constructing social presence through conversation and embodiment; and Negotiating boundaries and psychological load. For cross-site reporting, Japan themes/codes were subsequently mapped onto the UK-derived thematic framework (\u003cb\u003eSupplementary Table S2\u003c/b\u003e). Illustrative participant quotations are included to support each theme; UK focus group participants are labelled UK1\u0026ndash;UK18, and Japan participants are labelled JP1\u0026ndash;JP8.\u003c/p\u003e \u003cp\u003eTheme 1: Practical Use and Functionality\u003c/p\u003e \u003cp\u003eSubtheme 1.1. Operational Challenges and User Adaptation\u003c/p\u003e \u003cp\u003eAcross both contexts, participants considered that smooth interaction depended on basic usability features and learnable turn-taking. In the UK, brief hands-on exposure revealed that interaction latency and cue interpretation could disrupt conversational flow, with one participant noting: \u0026ldquo;You had to sit and wait for 10 seconds (for a response)\u0026mdash;it was 9 1/2 seconds too long. (UK2)\u0026rdquo; In addition to latency, participants sometimes struggled to understand when to speak or how to recover from breakdowns or interruptions to conversational flow. In the Japan home setting, participants similarly described uncertainty around timing and whether the robot had heard and understood their speech. Several reported adapting their own communication style over time (e.g., simplifying phrasing or choosing words more deliberately). One participant noted that a one‑week deployment limited this process: \u0026ldquo;This time it was only a week\u0026hellip; If it had been here all the time, I think I\u0026rsquo;d start to figure out.\u0026rdquo; (JP6). These accounts align with the descriptive SUS scores, indicating variable usability in both cohorts.\u003c/p\u003e \u003cp\u003eSubtheme 1.2. Suggestions and Expectations for Improved Design and Functionality\u003c/p\u003e \u003cp\u003eAcross both contexts, participants articulated design expectations oriented toward everyday usefulness, particularly features that would reduce effort and increase practical value beyond conversation. In the UK focus groups, participants discussed features that could make the robot more practically useful and better integrated with existing routines and technologies, rather than remaining a standalone novelty. Music playing was repeatedly framed as meaningful for loneliness support (e.g., \u0026ldquo;There are lonely people\u0026mdash;they like to listen to music. That\u0026rsquo;s important,\u0026rdquo; UK15). In the Japan study, improvement suggestions were particularly shaped by maintaining agency\u0026mdash;being able to decide when to engage and when to suspend interaction in the flow of household life (\u0026ldquo;When my daughter or the grandkids come over, I switch it off\u0026hellip; because it keeps talking the whole time\u0026rdquo;, JP7).\u003c/p\u003e \u003cp\u003eTheme 2: Emotional Connection and Engagement\u003c/p\u003e \u003cp\u003eSubtheme 2.1. Conversational Quality and Limitations\u003c/p\u003e \u003cp\u003eIn the UK focus groups, many participants reported that the robot\u0026rsquo;s conversational naturalness could be convincing at first (e.g., \u0026ldquo;It is definitely\u0026hellip; it makes you feel like you\u0026rsquo;re talking to somebody,\u0026rdquo; UK1). However, this was frequently tempered by perceived limitations in depth, specificity, and continuity, with responses sometimes experienced as generic or non-committal (e.g., \u0026ldquo;It sounds like a politician\u0026mdash;it won\u0026rsquo;t answer the question,\u0026rdquo; UK12). In Japan, conversational limitations were often described in terms of mismatch and interaction burden during everyday use. One participant noted intermittent misalignment: \u0026ldquo;There were times when the conversation didn\u0026rsquo;t quite line up\u0026hellip; maybe I didn\u0026rsquo;t explain well enough\u0026rdquo; (JP7). Another highlighted that the robot\u0026rsquo;s frequent, unsolicited talk could become excessive: \u0026ldquo;It doesn\u0026rsquo;t need to react every time I pass by\u0026hellip; It talks too much\u0026mdash;too much chatter\u0026rdquo; (JP1).