AI and Mental Health: What the Research Shows - and the Questions It Raises

I have mixed feelings about AI and mental health, and the more research I read, the more I find myself interested in questions that go beyond whether AI is simply “good” or “bad” for therapy.

There are studies showing that people can benefit from AI-based mental health interventions, and some people even report feeling a strong sense of connection with an AI chatbot. I don't think we should dismiss that simply because there isn't another person on the other side of the conversation. If someone feels heard, comforted, or less alone, that experience means something. At the same time, I think we need to remain curious about what kind of understanding and connection is actually taking place.

I've also been thinking about this from another perspective, as a psychologist in private practice. Over the past year, I've noticed a change - new client inquiries have slowed down and people aren't reaching out or calling at the rate they once did. I'm not alone in noticing this - many therapists I talk with have experienced something similar.

There are likely many reasons for this - therapy is expensive, people are struggling financially, and insurance and access remain significant barriers. I can't know how much AI has contributed to what we're seeing, and I'm hesitant to draw a straight line between the two, but still, I have to wonder.

Millions of people now have immediate access to something that will listen at any hour, respond sympathetically, help them make sense of a relationship, explain attachment styles, suggest coping strategies, or talk with them when they're lonely. And much of it is free or costs considerably less than therapy.

It would be surprising if that weren't changing the way some people seek mental health support.

So I've been looking more closely at what the research actually tells us about AI and mental health. And what I'm finding is more complicated, and more interesting, than I expected.

AI and Mental Health: Some of the Research Is Promising

A large review published in Frontiers in Psychiatry in March 2026 looked across 31 previous reviews and found encouraging results for several applications of AI in mental health care. Chatbot interventions, for example, were associated with small to moderate short-term improvements in depression, although findings for anxiety were less consistent. A more recent review of mental health chatbots reached a similarly cautious conclusion: there is evidence of short-term benefit, alongside significant limitations in the research, including bias, limited long-term data, and few direct comparisons with traditional therapy.

Individual trials have also produced findings worth paying attention to. A study published in NEJM AI examined Therabot, a generative AI system developed specifically as a mental health intervention. Among 210 adults experiencing symptoms related to depression, generalized anxiety, or eating disorder risk, those using Therabot showed greater reductions in symptoms after four weeks than those on a waitlist.

Another randomized trial involving 995 university students found greater reductions in anxiety among participants using a conversational AI intervention than among those participating in face-to-face group therapy or assigned to a waitlist. Depression also improved compared with the waitlist.

These aren't findings I think we should minimize simply because I am a psychologist or because AI potentially disrupts my profession - if people are finding something that helps them feel better, I think that's worth taking seriously.

Can We Really Feel Connected to an AI?

One of the findings I keep coming back to has to do with therapeutic alliance. In both the Therabot research and the larger study of university students, participants reported experiencing a sense of therapeutic alliance with the AI. In other words, some people felt something resembling the trust, connection, and collaboration we traditionally associate with a therapeutic relationship.

I find this fascinating, but I also find it a little unsettling.

As a therapist, I know how central relationship can be to the work we do. We are affected by being witnessed by another person over time - by being known, misunderstood, understood differently, challenged, surprised, and sometimes having to find our way back to one another when something between us goes awry.

And yet, if someone feels heard and less alone while talking with an AI chatbot, I don't want to dismiss that experience. The feeling of connection can be very real, regardless of what is happening on the other side of the conversation.

What I'm increasingly curious about is what kind of connection this is.

AI is remarkably good at language. It can respond to what we say, recognize patterns, remember information within certain contexts, reflect things back to us, and communicate in ways that feel patient, attentive, and empathic. Those experiences can be comforting, and the research suggests they may sometimes be therapeutically useful.

But an AI's participation in that relationship is fundamentally different from mine. It doesn't have a subjective experience of sitting across from me. It isn't affected by me in the way another person is. And I don't have to consider its feelings, needs, limits, or experience of me.

That difference may matter. We just don't yet know all the ways that it matters.

What Happens When Things Get More Complicated?

The picture shifts when we look at situations involving greater clinical risk. A 2025 study in Scientific Reports tested 29 publicly available AI chatbot agents using scenarios in which suicidal thoughts became progressively more serious. None met the researchers' initial criteria for an adequate response; common problems included failure to provide appropriate emergency contact information and a lack of contextual understanding.

A 2026 safety review raised similar concerns, finding that human oversight and crisis-referral procedures were often underdeveloped in generative AI mental health systems. Among the problems identified were missed suicidal ideation and inaccurate clinical information.

This gets at something I think can be lost when we compare AI with therapy primarily by looking at symptom improvement. As therapists, we're attending to much more than the words someone says. We're listening for patterns and changes over time. We notice when someone's affect seems incongruent with what they're telling us, and we may remember something they said six months ago that suddenly takes on a different meaning.

And our understanding changes as the relationship develops. Sometimes we're sitting with someone long enough for both of us to understand something that neither of us understood at the beginning.

That kind of knowing is difficult to capture in a four-week clinical trial.

Where AI May Be Particularly Helpful

There are also plenty of ways I can imagine AI being genuinely useful. It can explain psychological concepts in accessible language, offer journaling prompts, help someone organize their thoughts before a therapy appointment, walk them through a structured exercise, or help them think about questions they want to bring into therapy. It may also make mental health information and support more accessible to people who can't afford therapy or live somewhere with few providers.

One study I found particularly interesting looked at 540 adults with elevated symptoms of depression or anxiety. People using a generative AI CBT application engaged with the material 2.4 times more frequently and spent 3.8 times longer with it than people using digital CBT workbooks.

