The Empathetic Algorithm: Can AI Really Support Mental Wellness?

Conversational AI is now a routine part of mental health support: always available, never tired, free or nearly so. Whether it is any good depends on what you need from it. The research gives a clearer answer than the hype in either direction.

Key takeaways

In a 2023 meta-analysis of 15 randomised trials, AI conversational agents significantly reduced depression (Hedges' g = 0.64) and distress (g = 0.70). They worked better when generative, multimodal, and built into a mobile app (Li et al., 2023).
AI empathy is simulated: recognising cues and producing fitting responses, without experiencing anything. People can still find it helpful, but they also value knowing that empathy comes from a person (Rubin et al., 2024).
Ethics reviewers warn about dependency, privacy and the lack of safeguards for severe cases, and call for guidelines rather than bans (Fiske et al., 2019).
Use AI for the 2 a.m. moments, rehearsal and journaling; use humans for diagnosis, crisis and anything you need to be witnessed rather than answered.

Does it work?

Li, Zhang, Lee, Kraut and Mohr (2023) screened 7,834 records and analysed 35 studies, 15 of them randomised controlled trials. Read the meta-analysis. The pooled result: AI conversational agents significantly reduced symptoms of depression and psychological distress, with medium-to-large effects. The moderators are the useful part. Agents worked better when they used generative AI rather than scripted replies, when they were multimodal, and when they lived inside a mobile or messaging app people already used.

Long-term effects were less clear, and most trials targeted mild-to-moderate symptoms. A separate randomised trial of a generative-AI therapist, covered in our Therabot piece, found the same shape of result: meaningful symptom reductions, with the authors explicit that it supplements rather than replaces care.

Can it be empathetic?

Simulated vs. felt

A model learns what empathetic responses look like from millions of conversations and reproduces the form: reflecting feelings, validating, asking gently. It does not feel any of it. Rubin, Arnon, Huppert and Perry (2024) argue this matters most in therapy that works through the relationship itself, where knowing a person is choosing to understand you is part of the effect. Read the viewpoint.

Where the simulation is enough

For a lot of what people want at midnight (to be heard without judgment, to sort out what they feel, to rehearse a hard conversation) the form of empathy does most of the work. The meta-analysis results come largely from exactly these uses.

Where it is not

Fiske, Henningsen and Buyx (2019) flag the risks: emotional dependency on a system that cannot reciprocate, privacy of intimate disclosures, bias, and the absence of a clinician when things turn severe. Read the paper. Their conclusion is not to stop, but to set guidelines and keep humans in the loop.

Using AI support well

Name what you want from it. "Help me get through the next hour" and "tell me if I'm depressed" are different requests; AI is decent at the first and should not be trusted with the second.
Watch for substitution. If the AI conversation is replacing a friend you used to call, that is dependency, and it is the risk the ethics literature is most worried about.
Rehearse, then do it for real. Practising how to tell your partner you have been struggling is a good use; never telling them because the bot already listened is not.
Know the exit. If you have thoughts of harming yourself, an AI chat is not the place; call a crisis line or a person you trust.
Tell the people who matter. Using an AI for support is not shameful, and hiding it from a partner makes it a secret rather than a tool. Our comparison of AI and human relationship coaches covers where each fits.

Relationship-specific support is a subset of this. MosaicChats' assistant, Myrah, answers questions about a conversation you have uploaded, which makes it useful for "why does this feel worse lately" and the wrong tool for "am I okay." Keep the two apart.

Understand the conversation, not just the feeling

Upload a chat and MosaicChats shows how its tone has moved and who carries it. Myrah answers questions from that record, so you can see what changed and when.

Analyze your chat

Frequently asked questions

Can AI actually feel empathy?

No. AI recognises emotional cues and produces responses that fit them, which is a simulation of empathy rather than the experience of it. Research suggests the form is often enough for support, but people still value knowing that empathy comes from a person, especially in therapy that works through the relationship.

Do AI mental health chatbots work?

In trials, yes, for mild-to-moderate symptoms. A 2023 meta-analysis of 15 randomised controlled trials found significant reductions in depression (Hedges' g = 0.64) and distress (g = 0.70), with larger effects for generative, multimodal, app-integrated agents. Long-term effects are less clear.

Is it unhealthy to rely on an AI for emotional support?

It depends on what it replaces. Using AI to get through a hard hour or rehearse a conversation is fine. Using it instead of people you used to talk to is the dependency risk ethicists warn about. If it is crowding out human contact, that is the signal to rebalance.

Should I tell my partner I use an AI for support?

Yes. It is a reasonable tool, and keeping it secret turns it into one more thing you are not saying. Many people also use AI to rehearse exactly that conversation, which is one of its better uses.

References & Sources

Fiske, A., Henningsen, P., & Buyx, A. (2019). Your robot therapist will see you now: Ethical implications of embodied artificial intelligence in psychiatry, psychology, and psychotherapy. Journal of Medical Internet Research, 21(5), e13216. Source
Li, H., Zhang, R., Lee, Y.-C., Kraut, R. E., & Mohr, D. C. (2023). Systematic review and meta-analysis of AI-based conversational agents for promoting mental health and well-being. npj Digital Medicine, 6, 236. Source
Rubin, M., Arnon, H., Huppert, J. D., & Perry, A. (2024). Considering the role of human empathy in AI-driven therapy. JMIR Mental Health, 11, e56529. Source

The Empathetic AlgorithmAI in Mental Wellness