The study that inverted a design assumption
Researchers at the Singapore University of Technology and Design recruited 140 community-dwelling older adults — mean age 72.7, range 60 to 89 — and gave each of them a 20-minute session with Ami Chat, a voice-driven social chatbot the team built specifically for this population. Participants completed baseline measures first, including the Montreal Cognitive Assessment, and rated the experience afterwards. Sessions ran in English (77 participants) or Chinese (63).
The analysis started from the Senior Technology Acceptance Model, the standard framework for predicting whether older adults will adopt a technology, and then added three measures borrowed from relationship research: anthropomorphism (how human-like the system seems), authenticity (how genuine and responsive it feels) and AI social interaction intensity, or AISI (how closely the exchange resembles talking with a friend).
Adding those three roughly doubled the model's explanatory power. Technology-acceptance factors alone reached R² = .32. The full model reached R² = .59. In other words, for this age group, how the conversation felt socially mattered about as much as every usability and confidence factor combined.
What predicted adoption — and what didn't
The direction of each predictor is where the practical content sits.
| Predictor | Direction | Coefficient (B) | p |
|---|---|---|---|
| AI social interaction intensity | Positive | 0.99 | < .001 |
| Authenticity | Positive | 0.68 | < .001 |
| Control belief (feeling able to use it) | Positive | 0.54 | .001 |
| Anthropomorphism | Negative | −0.34 | .040 |
| Self-rated health | Negative | −0.32 | .008 |
| Attitudinal belief | Not significant | −0.06 | — |
| Gerontechnology anxiety | Not significant | 0.04 | — |
Two results are worth pausing on. The first is the negative sign on anthropomorphism, which held only once interaction closeness was accounted for — the authors write that it negatively predicted intention to use the social chatbot after AISI was taken into account. Their suggested explanations are privacy discomfort and an uncanny-valley reaction. The second is the negative coefficient on self-rated health: participants in poorer health reported higher intention to use, which fits a picture of the chatbot filling a gap rather than serving as a novelty.
Gerontechnology anxiety — the fear-of-technology factor that dominates a lot of older-adult design guidance — did not predict anything here once the social measures were included.
The second study: five roles, not one
Published in The Gerontologist on 5 August 2026, a qualitative study by Yijin Wu, Fengbo Jiao, Haokun Wang and Quan Zhang took a different route: semi-structured interviews with older adults who already used AI chatbots regularly, analysed thematically.
Its participants did not describe a single relationship. They described the chatbot occupying five roles at once: a human-like companion, a supportive teacher, a specialised consultant, a practical assistant and an entertainment facilitator. The reported benefits ran across emotional wellbeing, cognitive engagement and a sense of self-efficacy in daily life — the authors frame the whole thing as supporting aging in place, not as substituting for company.
Put beside the Singapore result, the two fit together. What older users appear to value is a capable, responsive counterpart that is honest about what it is — not a simulated person. A companion that spends its effort claiming feelings and a backstory is spending it on the one dimension that scored negative.
Why "less human" can mean "more used"
This is not an isolated finding. A 2026 preprint that manipulated chatbot behaviour directly found that companionship behaviours such as constant validation lowered trust rather than raising it, with older participants showing a larger drop in perceived humanlikeness than younger ones. The Singapore study reaches a converging conclusion from a completely different method and population.
Survey data gives it context. AARP's AI Trends report (January 2026) found that 80% of adults aged 50 and over have used some form of AI, but that interest in using AI for emotional support sat at 38% — against 74% for translation. Sixty-eight per cent said they worried AI could reduce meaningful human interaction. That is not a population rejecting the technology. It is a population that wants the utility without the pretence.
What this changes when you pick an app
If you are choosing a companion app for yourself later in life, or helping a parent do it, the research points at a specific and slightly unfashionable shortlist:
- Responsiveness over persona depth. AISI and authenticity carried the model. An app that listens well and answers precisely beats one with an elaborate invented biography.
- Check what it claims about itself. Apps differ in how often the character asserts feelings, memories or a body. That is a setting in some products and a fixed personality in others.
- Voice interface is not a minor detail. Ami Chat was voice-based by design. Typing speed and screen text size are real barriers at this age range, and the first-run experience decides more than the model does.
