What the study actually did
The paper asks a narrower question than the headlines it produced: in everyday, nonclinical situations, can a language model do what psychologists call extrinsic interpersonal emotion regulation — one person helping another manage a feeling — as well as a person can?
The team at the universities of Manchester and Durham built scenarios from ordinary life: a job rejection, a messy shared house, the kind of thing that makes someone angry, sad or afraid. Participants read a comfort message without being told who or what wrote it, then reported how they felt and how supported they felt.
| Experiment | N | What it tested |
|---|---|---|
| Study 1a | 279 | Thematic analysis: do LLM replies mirror human regulation strategies? Largely yes, with differences between models. |
| Study 1b | 390 | Head-to-head regulatory effects. LLMs sometimes stronger — but effects varied by emotion and by outcome. |
| Study 2 | 196 | Is emotional validation the explanation? It was not. |
| Study 3a | 188 | Is actionable support the explanation? Evidence said yes, across sources. |
| Study 3b | 180 | Same test with message length controlled. Result held. |
That adds to 1,233 participants. The models compared were ChatGPT 4o, Claude 3.5 Sonnet and Gemini 1.5 Pro.
Where AI came out ahead, and where it did not
In the 390-participant experiment, participants rated the AI messages as more emotionally supportive for the anger and fear scenarios. For sadness there was no reliable difference. That much matches the coverage.
What the coverage mostly drops is the second half. The abstract records no regulatory effects on fear, and the authors explain what that means: in the fear scenario the AI messages increased calm more than human messages did, but they did not produce a greater reduction in fear itself.
| Emotion | Rated support | Measured emotional change |
|---|---|---|
| Anger | AI rated more supportive | LLM advantage present in some outcomes |
| Fear | AI rated more supportive | Calm increased; fear itself not reduced more than by humans |
| Sadness | No reliable difference | No difference reported |
The gap the headlines skip: feeling supported is not feeling better
Two different things were measured, and they came apart. One is a judgement — this message seems understanding. The other is a state — I am less afraid than I was. A message can win the first and lose the second, and in the fear condition that is exactly what happened.
This matters more for companion apps than for chatbots in general, because a companion app is optimised against the first measure. Perceived warmth is what gets rated, screenshotted and retained. If you are using an app to actually feel different at the end of a bad evening, the rating is the wrong signal to trust.
Validation was not the explanation
The obvious hypothesis is that language models are simply better at saying that sounds really hard. Study 2 tested it directly. Explicitly acknowledging and validating the recipient's feelings did not account for the greater emotional improvement associated with the AI messages.
That is a useful negative result, and it lines up with a separate strand of 2026 work we covered in our piece on whether companion validation backfires. Affirmation is the easiest thing for a language model to generate at scale, which is precisely why it is a weak explanation for anything.
What did explain it: actionable support
Studies 3a and 3b pointed at something more mundane. Messages containing actionable support — specific, realistic, implementable suggestions rather than general reassurance — were rated as more comforting and improved some emotional responses. The effect appeared regardless of whether a person or a model wrote the message, and it survived controlling for message length, so it was not an artefact of AI replies simply being longer.
Read the other way round, this deflates the “AI advantage” considerably. When human messages offered comparable practical help, they were judged just as supportive. The model was not doing something humans cannot do. It was doing reliably, on demand, what a tired friend often forgets to do.
The authors add a caution against overcorrecting. They cite 2024 work finding that excessive suggestions can be less effective at making people feel heard, and suggest asking whether advice is wanted before offering it — then offering one manageable step rather than a list.
Why people still prefer a human
Source attribution changes reception. Across nine further studies involving 6,282 people, the same AI-generated responses were rated as more empathic and supportive when described as human-written than when attributed to AI. Participants preferred human interaction when seeking emotional engagement even when choosing a person meant waiting longer.
Nobody has established why. One reading is that a human reply signals a willingness to spend time and effort, and arrives inside a relationship that continues after the message. As the authors put it, gen AI can suggest going for a walk, but it cannot come along.
