Can a social robot improve older adults’ wellbeing?

Two phases, with very different denominators

In the first phase, 55 Japanese adults aged 65 to 74 rated imagined encouraging responses in an online scenario. The researchers used those ratings to select 'offering specific actions' as a high-perceived-efficacy strategy and 'distraction' as a lower-efficacy strategy. This phase identified candidate conversation styles; participants did not interact physically with the robot, and their age range differed from the later experimental group.

For the six-week phase, 21 people were recruited, one did not pass screening and 20 were randomised equally to four cells: human partner with higher-efficacy encouragement, human with lower-efficacy encouragement, robot with higher-efficacy encouragement and robot with lower-efficacy encouragement. Two participants withdrew. The final analysis included 18 people—eight men and 10 women—aged 62 to 90, with a mean age of 74.7. Final cell sizes were four, four, five and five.

Participants held weekly conversations. Early sessions focused on relationship-building and reminiscence; weeks three to six invited discussion of worries and the assigned form of encouragement. The robot was Sharp's RoBoHoN. The study combined scripted structure with participant diaries, interviews and open-ended responses. That mixed approach is useful for learning how an intervention feels, but it does not compensate statistically for four or five observations in each comparison cell.[1]

What changed on the two wellbeing measures

Researchers measured affective wellbeing with the five-item World Health Organization index, WHO-5, and cognitive evaluation of life with the Satisfaction With Life Scale, SWLS. Scores were recorded before the intervention and at weeks two, four and six. These are recognised self-report measures, but they are outcomes reported by participants who knew whether they were talking to a person or a robot; the study was not blinded.

On WHO-5, the average pre-to-week-six change was plus 2.3 points for human/high encouragement, plus 1.8 for human/low, plus 4.2 for robot/high and plus 1.4 for robot/low. Yet the robot/high group began at 15.4, compared with 18.4 in robot/low, and finished at 19.6 versus 19.8. The apparently larger rise may therefore reflect its lower starting point rather than a superior intervention.

On SWLS, the corresponding mean changes were plus 3.3, plus 1.3, plus 2.2 and plus 2.0 points. Every group moved upward on average. The researchers appropriately did not report between-group inferential tests, because the sample was too small for credible comparisons. These descriptive trajectories can generate hypotheses, but they cannot tell us whether any difference is larger than sampling noise or meaningful to an individual older person.[1]

Why the missing control changes the interpretation

There was no group receiving no weekly conversation, usual activity alone or neutral listening without encouragement. As a result, improvement could reflect being heard, expecting benefit, repeated measurement, unrelated life events or a return toward typical scores after a low baseline. The study also cannot cleanly separate the identity of the conversation partner from the assigned encouragement strategy with such small cells.

Human delivery was uneven. The two human partners were both women, and one conducted seven participants' sessions while the other conducted one. Their warmth, pacing and skill may have mattered. On the technology side, results from one compact robot and one dialogue design do not establish a general effect for social robots or conversational AI. A different voice, embodiment, error rate or privacy practice could change both engagement and harm.

The participants were generally living in the community and many were in comparatively good health or already involved in social activities. People with cognitive impairment, severe depression, hearing difficulties, limited Japanese fluency or acute social-care needs may respond differently and may require safeguards not tested here. No clinical diagnosis or treatment outcome was evaluated.[1]

What the qualitative material can—and cannot—add

Diaries and interviews can reveal why a conversation felt supportive or awkward. They may help developers understand whether specific suggestions invite action, whether distraction feels dismissive, and when repetition or limited responses expose the robot's constraints. In an exploratory study, that information is valuable for refining scripts and deciding what a larger trial should measure.

Qualitative enthusiasm is not a clinical effect size. Participants may value novelty, attention from researchers or the ritual of a weekly appointment. People who withdrew may differ from those who completed the study. Because the paper explores sustained interaction over six weeks rather than a one-off encounter, it moves beyond a laboratory demonstration, but six weeks is still too short to establish durable wellbeing gains, reduced service use or protection against isolation.[1]

Implications for families, clinicians and public services

A cautious deployment would position a robot as an optional conversational aid, not a substitute for a visitor, therapist, nurse or social worker. Older adults should be able to decline it without losing services. They need a plain explanation of what is recorded, where data go, who can review conversations and what happens when the system misunderstands. A device that elicits worries also needs an escalation path for distress, abuse, self-harm, medication problems or urgent health symptoms.

Clinicians and commissioners should demand outcomes beyond engagement minutes. Useful trials would measure loneliness, mood, quality of life, adverse experiences, dropout, human-contact displacement and caregiver burden, with prespecified thresholds for meaningful change. Costs, maintenance, connectivity and staff time matter too. A robot that appears inexpensive but shifts monitoring work onto family members or fails in homes with poor connectivity may widen rather than reduce inequality.

For families, the study offers no basis to reduce contact because a device is present. The more constructive use is to test whether a tool helps a person prepare topics, remember an activity or bridge gaps between human visits. The person—not the device or provider—should retain control over when interaction starts and stops.[1]

Funding, disclosure and what would change the assessment

The study was supported by JST SPRING grant JPMJSP2128 and a Waseda University laboratory work fee. The authors declared no commercial or financial conflicts. Waseda University's ethics committee approved the work under reference 2024-603. The paper reports that GPT-5.2 was used for English grammar editing. These disclosures are useful, but ethical approval and peer review do not turn exploratory descriptive changes into causal evidence.

Confidence would change with a preregistered, adequately powered trial that includes a no-intervention or attention-matched control, balances human partners, masks outcome assessment where possible and reports both absolute scores and clinically meaningful change. It should recruit people with varied health, housing, income and digital experience, follow them for longer and record adverse effects and displacement of human contact. Until then, this pilot supports further research and careful co-design—not a claim that social robots improve older adults' wellbeing.[1]