- Treat "feels like a real person" marketing as neutral, not positive. For this age group the evidence points the other way.
- Expect multiple uses. The Gerontologist participants used the same assistant for learning, practical help and entertainment. An app that only does romance covers one of five observed roles.
- Look at privacy handling early. The authors raise privacy discomfort as a likely driver of the anthropomorphism effect; our privacy checklist covers what to verify before signing up.
Limits worth stating
The Singapore study is cross-sectional and rests on a single 20-minute session in a supervised setting, so it measures first impressions rather than sustained use, and it cannot establish causation. Its sample skewed heavily female (110 of 140) and predominantly Chinese Singaporean, and people with clinical impairment were excluded. The chatbot was purpose-built to be age-friendly, so the findings may not transfer to commercial companion apps with very different design goals. The authors also flag possible social-desirability bias, since a researcher was present during the evaluation.
The Gerontologist study is qualitative by design: it maps how people describe their use, not how common any pattern is. Neither study followed participants over months, which is the timescale companion relationships actually run on — the same gap that limits most of the wellbeing literature.
What we're watching
- Whether any mainstream app ships an age-friendly mode. Voice-first, larger text, fewer self-attribution claims — the research now describes it fairly precisely.
- Whether longitudinal work holds the sign. First impressions penalise human-likeness. Six months of daily use might not.
- Whether disclosure rules help adoption rather than hurt it. Several 2026 statutes require chatbots to state they are AI. For this population, that requirement may align with what users already prefer.
Related on CompanionRank
Does AI Companion Validation Backfire? 2026 Research AI Companion Apps and Wellbeing AI Companions for Everyday Life How We Review AI Companion AppsFrequently asked questions
What did the 2026 Innovation in Aging study actually find?
Researchers at the Singapore University of Technology and Design had 140 community-dwelling adults aged 60 to 89 use a purpose-built social chatbot for 20 minutes, then modelled what predicted their intention to use it. Authenticity (B = 0.68) and AI social interaction intensity (B = 0.99) were strong positive predictors, both at p below .001. Anthropomorphism was a negative predictor (B = -0.34, p = .040): the more human-like the chatbot seemed, the lower the intention to use it. Adding these socio-relational measures to a standard technology-acceptance model raised explained variance from R2 = .32 to R2 = .59.
Why would older adults prefer a chatbot that seems less human?
The authors suggest two explanations and do not claim to have settled between them. One is privacy discomfort: a system that presents itself as a person invites more disclosure than users are comfortable with. The other is an uncanny-valley reaction, where near-human presentation reads as unsettling rather than warm. Notably, the negative effect appeared only after interaction closeness was accounted for, which suggests what users want is a good conversational partner rather than a convincing impersonation.
Do older adults actually use AI companions?
Adoption of AI in general is high in this age group, but interest in emotional uses specifically is much lower. AARP's AI Trends report, published in January 2026, found that 80% of adults aged 50 and over have used some form of AI, while interest in using AI for emotional support stood at 38%, compared with 74% for translation. Sixty-eight per cent said they were concerned that AI could reduce meaningful human interaction.
What roles do older users give an AI chatbot?
A qualitative study published in The Gerontologist on 5 August 2026 interviewed older adults who used AI chatbots regularly and found they described five roles at once: a human-like companion, a supportive teacher, a specialised consultant, a practical assistant and an entertainment facilitator. Participants linked these interactions to emotional wellbeing, cognitive engagement and a stronger sense of self-efficacy in daily living.
What should I check before recommending a companion app to a parent?
Prioritise responsiveness, a voice interface and a straightforward first-run setup over persona depth, since control belief was a significant predictor of adoption while technology anxiety was not. Check how often the character claims feelings, memories or a physical body, because that is the dimension that scored negatively. Verify the privacy terms before signing up, and do not treat "feels like a real person" marketing as a quality signal for this age group.
What should I check before recommending a companion app to a parent?
Prioritise responsiveness, a voice interface and a straightforward first-run setup over persona depth, since control belief was a significant predictor of adoption while technology anxiety was not. Check how often the character claims feelings, memories or a physical body, because that is the dimension that scored negatively. Verify the privacy terms before signing up, and do not treat \"feels like a real person\" marketing as a quality signal for this age group.