How to use this when picking a companion app
The finding converts into a test you can run on any app in five minutes. Send it something genuinely mildly annoying from your week and read the reply against this list.
- Does it offer one specific next step? Not “take care of yourself” — something with a verb and a scope you could do today.
- Is the step realistic for your actual situation? Generic advice that ignores the constraints you stated is decoration.
- Does it ask before advising? The research supports checking whether advice is wanted; an app that always solves is as mis-calibrated as one that never does.
- Is it one step or a pile? More suggestions is not more help, and can make people feel less heard.
- Strip the warmth and re-read it. If nothing is left once the affirming sentences are removed, you are buying tone, not support.
If you are still narrowing a shortlist, our guide to choosing an AI companion app covers the rest of the criteria, and our wellbeing overview collects what the 2026 evidence does and does not support.
What the study cannot tell you
The authors are explicit about the boundaries, and they are worth restating because this is the point where research about single messages gets stretched into claims about relationships.
- The scenarios were imagined everyday situations in nonclinical contexts, and the measures were immediate reactions to brief or one-off exchanges.
- It cannot say what happens when someone relies on a chatbot over months, or uses one during a mental health crisis.
- The human messages were generally written by strangers or built as representative examples — not by a friend who knew the recipient and the wider circumstances.
- Findings were associations under experimental comparison, not evidence that a particular style of support causes better relationships.
For longer-horizon questions, the studies that follow people over time are a better guide than this one — see our summary of the 2026 loneliness and social-anxiety findings.
The bottom line
The durable result here is not “AI comforts better than people.” It is that the ingredient which moved the needle was a concrete, doable suggestion, and that it worked whoever supplied it. For anyone choosing a companion app, that turns a vague question about warmth into a testable one. For everyone else, it is a free lesson in how to answer a friend who is having a bad day.
Related on CompanionRank
Does AI Companion Validation Backfire? 2026 Research Does AI Chat Support Ease Loneliness? New 2026 Study AI Companion Mood Systems and Emotional Wellbeing How to Choose an AI Companion AppFrequently asked questions
What did the 2026 Emotion study on AI emotional support find?
Across five experiments with 1,233 participants, large language model messages sometimes produced stronger emotion-regulation effects than human-written ones, but the advantage depended on the emotion and on which outcome was measured. There was no difference for sadness, and in the fear scenario the AI messages raised calm without reducing fear itself. The authors identified actionable support — specific, implementable suggestions — as the feature that improved effectiveness regardless of who wrote the message.
Is AI better than a human at emotional support?
Not as a general claim. In the 390-participant experiment, participants rated the AI messages as more emotionally supportive for anger and fear scenarios, with no reliable difference for sadness. But when human messages offered comparable actionable support, they were judged just as supportive. The study compared isolated messages about imagined everyday situations, not support from someone who knows you.
Why did AI messages score higher on perceived support?
The researchers tested emotional validation — explicitly acknowledging the person's feelings — and found it did not account for the greater emotional improvement. Two later experiments pointed instead to actionable support, and the result held when message length was controlled, so it was not simply that AI replies were longer.
Does knowing a message came from AI change how it lands?
Yes. The authors cite nine further studies involving 6,282 people in which the same AI-generated responses were rated as more empathic and supportive when described as human-written than when attributed to AI. Participants also preferred human interaction when seeking emotional engagement, even when that meant waiting longer.
What does this mean for choosing an AI companion app?
Judge replies on whether they offer one specific, realistic next step rather than on how warm they sound. Warmth is cheap for a language model to produce; a concrete suggestion you can actually act on is the part the research tied to emotional improvement. An app that only mirrors your feelings back is optimising for the metric that did not explain the effect.
What can this study not tell us?
It tested brief or one-off exchanges, mostly about imagined everyday situations, and measured immediate reactions in nonclinical contexts. It cannot say what happens when someone leans on a chatbot for months, or during a mental health crisis. The human comparison messages were written by strangers or built as representative examples, not by friends who knew the recipient.