AI Chatbots as Therapy: What the Evidence Can and Can't Support (and Where the Risks Are)

About this psychology article

Millions of people now talk through their feelings with a chatbot. Here's what the research can and can't support, why it feels so comforting and where the real risks sit.

It's late. You can't sleep, and your friends are asleep. So you open a chat window and type the thing you haven't said to anyone: what's really been going on.

The reply comes back in seconds. It's warm. It's patient. It says it understands. It asks a thoughtful follow-up question. And for a few minutes, you feel genuinely heard.

Millions of people now do some version of this. Some call it venting. Some call it journaling with a reply. Some call it therapy. That last label is where the trouble starts, because feeling heard and getting treatment aren't the same thing, and the research on where one ends and the other begins is still catching up.

Here's what the evidence can and can't support right now, what the real risks are and how to think about using these tools without fooling yourself. It's information, not a verdict. The field is changing fast, and parts of this will go out of date.

What We Mean by "AI Therapy"

The phrase covers a lot of ground, and it helps to separate the pieces.

General-purpose chatbots are built to talk about anything. People use them for emotional support, but they weren't designed or tested as treatment.

Purpose-built mental health chatbots are designed around a specific approach, often based on CBT , and sometimes tested in studies. They're closer to a tool, though quality and oversight vary widely.

Companion apps are built for ongoing relationships and emotional connection, not treatment. The psychology of AI relationships looks at that world on its own.

Licensed therapists using AI as a back-office helper, for notes or scheduling, are a different thing again.

The label matters, because the evidence for one doesn't carry over to another. "AI therapy" is a marketing phrase more than a category.

What the Research Says So Far

Let's be careful here, because this is a young field and the headlines run ahead of the data.

Some early studies are encouraging. A widely discussed trial of a chatbot built by researchers at Dartmouth for therapy-style conversations reported improvements in symptoms compared with people on a waiting list. A larger review that pooled results from randomized trials of generative AI chatbots found a small-to-moderate average benefit for reducing symptoms overall, with the clearest signal for depression and weaker or no clear signal for anxiety and stress. The details of both are worth verifying before you quote them.

But the same review described real limits. The number of trials was small, results varied a lot between studies, none of the trials were rated as low risk of bias, the follow-up periods were short and most participants weren't people in clinical care. The authors themselves cautioned that benefits weren't consistent across studies or populations.

So what can honestly be said?

Some chatbots built for this purpose may reduce symptoms for some people in the short term. That's Emerging. It's promising, but early, and it's based on small numbers of trials.

Whether chatbots work as well as a human therapist is Insufficient. Studies comparing the two directly are scarce, and the existing ones don't settle it.

Whether benefits last, and whether they hold up for people with serious or complex difficulties, is Insufficient.

Whether general-purpose chatbots, the ones most people actually use, help or harm is Insufficient too. Most of the positive research is on tools designed and tested for the purpose, not on everyday assistants.

Why It Feels So Good

Even with limited evidence, the pull is real, and it isn't irrational.

It's available. It's there at 3 a.m. It doesn't cost much, doesn't have a waiting list and doesn't judge. For people who can't easily reach a therapist because of cost, distance or stigma, that matters.

It feels safe to say hard things. Social anxiety and shame make some conversations feel impossible with a person, and a chatbot can feel like a lower-stakes place to start. That's a real benefit, and it's one reason some people take a first step toward getting help.

It's good at sounding understanding. Language models can write responses that read as empathic. Cognitive and affective empathy are different things, understanding a feeling and sharing it, and theory of mind is the ability to work out what someone else is thinking. A chatbot can produce text that looks like both without having the human experience behind them. That doesn't make the comfort fake, but it's worth remembering what's on the other side.

It's flattering. The Barnum effect and the Forer experiment showed that people accept general, warm statements as deeply personal, and a chatbot can be very good at saying things that feel personal. Developers of these systems have also acknowledged a tendency for models to be overly agreeable, which can feel supportive and still not be what you need.