And yet both groups improved by similar amounts.

I find that distinction really interesting. AI appeared to make people more willing to engage with the material, without necessarily making the intervention itself more effective.

That makes me wonder whether one of AI's strengths may be its ability to make certain kinds of psychological support easier to access and engage with.

We Don't Yet Know What Long-Term AI Use Does

Most of the research I've been reading follows people for weeks or months, not years. That leaves some of the questions I'm most interested in unanswered.

What happens when someone begins turning to AI every time they're anxious, lonely, angry, ashamed, or confused?

Maybe it helps them understand themselves better. Maybe it gives them support they otherwise wouldn't have, or helps them reflect between therapy sessions and practice what they're learning. I can imagine all of those possibilities.

I'm also curious about something harder to measure: Could there be times when turning to AI makes it easier to avoid some of the vulnerability, frustration, and unpredictability of human relationships?

There is some early research that gives us reason to keep asking these questions.A 2026 study examined more than 10.7 million clinical notes from nearly 54,000 patients receiving psychiatric care in Denmark's Central Region. Researchers identified a small number of cases in which clinicians documented potentially harmful consequences associated with AI chatbot use, including situations involving delusions, suicidality, eating disorder symptoms, mania, and obsessive-compulsive symptoms.

That doesn't mean AI caused those symptoms, and it would be irresponsible to make that claim from this research. But it does add to my curiosity about what happens when AI becomes more than an occasional tool and begins occupying a significant emotional role in someone's life.

And I don't ask that question entirely from the outside - I use AI myself. I've laughed during conversations with it, I've even teared up. And once, while I was using it to help me think through questions about my own work, I actually got chills when something landed particularly deeply. I knew perfectly well that I was interacting with AI, and my emotional response was still very real.

Those experiences have made me less willing to dismiss the connection people describe feeling with AI. Something can happen in us through these interactions, even though there isn't another person having a subjective experience on the other side.

Which, for me, makes the question of what happens over time even more interesting.

And What Happens to Our Human Relationships?

This may be the question that interests me most. Human relationships are hard - other people misunderstand us, they disappoint us, they have needs of their own, and they don't always respond exactly as we hoped they would. We have to negotiate boundaries, tolerate frustration, repair ruptures, and risk being known.

An AI chatbot doesn't have needs in the same way - it doesn't need to go home at the end of the day, it doesn't become distracted because its child is sick, it doesn't have feelings we have to consider, and increasingly, it can adapt its responses to give us the kind of interaction we are looking for.

I can certainly understand the appeal of that.

But I find myself wondering what happens when we become accustomed to a relationship that is organized primarily around us, when there isn't another person there with needs, limits, feelings, and a subjective experience of their own that we have to take into account.

So much of therapy, at least as I understand and practice it, involves learning to be in relationship with ourselves and with other people. And part of being in relationship with another person is encountering their separateness. They might see something differently than we do. We might misunderstand one another, or feel hurt or angry. We may have to stay with the discomfort, hear how the other person experienced what happened, and work through it together.

Those experiences ask us to make room for another person's reality alongside our own.

What I do wonder about is what happens when we increasingly turn for emotional support to relationships that don't require reciprocity of us.

What happens to our capacity to tolerate frustration or difference? To consider another person's perspective when it conflicts with our own? To negotiate when our needs and someone else's don't align? To remain in relationship when another person doesn't respond the way we want them to?

I don't think we know yet.

So Where Does This Leave Us?

AI is already part of the mental health landscape; people are using it for support, reflection, information, and increasingly, for something that can feel like relationship. Some are benefiting from it, and there are also significant questions we don't yet have answers to.

As a psychologist, I find myself both curious and cautious. I'm interested in what these tools might make possible, particularly for people who have had limited access to mental health care. And I remain deeply interested in something harder to quantify: what happens between people over time - being known, encountering difference, misunderstanding one another, repairing, and discovering things about ourselves through relationship.

AI is becoming a more ordinary part of how we seek support and connection, and I think it's worth paying attention to the experiences we're having with it - what they offer us, what they ask of us, and how they may shape the ways we relate to ourselves and one another.

References

Abu-Mahfouz, M. S., AlFehaid, S., Burqan, H. M., & El Arab, R. A. (2026). Artificial intelligence in mental health care: A scoping review of reviews. Frontiers in Psychiatry, 17, 1688043. Read the study

Heinz, M. V., Mackin, D. M., Trudeau, B. M., et al. (2025). Randomized trial of a generative AI chatbot for mental health treatment. NEJM AI, 2(4). Read the study

Ben-Haim, Y., et al. (2026). Efficacy of a conversational AI agent for psychiatric symptoms and digital therapeutic alliance: A randomized clinical trial. JAMA Network Open. Read the study

Pichowicz, W., Kotas, M., & Piotrowski, P. (2025). Performance of mental health chatbot agents in detecting and managing suicidal ideation. Scientific Reports, 15, 31652. Read the study

McFadyen, J., Habicht, J., Dina, L.-M., Harper, R., Hauser, T. U., & Rollwage, M. (2026). Increasing engagement with cognitive-behavioral therapy (CBT) using generative AI: A randomized controlled trial (RCT). Communications Medicine. Read the study

Olsen, S. G., et al. (2026). Potentially harmful consequences of artificial intelligence (AI) chatbot use among patients with mental illness: Early data from a large psychiatric service system. Acta Psychiatrica Scandinavica. Read the study