\u003c/p\u003e \u003cp\u003eSubtheme 2.2. Expectations for the Robot as a Companion\u003c/p\u003e \u003cp\u003eAcross both contexts, participants\u0026rsquo; expectations of \u0026ldquo;companionship\u0026rdquo; were strongly tied to embodiment and social presence\u0026mdash;the sense that the robot is there as a responsive other, not merely a voice interface. In the UK, several participants suggested that physical movement and attentional signalling could make the robot feel more present than disembodied assistants (\u0026ldquo;The movement\u0026hellip; would be better than Alexa. Because it\u0026rsquo;s more sort of\u0026hellip; present,\u0026rdquo; UK11). Some also expected that companionability would depend on longer-term adaptation, including memory of prior conversations and preferences, enabling a relationship that could develop over time. In Japan, companion framing was often expressed through everyday relational talk directed toward the robot, akin to speaking to a pet or housemate. One participant described: \u0026ldquo;I talked to it like I would with a normal person\u0026hellip; Even when I went to day care, I\u0026rsquo;d say, \u0026lsquo;I\u0026rsquo;m heading to day care now,\u0026rsquo; or \u0026lsquo;Please watch the house\u0026rsquo;\u0026rdquo; (JP2). A small number also reported emerging attachment and reluctance to part after the one-week installation.\u003c/p\u003e \u003cp\u003eTheme 3: Ethical and Societal Reflections\u003c/p\u003e \u003cp\u003eSubtheme 3.1: Concerns about Access and Ethical Use\u003c/p\u003e \u003cp\u003eIn the UK focus groups, ethical acceptability was frequently discussed through the topics of privacy, profiling, and downstream misuse. Participants expressed unease that an AI companion could accumulate more intimate information than might be shared with clinicians and that such data could be repurposed or commercialised: \u0026ldquo;It\u0026rsquo;s going to have more accurate profiles on us\u0026hellip; would you want your psychologist to be selling your data?\u0026rdquo; (UK14). Several extended this to crime-related vulnerability, noting that older people with pensions and assets could become \u0026ldquo;perfect targets\u0026rdquo; if profiling enabled exploitation. Some extended this to heightened vulnerability to scams or exploitation, and a broader dislike or scepticism toward \u0026ldquo;unnecessary\u0026rdquo; technology. In the Japan home-installation interviews, comparatively few participants raised explicit access or ethical-use concerns. Aside from a single privacy-boundary comment noted in the codebook, discussion in this subtheme was limited; the present dataset does not allow firm attribution to sample, setting, or cultural factors.\u003c/p\u003e \u003cp\u003eSubtheme 3.2. Views on the Psychological Supportive Role of Robots\u003c/p\u003e \u003cp\u003eAcross both contexts, participants generally recognised potential benefit, but framed the robot\u0026rsquo;s role as conditional and situational, particularly for people who are isolated or have limited opportunities for social contact (e.g., \u0026ldquo;Someone can\u0026rsquo;t get out\u0026mdash;it could be a godsend,\u0026rdquo; UK13). In the UK focus groups, this conditional framing was paired with concern about how the technology might be positioned within services: \u0026ldquo;If this is a partial solution, I would hate to see it being the only solution\u0026rdquo; (UK2). In the Japan home-installation interviews, participants likewise described context-dependent usefulness (e.g., for someone living alone), while a small number also noted psychological load (e.g., post-interaction rumination affecting sleep, JP8). Overall, these accounts suggest that supportive value depends on aligning conversational depth and initiative with users\u0026rsquo; circumstances and capacity, rather than assuming uniform benefit.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this UK\u0026ndash;Japan qualitative study, participants in both component sub-studies broadly recognised the potential value of an LLM-enabled conversational robot as a conditional form of support for loneliness in later life, particularly for people who are isolated or have limited opportunities for social contact. The UK focus groups, conducted after brief hands-on interaction, generated active discussion around desired functions and usability, the social acceptability of using a robot for companionship, and ethical concerns such as privacy and potential misuse. In contrast, the Japan home-installation interviews (analysed independently into four themes: becoming fluent in robot-mediated conversation; maintaining agency over engagement and interruption; constructing social presence through conversation and embodiment; and negotiating boundaries and psychological load) foregrounded the practical realities of use at home\u0026mdash;especially the time needed to learn interaction routines and the importance of maintaining agency in everyday contexts. Together, the two components provide complementary insights for designing and governing LLM-enabled companion robots for older people. Engagement logs in the Japan cohort indicated use on 78.6% of full deployment days and a mean daily dialogue-mode time of 68.2 minutes/day (SD 41.7), suggesting that meaningful day-to-day use was feasible for most participants during the one-week deployment.