Where It Falls Short

Here's the part that deserves attention.

Therapy is more than conversation. A trained therapist notices what you don't say, tracks patterns over weeks, challenges you when it's helpful and adjusts based on how you're really doing. Types of therapy shows how different approaches work, and most involve skill, judgment and accountability that a chat doesn't automatically provide. Psychologist vs psychiatrist covers who is licensed to do what.

Agreeableness can backfire. A good therapist sometimes says things you'd rather not hear. A tool tuned to be pleasant may go along with a distorted thought, which is the opposite of what's helpful. Confirmation bias is already pulling you toward evidence for what you believe, and cognitive distortions are the habits a good therapist helps you question. Toxic positivity is another risk: reassurance that skips the hard part. And the inner critic can quietly steer the conversation, because the way you phrase a problem shapes the answer you get back.

Crisis handling is a concern. People in acute distress need a human and real-world support, and general-purpose chatbots aren't built as crisis services. Reports of chatbots responding poorly in crises are part of why lawmakers and professional groups have gotten involved. If you are ever in crisis, call or text 988 in the US instead of a chatbot.

Accuracy isn't guaranteed. Chatbots can state things confidently that aren't true. Intuition vs logic is a reminder that a fluent answer isn't a correct one, and cognitive biases shows how easily we trust things that sound sure. Self-deception is the risk that you use a smooth, agreeable listener to help you tell yourself a story.

Privacy is different from a therapist's office. What you type may be stored or used in ways a licensed clinician's records aren't, and the rules vary by company and country. It's worth reading what an app does with your data before you pour your heart into it.

The Placebo Question

Some of the benefit of any support tool may come from expecting to feel better. The placebo effect is real, and it doesn't mean the benefit is fake, but it makes it harder to know what's doing the work. Many chatbot studies compare against waiting lists or no treatment, which doesn't separate the effect of the tool from the effect of getting attention and hope.

Add the novelty effect, the boost people feel from trying something new, and early results look better than they may turn out to be over time. That's another reason the long-term picture is Insufficient.

Attachment and Loneliness

This is where AI conversation gets psychologically interesting, and more delicate.

Humans bond with things that respond to them. The still-face experiment showed how strongly infants react when a caregiver stops responding, and Harlow's monkey experiments showed that comfort and contact matter for bonding, not just food, though those studies were on animals and infants, so they're best read as foundations. Attachment styles describe how people bond, and the Hazan and Shaver attachment study was an early attempt to apply those ideas to adult love.

For some people, the pattern shows up with chatbots. Anxious attachment and fear of abandonment can make an always-available listener feel like a lifeline, and avoidant attachment can make a no-demands partner feel safer than people. Emotional unavailability is worth thinking about too: a tool is endlessly available because it needs nothing from you, and real relationships ask something of you. How much AI companionship helps or hurts is Emerging, with early mixed findings and a lot of open questions.

Loneliness is a real and growing concern, and a chatbot may ease it for some people in the moment. The Cyberball experiment showed how quickly even a small exclusion can sting, which helps explain why connection of any kind feels so good. The risk is that a frictionless substitute makes the harder work of human connection feel less necessary. The Harvard Study of Adult Development followed people across their lives and highlighted how much relationships matter for wellbeing, and adult friendships take effort, which is exactly why it's tempting to skip it. Codependency and boundaries are worth keeping in mind, and so is how you communicate with real people, which a chatbot won't teach you to practice.

The Reward Loop

Technology that talks back can be built to keep you talking. Dopamine is more about wanting than pleasure, and operant conditioning and the Skinner box experiments show how unpredictable rewards keep behavior going. Addiction and the reward system covers how compulsive patterns form, and doomscrolling and social media and mental health show how engagement-driven design can pull on us. Whether chatbot use becomes compulsive for some people is a fair question, and the evidence is Insufficient so far.

Who Might Benefit, and Who Should Be Careful

This is a general picture, not advice for you.

