Do emotional companion robots help? The evidence, with effect sizes. Offbrand Human

An emotional companion robot will measurably reduce loneliness in an older adult. That claim now has a real evidence base under it: five research syntheses published between 2024 and 2026, dozens of trials, a few thousand participants between them. Almost everything else these machines are marketed on is unsupported, and the distance between the two is where families waste money.

Outcome by outcome, what pooling the studies gives you

Outcome	Pooled effect	Source	How to read it
Loneliness	-0.59, 95% CI -0.92 to -0.26 (19 studies, 1,083 people)	The Gerontologist, 2025	Large, probably somewhat inflated
Loneliness	SMD -0.69, 95% CI -0.89 to -0.40 (16 studies, 781 people)	Arch Gerontol Geriatr, Jan 2026	Independent replication of the above
Depression	SMD -0.42, 95% CI -0.72 to -0.12	Same 2026 review	Modest, statistically real
Quality of life	SMD 0.24, 95% CI -1.0 to 0.58	Same 2026 review	Nothing demonstrated
Happiness	SMD 0.63, 95% CI -1.09 to 2.35	Same 2026 review	Nothing demonstrated
Depression and loneliness in long-term care	Significant, described as large (8 RCTs)	JAMDA, Jun 2024	The setting with the best results
Quality of life, anxiety, depression, all ages	Pooled estimates not significant (35 reviews)	Int J Nurs Stud, May 2024	The cold shower

Two loneliness numbers from two independent teams landing at -0.59 and -0.69 is about as much agreement as this literature offers. Everything below the loneliness rows gets thinner fast.

A fifth synthesis, in Psychological Medicine (January 2026), looked at conversational agents specifically for older people without cognitive impairment and found smaller movement on loneliness (standardized mean change 0.350, 95% CI 0.180 to 0.520) and more on depression (0.464, 0.327 to 0.602). It also carries the weakest design in the set: too few good comparison studies existed, so every pooled figure is a before-and-after within one group, with no blinded outcome assessment. Numbers built that way run high by default.

Two-column board splitting the pooled findings by whether the 95% interval clears zero. Clearing zero: loneliness -0.59 (CI -0.92 to -0.26) across 19 studies and 1,083 older adults, loneliness -0.69 (CI -0.89 to -0.40) across 16 studies and 781 people, depression -0.42 (CI -0.72 to -0.12), depression and loneliness in long-term care described as large across 8 randomized trials, and a stronger -0.878 in the eight controlled studies against -0.403 in the eleven uncontrolled ones. Crossing zero or otherwise discounted: quality of life 0.24 (CI -1.0 to 0.58), happiness 0.63 (CI -1.09 to 2.35), pooled quality of life, anxiety and depression across all ages not significant over 35 reviews, publication bias with an Egger's intercept of -0.825 at p below .01, and living independently at home as the weakest setting in the analysis

The strongest single number, and its fine print

The best-sourced figure in the category comes from "Wired for companionship," a meta-analysis by Mehrabi and Ghezelbash in The Gerontologist (2025). Nineteen studies, 42 effect sizes, 1,083 older adults, literature searched through October 2024. Pooled effect on loneliness: -0.590, 95% CI -0.919 to -0.261.

Four details from the full text change how you should use that number.

Study quality moves it a lot. In the eight studies with a control group the effect was -0.878. In the 11 without one it was -0.403. Weaker designs usually flatter an intervention, so having the comparison-group studies come out stronger is a point in the category's favor.
Setting moves it more. Relative to older adults living independently, effects were stronger in nursing homes, stronger again in dementia care, and strongest in hospitals. Living alone at home is the weakest setting in the analysis and the one most buyers are shopping for.
The robot's form did not matter. No significant difference between pet robots, humanoid robots and voice assistants. Nothing in the data says the expensive one works better.
Publication bias is present. Egger's test returned a significant intercept (-0.825, p < .01). The authors report it plainly, so read -0.59 as real but flattered.

The same paper names the gap that matters most to anyone spending money: it remains unclear whether the effects persist over the long term or fade as the novelty wears off. Trials in this field run for weeks, not years.

The largest trial put the robot up against a plush toy

The field's flagship study is a cluster-randomized trial by Wendy Moyle and colleagues, published in JAMDA in 2017: 415 residents aged 60 and over across 28 long-term care facilities in Queensland, Australia. Three arms, 15-minute sessions three times a week for 10 weeks. One group got PARO, the therapeutic robot seal. One got an identical plush toy with the robotics switched off. One got usual care.