\u003c/p\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003ePractical usability and the burden of adaptation\u003c/h2\u003e \u003cp\u003eFor a conversational companion robot to be sustainable in people\u0026rsquo;s homes, basic dialogue mechanics must come first: turn-taking stability, low end-to-end latency, and robust speech recognition. Our iterative refinement work found these are not merely desirable but essential prerequisites for perceived usability, consistent with LLM-robot reports that highlight both timing and ASR breakdowns as prominent barriers (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e). In the present study, participants likewise framed usability as something that requires effort and learning, indicating that shortcomings in timing and recognition translate into a burden of adaptation rather than a minor inconvenience (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eSecond, participants also expressed a wish for additional functions (e.g., music for comfort/enjoyment, multilingual support, and safety-oriented features such as contacting family after a fall), echoing the findings of our earlier home work (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e). However, adding functions can increase configuration demands and discoverability burdens; both our field experience and general usability evidence suggest that introducing new, untested components can reduce usability unless the interface and workflow are stabilised (\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e). A staged strategy therefore seems safer: establish a feasible \u0026ldquo;core companion\u0026rdquo; with reliable conversational fluency first, then expanding functionality incrementally with user testing\u0026mdash;especially when designing for users with cognitive impairment.\u003c/p\u003e \u003cp\u003eThird, sustained domestic use depends on context-sensitive control and agency: users need to be able to initiate, pause, silence, or limit interaction in ways that fit daily routines and social context. In our Japan home installation, \u0026ldquo;maintaining agency\u0026rdquo; emerged as a salient requirement (e.g., the need to suppress interaction when family members visit), suggesting that mode control and interruptibility are core usability features rather than optional extras. This complements acceptance models in older people, where perceived controllability and ease of use are central determinants of intention to use and continued use (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eEmotional engagement, companionship, and the importance of control\u003c/h2\u003e \u003cp\u003eA key advance in the present work is that introducing an LLM enabled a level of conversational naturalness that is difficult to achieve with scenario-based dialogue alone. For example, in our earlier home study with RoBoHoN, conversation was explicitly scenario-based and could not respond to more complex speech, highlighting inherent constraints of scripted pathways (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e). In contrast, participants in both the UK and Japan components described the LLM-enabled interaction as feeling close to human conversation (e.g., \u0026ldquo;it makes you feel like you\u0026rsquo;re talking to somebody,\u0026rdquo; and \u0026ldquo;it really felt like talking with a person\u0026rdquo;). This improvement is important because conversational fluency is a prerequisite for companionship, but it was also clear that fluency alone is not sufficient for this.\u003c/p\u003e \u003cp\u003eCompanionship was repeatedly linked to presence, namely whether the robot is experienced as a socially present counterpart, which embodiment plausibly supports (e.g., movement being \u0026ldquo;more\u0026hellip; present\u0026rdquo; than a disembodied assistant). Prior acceptance work similarly positions social presence as a pathway to more positive experience and higher acceptance of companion robots (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e). The Japan home installation further suggested that repeated everyday contact may foster attachment (including reluctance to part after only one week), echoing longitudinal HRI evidence that attachment-related responses can emerge with ongoing exposure, while remaining context-dependent rather than inevitable (\u003cspan additionalcitationids=\"CR26\" citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eDespite high ratings of naturalness, UK participants frequently pointed to insufficient depth (generic or non-committal answers and limited contextual continuity). This raises a design question: whether depth is primarily a limitation of open-domain LLM chat, or can be improved through prompt design and system-level scaffolding. Dialogue research has argued that unconstrained \u0026ldquo;chat about anything\u0026rdquo; often produces shallow interaction unless systems establish common ground and structured purposes (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e). For LLM-enabled companion robots, recent work likewise emphasises that improving perceived depth likely requires architecture-level strategies (e.g., grounding, retrieval, structured interaction, and memory) rather than generation alone (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e). At the same time, increasing \u0026ldquo;personality\u0026rdquo; or opinionatedness to deepen companionship introduces ethical tension: stronger social presence and more assertive dialogue may increase persuasive influence, and sustained emotionally intense engagement with conversational AI has been discussed as a potential risk factor for delusion-like experiences in vulnerable individuals (\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e), underscoring the need for safety-by-design and clear governance in socially assistive AI systems(\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFinally, Japan participants\u0026rsquo; concerns about the robot talking too much highlight a complementary design challenge: unsolicited talk can feel intrusive in everyday home contexts, particularly when users are moving around or when other people are present and may undermine long-term acceptability. This reinforces recommendations that talkativeness and initiative should be adjustable to context and user preference, with clear controls for pausing, silencing, and interrupting interaction rather than treated as uniformly desirable (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e). In-home engagement also showed substantial interaction time, suggesting the importance of calibrating interaction intensity so that it remains comfortable rather than burdensome.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eEthical and psychological appropriateness in psychogeriatric contexts\u003c/h2\u003e \u003cp\u003eBroader ethical concerns were salient, particularly around privacy, data governance, equitable access, and the fear that technology might be used to substitute for human contact. Such concerns are well recognised in the sociology and ethics of care robotics, including anxieties about deception, dependence, and the reconfiguration of care obligations (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). These issues matter acutely in psychogeriatrics because loneliness interventions often operate within care ecosystems characterised by resource constraints, service gaps, and vulnerability to inappropriate substitution. Participants\u0026rsquo; accounts suggest that acceptability depends on clear framing of companion robots as adjunctive supports\u0026mdash;potentially helpful in specific contexts (e.g., living alone, limited day-to-day contact)\u0026mdash;rather than an \u0026ldquo;only solution.\u0026rdquo;\u003c/p\u003e \u003cp\u003eA particularly clinically relevant contribution from the Osaka in-home interviews was the articulation of psychological intensity and interaction burden. Some participants described pressure to respond \u0026ldquo;seriously\u0026rdquo; or experiencing discomfort when the robot\u0026rsquo;s questioning felt as though it was \u0026ldquo;seeing through\u0026rdquo; them, with effects that could persist beyond the interaction (e.g., rumination). Together, these findings suggest that generative AI systems, if not carefully calibrated, may drift toward a quasi-therapeutic or probing interaction style that some users experience as intrusive\u0026mdash;particularly those with cognitive impairment or heightened anxiety. In practice, this supports several safeguards: transparent role framing (what the robot is and is not), user-selectable conversation modes (light chat versus reflective talk), explicit consent cues before sensitive topics, and de-escalation pathways when distress is detected. These design priorities also align with broader discussions of foundation-model risks\u0026mdash;where hallucination, opacity, and miscalibrated social persuasion are salient\u0026mdash;and underscore the importance of risk-calibrated oversight and conservative defaults in clinical-adjacent deployments (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e). It is also possible that differences in study format and participant characteristics (cognitively unimpaired focus groups vs brief in-home interviews with outpatients with MCI) shaped what was foregrounded, with broader ethical considerations potentially less salient or harder to articulate in the latter context. This further supports conservative defaults and clear governance in psychogeriatric deployments, where downstream privacy, profiling, or substitution risks may be consequential even when not spontaneously raised.