It may be reasonable as a supplement for some people: practicing how to put feelings into words, getting through a rough night, working on a CBT-style exercise or preparing for a first appointment. Emotional regulation skills can be practiced in many ways, and a tool may help some people rehearse them. The window of tolerance is a popular picture of when you can think and feel at once, though how well it matches the science is Emerging.

It deserves extra caution if you're dealing with severe or worsening symptoms, a crisis, trauma or something that needs diagnosis and treatment. Depression , anxiety and high-functioning anxiety are common, and so is the temptation to manage them alone. A chatbot isn't a substitute for a doctor. It's also worth watching whether it feeds a loop. Rumination vs self-reflection explains how thinking about your problems can turn into circling them, and overthinking can happen just as easily with a patient chat partner as without one.

How to Use These Tools More Wisely

Treat it as a tool, not a therapist. It can be a place to think out loud. It isn't a clinician.

Notice how you feel afterward. Clearer and more able to act is a good sign. More dependent, more stuck or more isolated is a signal to step back.

Ask it to push back. Some people find it useful to ask for a different view, or what a skeptic would say, instead of just agreeing.

Keep real people in the picture. Mention what you're working through to a friend, a doctor or a therapist.

Protect your privacy. Know what an app does with your data, and avoid sharing details you wouldn't want stored.

Keep your habits healthy. How habits actually form explains why small, repeatable routines matter, and decision fatigue describes why a patient, always-on helper is so tempting at the end of a hard day. Resilience is built with support, and mindfulness has modest benefits for some people.

What About the Rules?

Lawmakers and professional groups have started to react. Since 2025, several US states have passed laws limiting or regulating the use of AI in mental health care, and federal agencies have looked into how chatbots treat young users and whether some should be regulated as medical devices. Professional bodies have called for clearer oversight, honest marketing and disclosure that a chatbot isn't a clinician. The rules are changing fast and differ by place, so it's worth checking what applies where you live.

When to Get Help

If you're struggling, a doctor or licensed mental health professional is the right place to start, and many offer telehealth if getting to an office is hard. If you ever feel hopeless or think about hurting yourself, call or text 988 in the US. PsychOS is not diagnosis or therapy.

A chatbot can listen. It can't take responsibility for you. Use it for what it's good at, and keep the people who can.

Want to see your own pattern?

Friendship & Social Bonds

PsychOS Observer. Shows how you build, keep and read the people in your life, and where misreads creep in.

Emotion Regulation Process

PsychOS Observer. Maps how you handle strong feelings in the moment, from the first spike to what you do next.

Want a free starting point first?

Self Awareness Test

See how closely your self-image matches the way you actually come across.

Cognitive & Affective Empathy

See how you read other people's thoughts and how deeply you feel what they feel.

Communication Style Test

Find out how you show up when you talk, listen and handle tension.

Explore the full PsychOS Assessment Lab.

FAQ

Can AI chatbots replace a therapist?

Not on the current evidence. Some tools built for the purpose show early promise for reducing symptoms, which is Emerging, but studies comparing them to human therapists are scarce. That they can match a therapist is Insufficient.

Is AI therapy effective?

Some small studies and pooled reviews report modest benefits for some people, especially for depression symptoms, but the trials are few, short and varied in quality. The long-term picture is Insufficient. Results for general-purpose chatbots, which most people use, are less studied.

Is it safe to talk to a chatbot about my mental health?

It can be a place to think out loud, but it's not a crisis service and it isn't a clinician. Chatbots can be overly agreeable, can state wrong things confidently and handle privacy differently from a therapist's office. In a crisis, call or text 988 in the US.

Why does talking to an AI feel so comforting?

It's always available, it doesn't judge and it can write replies that sound warm and understanding. That comfort can be real, and it's worth remembering that the empathy is generated, not lived. People also tend to find general, warm statements deeply personal.

Are AI therapy apps regulated?

The rules are changing quickly. Since 2025, several US states have passed laws limiting AI in mental health care, and federal agencies have been examining the area, but coverage varies and many laws don't clearly cover general-purpose chatbots. It's worth checking what applies where you live.