PARO beat the plush toy on attention: verbal engagement higher by 3.61 points (p = .011), visual engagement by 13.06 points (p < .0001). Against usual care it reduced neutral affect (3.09 points, p = .022) and increased pleasure (1.12 points, p = .008). On agitation it improved one measure (3.33 points, p = .008) and showed no difference on the standard agitation inventory.

Moyle herself on the findings above, in a film her university posted in August 2017: the robotic seal improved verbal and visual engagement and pleasure in people with dementia.

Then the same team costed it. In their 2018 economic evaluation, PARO ran $50.47 more per resident than usual care and the plush toy $37.26 more, for incremental cost-effectiveness ratios of $13.01 per point of agitation averted for PARO and $12.85 for the plush toy (figures as published, in a trial costed in Australia). Their conclusion, in their words, is that the plush toy offered marginally greater value for money than PARO for agitation. The 2024 umbrella review reports the same pattern from other trials: several found no significant difference between a socially assistive robot and a plush toy.

That does not make the robot pointless. It holds attention in a way a switched-off toy cannot, and attention is the mechanism the whole category rests on. It does mean the burden of proof sits on the price difference. PARO carries no published consumer price and sells by quote into clinical channels, while a Joy for All cat is $159.99.

Side by side comparison of the two active arms of the Moyle 2017 trial, 415 residents aged 60 and over across 28 long-term care facilities over 10 weeks: PARO the therapeutic robot seal against an identical plush toy with the robotics switched off, both in 15 minute sessions three times a week. PARO held attention better, verbal engagement higher by 3.61 points at p equals .011 and visual engagement higher by 13.06 points at p below .0001, but cost $50.47 more per resident above usual care against the plush toy's $37.26, and $13.01 per point of agitation averted against the plush toy's $12.85, which the authors called marginally greater value for money for the plush

Where the evidence simply stops

People living alone at home. Covered above: it is the weakest setting in the meta-analysis, and it is the setting almost every consumer companion robot is sold into. Companion robots for seniors sets out what that means for a purchase.

Anyone under about 65. Searching PubMed for companion or social robot studies on loneliness naming young adults, students or working-age people returned a single record, and that record turned out to be a design study with older adults. If you are buying a companion robot for a lonely thirty-year-old, you are ahead of the research, not behind it.

The desk-companion class. PubMed holds no record containing the word Eilik. The pet-robot and companion-robot literature that exists is overwhelmingly care homes, dementia units and community programs for elders. Eilik costs $139.99, and whatever it does for the mood of the person it sits in front of, nobody has measured it. What those machines actually run is covered in what the AI in a companion robot actually does.

Anything past a few months. The trials are short. The consumer robot graveyard is the practical version of that gap: interest fades in some households, and in others the company folds and the robot stops working while the interest is still there.

Program statistics are a different kind of evidence

New York's Office for the Aging made roughly 900 ElliQ units available to older adults in the community and reports a 95% reduction in loneliness from its 2023 evaluation, with users interacting 30-plus times a day, six days a week. A Wisconsin program reported roughly double that interaction rate, 88 a day, to WEAU in April 2026.

Those are program evaluations run with the manufacturer, covering people who volunteered and stayed enrolled, with no control group. They are useful for one thing: they show that older adults keep using these devices, which is not a given in a category littered with abandoned gadgets. A 95% from an uncontrolled program report and a -0.59 from pooled trials are different species of claim, and stacking them side by side flatters the weaker one.

Four things separate the two. Whether there was a control group. Whether loneliness was measured on a validated scale instead of a satisfaction survey. How long it ran. Who funded the evaluation and who published it.

What to do with all this

If you are deciding for a parent, the research supports a specific, cheap position. Loneliness responds. Depression responds a little. Quality of life and anxiety have not been shown to move. Form factor made no measurable difference, so start at the bottom of the price ladder unless the person needs something the cheap tier cannot do, which usually means medication reminders, calendar prompts and family check-ins. That is the case for stepping up to ElliQ at $249 plus $39 to $59 a month, and ElliQ explained covers the state aging agencies that hand it over for nothing.

The study that would settle the question, a year long, in people's own homes, against a control group, has not been run. Until it is, spend at a level you would not resent if the robot is in a cupboard by spring.