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eStrengths and limitations\u003c/h2\u003e \u003cp\u003eThis study integrates perspectives from two complementary contexts\u0026mdash;hands-on group interaction in the UK and situated in-home use in Japan\u0026mdash;includes a clinical outpatient sample (MCI), and combines qualitative accounts with brief subjective measures and planned in-home log metrics.\u003c/p\u003e \u003cp\u003eKey limitations should be noted. First, recruitment occurred through specific channels, limiting generalisability and increasing the likelihood of selection bias. Second, the UK and Japan components differed substantially in clinical characteristics, cultural context, and study format (hands-on focus groups vs in-home deployment; cognitively unimpaired participants vs outpatients with MCI); while informative, these differences constrain direct comparison and mean that observed differences should not be attributed to any single factor. Third, the Osaka cohort was small and clinically heterogeneous (MCI due to Alzheimer\u0026rsquo;s disease and Lewy body disease), and the one-week deployment cannot establish longer-term trajectories such as habituation, sustained attachment, disengagement, or adverse psychological effects. Fourth, qualitative findings are sensitive to interview context and phrasing; cross-language reporting adds additional interpretive layers because only selected excerpts were translated. Although this was mitigated through careful anonymisation and team review, future work would benefit from more formalised translation procedures and bilingual audit trails.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eGenerative AI-supported companion robots may provide older people with interactions relevant to loneliness support, but acceptability depends on practical usability, user control, and calibration of conversational depth and initiative. Cross-context user perspectives\u0026mdash;particularly from in-home use among individuals with MCI\u0026mdash;highlight design tensions and safeguards needed for clinically responsible implementation. Clinically, these findings support cautious optimism for conditional use where social contact is limited, provided privacy governance and psychological appropriateness are addressed. Future studies should test longer deployments and evaluate clinically meaningful outcomes, while examining how design choices influence both benefit and burden.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eConflict of interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAuthors AT, IE and KU are employees of NTT West, which provided in-kind support for this study, including the loan of the Sota robot(s), server infrastructure, and payment of Microsoft Azure OpenAI service usage fees. The remaining authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDescription of author’s role\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eYS designed the studies, led participant recruitment, and conducted data collection and analysis. YS also wrote the first draft of the manuscript. MN, CY, KK and IE supported data collection and data curation, with CY and KK additionally assisting with recruitment. IE, AT and KU developed the robot’s dialogue system and contributed to improvements in user experience. PR provided oversight of qualitative data collection and analysis. NB contributed to the ethics approval process and advised on refinements to the robot’s UX design. MI obtained funding and supervised the study design. RH provided overall project leadership and supervision and guided manuscript development. All authors critically reviewed the manuscript and approved the final version.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the SENSHIN Medical Research Foundation (Overseas Study grant), Daiichi-Sankyo ”Habataku” Support Program for the Next Generation of Researchers and by the EPSRC-funded Sensing to Collaboration Programme Grant (EP/V000748/1).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of generative AI and AI-assisted technologies in the manuscript preparation process\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDuring the preparation of this work, the authors used ChatGPT (OpenAI) to support language editing and manuscript organisation. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eHolt-Lunstad J (2024) Social connection as a critical factor for mental and physical health: evidence, trends, challenges, and future implications. World Psychiatry 23(3):312\u0026ndash;332. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/wps.21224\u003c/span\u003e\u003cspan address=\"10.1002/wps.21224\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee SL, Pearce E, Ajnakina O, Johnson S, Lewis G, Mann F et al (2021) The association between loneliness and depressive symptoms among adults aged 50 years and older: a 12-year population-based cohort study. Lancet Psychiatry 8(1):48\u0026ndash;57. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S2215-0366(20)30383-7\u003c/span\u003e\u003cspan address=\"10.1016/S2215-0366(20)30383-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSalari N, Najafi H, Rasoulpoor S, Canbary Z, Heidarian P, Mohammadi M (2025) The global prevalence and associated factors of loneliness in older adults: a systematic review and meta-analysis. Humanit Soc Sci Commun 12(1):985. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1057/s41599-025-05304-x\u003c/span\u003e\u003cspan address=\"10.1057/s41599-025-05304-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLuhmann M, Hawkley LC (2016) Age Differences in Loneliness from Late Adolescence to Oldest Old Age. Dev Psychol 52(6):943\u0026ndash;959. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1037/dev0000117\u003c/span\u003e\u003cspan address=\"10.1037/dev0000117\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGardiner C, Geldenhuys G, Gott M (2018) Interventions to reduce social isolation and loneliness among older people: an integrative review. Health Soc Care Community 26(2):147\u0026ndash;157. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/hsc.12367\u003c/span\u003e\u003cspan address=\"10.1111/hsc.12367\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGilbody S, Littlewood E, McMillan D, Atha L, Bailey D, Baird K et al (2024) Behavioural activation to mitigate the psychological impacts of COVID-19 restrictions on older people in England and Wales (BASIL+): a pragmatic randomised controlled trial. Lancet Healthy Longev 5(2):e97\u0026ndash;107. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S2666-7568(23)00238-6\u003c/span\u003e\u003cspan address=\"10.1016/S2666-7568(23)00238-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHoang P, King JA, Moore S, Moore K, Reich K, Sidhu H et al (2022) Interventions Associated With Reduced Loneliness and Social Isolation in Older Adults: A Systematic Review and Meta-analysis. JAMA Netw Open 5(10):e2236676. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1001/jamanetworkopen.2022.36676\u003c/span\u003e\u003cspan address=\"10.1001/jamanetworkopen.2022.36676\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGasteiger N, Loveys K, Law M, Broadbent E (2021) Friends from the Future: A Scoping Review of Research into Robots and Computer Agents to Combat Loneliness in Older People. Clin Interv Aging 16:941\u0026ndash;971. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2147/CIA.S282709\u003c/span\u003e\u003cspan address=\"10.2147/CIA.S282709\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePu L, Moyle W, Jones C, Todorovic M (2019) The Effectiveness of Social Robots for Older Adults: A Systematic Review and Meta-Analysis of Randomized Controlled Studies. Gerontologist 59(1):e37\u0026ndash;51. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/geront/gny046\u003c/span\u003e\u003cspan address=\"10.1093/geront/gny046\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSatake Y, Costello H, Naran N, Ishimaru D, Ikeda M, Howard R (2026) Autonomous conversational agents for loneliness, social isolation, depression, and anxiety in older people without cognitive impairment: Systematic review and meta-analysis. Psychol Med 56:e27. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1017/S0033291725103073\u003c/span\u003e\u003cspan address=\"10.1017/S0033291725103073\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNagata Y, Satake Y, Yamazaki R, Nishio S, Ishiguro H, Ikeda M (2026) Case report: A five\u0026ndash;case series of 18\u0026thinsp;\u0026ndash;\u0026thinsp;month home use of a conversational companion robot for psychological support in older people with mild cognitive impairment or late\u0026ndash;onset psychosis. Front Psychiatry 16:1700340. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fpsyt.2025.1700340\u003c/span\u003e\u003cspan address=\"10.3389/fpsyt.2025.1700340\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNagata Y, Satake Y, Yamazaki R, Nishio S, Suzuki M, Kanemoto H et al (2025) A Conversational Robot for Cognitively Impaired Older People Who Live Alone: An Exploratory Feasibility Study. Psychogeriatrics 25(5):e70076. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/psyg.70076\u003c/span\u003e\u003cspan address=\"10.1111/psyg.70076\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerridge C, Zhou Y, Robillard JM, Kaye J (2023) Companion robots to mitigate loneliness among older adults: Perceptions of benefit and possible deception. Front Psychol 14:1106633. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fpsyg.2023.1106633\u003c/span\u003e\u003cspan address=\"10.3389/fpsyg.2023.1106633\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHeerink M, Kr\u0026ouml;se B, Evers V, Wielinga B (2010) Assessing Acceptance of Assistive Social Agent Technology by Older Adults: the Almere Model. Int J Soc Robot 2(4):361\u0026ndash;375. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s12369-010-0068-5\u003c/span\u003e\u003cspan address=\"10.1007/s12369-010-0068-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBender EM, Gebru T, McMillan-Major A, Shmitchell S (2021) On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 列. In: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency [Internet]. Virtual Event Canada: ACM; pp. 610\u0026ndash;23. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/3442188.3445922\u003c/span\u003e\u003cspan address=\"10.1145/3442188.3445922\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBommasani R, Hudson DA, Adeli E, Altman R, Arora S, von Arx S et al (2022) On the Opportunities and Risks of Foundation Models [Internet]. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.2108.07258\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2108.07258\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. arXiv\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIrfan B, Kuoppam\u0026auml;ki S, Hosseini A, Skantze G (2025) Between reality and delusion: challenges of applying large language models to companion robots for open-domain dialogues with older adults. Auton Robots 49(1):9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s10514-025-10190-y\u003c/span\u003e\u003cspan address=\"10.1007/s10514-025-10190-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIrfan B, Kuoppam\u0026auml;ki S, Skantze G (2024) Recommendations for designing conversational companion robots with older adults through foundation models. Front Robot AI 11:1363713. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/frobt.2024.1363713\u003c/span\u003e\u003cspan address=\"10.3389/frobt.2024.1363713\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSatake Y, Ishimaru D, Katsuki K, Rapaport P, Berthouze N, Howard R et al (2025) A Week With a Conversational Large Language Model Companion Robot. Am J Geriatr Psychiatry 33(7):799\u0026ndash;800. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jagp.2025.03.010\u003c/span\u003e\u003cspan address=\"10.1016/j.jagp.2025.03.010\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBartneck C, Kulić D, Croft E, Zoghbi S (2009) Measurement Instruments for the Anthropomorphism, Animacy, Likeability, Perceived Intelligence, and Perceived Safety of Robots. Int J Soc Robot 1(1):71\u0026ndash;81. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s12369-008-0001-3\u003c/span\u003e\u003cspan address=\"10.1007/s12369-008-0001-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrooke J (1996) SUS - A quick and dirty usability scale. Usability Evaluation in Industry. Taylor \u0026amp; Francis, pp 189\u0026ndash;194\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBangor A, Kortum PT, Miller JT (2008) An Empirical Evaluation of the System Usability Scale. Int J Hum-Comput Interact 24(6):574\u0026ndash;594. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/10447310802205776\u003c/span\u003e\u003cspan address=\"10.1080/10447310802205776\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBraun V, Clarke V (2021) Thematic analysis: a practical guide. SAGE, London; Thousand Oaks, California, p 376\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHeerink M, Kr\u0026ouml;se B, Wielinga B, Evers V (2008) Enjoyment intention to use and actual use of a conversational robot by elderly people. In: Proceedings of the 3rd ACM/IEEE international conference on Human robot interaction [Internet]. New York, NY, USA: Association for Computing Machinery; pp. 113\u0026ndash;20. (HRI \u0026rsquo;08). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/1349822.1349838\u003c/span\u003e\u003cspan address=\"10.1145/1349822.1349838\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDziergwa M, Kaczmarek M, Kaczmarek P, Kędzierski J, Wadas-Szydłowska K (2018) Long-Term Cohabitation with a Social Robot: A Case Study of the Influence of Human Attachment Patterns. Int J Soc Robot 10(1):163\u0026ndash;176. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s12369-017-0439-2\u003c/span\u003e\u003cspan address=\"10.1007/s12369-017-0439-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan Maris A, Zook N, Caleb-Solly P, Studley M, Winfield A, Dogramadzi S (2020) Designing Ethical Social Robots\u0026mdash;A Longitudinal Field Study With Older Adults. Front Robot AI 7:1. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/frobt.2020.00001\u003c/span\u003e\u003cspan address=\"10.3389/frobt.2020.00001\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYamazaki R, Nishio S, Nagata Y, Satake Y, Suzuki M, Kanemoto H et al (2023) Long-term effect of the absence of a companion robot on older adults: A preliminary pilot study. Front Comput Sci 5:1129506. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fcomp.2023.1129506\u003c/span\u003e\u003cspan address=\"10.3389/fcomp.2023.1129506\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSkantze G, Doğru\u0026ouml;z AS (2023) The Open-domain Paradox for Chatbots: Common Ground as the Basis for Human-like Dialogue. In: Stoyanchev S, Joty S, Schlangen D, Dusek O, Kennington C, Alikhani M, editors. Proceedings of the 24th Annual Meeting of the Special Interest Group on Discourse and Dialogue [Internet]. Prague, Czechia: Association for Computational Linguistics; pp. 605\u0026ndash;14. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18653/v1/2023.sigdial-1.57\u003c/span\u003e\u003cspan address=\"10.18653/v1/2023.sigdial-1.57\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePan Z, Wu Q, Jiang H, Luo X, Cheng H, Li D et al (2025) On Memory Construction and Retrieval for Personalized Conversational Agents [Internet]. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/ARXIV.2502.05589\u003c/span\u003e\u003cspan address=\"10.48550/ARXIV.2502.05589\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. arXiv\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHudon A, Stip E (2025) Delusional Experiences Emerging From AI Chatbot Interactions or AI Psychosis. JMIR Ment Health 12(1):e85799. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2196/85799\u003c/span\u003e\u003cspan address=\"10.2196/85799\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRanisch R, Haltaufderheide J (2025) Rapid Integration of LLMs in Healthcare Raises Ethical Concerns: An Investigation into Deceptive Patterns in Social Robots. Digit Soc 4(1):7. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s44206-025-00161-2\u003c/span\u003e\u003cspan address=\"10.1007/s44206-025-00161-2\" 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":"University College London","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":"loneliness, robots, large language models, AI, mild cognitive impairment, community-dwelling, qualitative research","lastPublishedDoi":"10.21203/rs.3.rs-9093780/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9093780/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eBackground: Loneliness is a critical social issue in older adults, with clinical implications. Conversational companion robots have been studied as one approach to ameliorating this issue. Large language models (LLMs) can enable flexible conversational ability in companion robots, but acceptability and suitability remain uncertain. We explored older people’s expectations and concerns regarding an LLM-supported companion robot for loneliness support.\u003c/p\u003e\n\u003cp\u003eMethods: We conducted a UK–Japan qualitative study comprising hands-on focus groups for community-dwelling older adults in London (n=17) and a one-week in-home use with follow-up interviews in Osaka among outpatients with mild cognitive impairment (MCI; n=8). Transcripts were analysed using reflexive thematic analysis; for cross-site reporting, Japan themes/codes were mapped onto the thematic structure generated from the larger UK dataset. Descriptive questionnaire measures and at-home conversational log metrics were collected to contextualise qualitative findings.\u003c/p\u003e\n\u003cp\u003eResults: Participants saw value of the companion robot as a support for older people with loneliness but emphasised that acceptability depends on interaction mechanics and user agency. Three cross-context themes were identified: (1) Practical use and functionality (response latency, turn-taking, desired features, and controllability in home use); (2) Emotional connection and engagement (social presence alongside perceived limits in conversational fit and depth); and (3) Ethical and societal reflections (privacy/data governance, access, and concerns about substituting for human contact).\u003c/p\u003e\n\u003cp\u003eConclusions: LLM-supported companion robots may provide acceptable low-intensity support for some older people, including those with MCI, provided that usability, user-adjustable control and ethical governance are prioritised. Longer deployments are needed to evaluate potential sustained benefit and burden.\u003c/p\u003e","manuscriptTitle":"Large language model-supported companion robots for loneliness in older people: A UK–Japan qualitative study integrating focus groups and in-home deployment","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-13 15:17:31","doi":"10.21203/rs.3.rs-9093780/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":"e84dd574-55d7-4e24-a500-eaea691a1d61","owner":[],"postedDate":"March 13th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-13T15:17:31+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-13 15:17:31","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9093780","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9093780","identity":"rs-9093780","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}