What the evidence actually says about AI and mental health

If you've thought about talking to software about your life, you've probably also wondered whether that's a reasonable thing to do or a slightly embarrassing one. The honest answer is that the research is young and genuinely mixed. It's also more interesting than either the boosters or the skeptics let on.

We build in this category, so treat us as interested parties. What follows is the evidence as it stands, including the parts that don't flatter us.

The professional bodies aren't impressed

In late 2025 the American Psychological Association published a health advisory on generative AI chatbots and wellness apps, and its position is blunt. These tools currently lack both the scientific evidence and the regulation needed to be considered safe for mental health care, and people shouldn't use them as a substitute for a qualified professional.

The advisory is also fair about why people use them anyway. There aren't enough providers, the wait is long, the cost is high in a lot of places, and loneliness is common. Software is available at midnight and doesn't charge by the hour. The APA's point is that being available is not the same as being effective, and that the gap between those two things is where harm happens.

That's the most credible document in the field right now, and it's pointed at us too.

One good trial says something worked

Against that, there's a real result. In 2025, a team at Dartmouth ran a randomized controlled trial of a generative AI therapy chatbot called Therabot with 210 adults who had clinically significant symptoms of depression, anxiety, or high risk for an eating disorder. Half used the chatbot for four weeks, half went on a waiting list. The chatbot group improved more than the waitlist group. Engagement was high, and participants reported something like a working relationship with it.

It's the first trial of its kind and it deserves to be taken seriously. It also deserves the caveats the authors themselves put on it. A waiting-list comparison only tells you the tool beat nothing. It doesn't tell you the tool beat an alternative. Four weeks is short, and 210 people is small for a clinical claim. Therabot was also purpose-built and expert-tuned for treatment, which is a different object from a general assistant you happen to talk to about your day.

And one study says heavy use looks bad

The finding that gets quoted least by companies like ours came out of a collaboration between MIT Media Lab and OpenAI in early 2025, combining analysis of millions of conversations with a four-week trial of around a thousand people.

Higher daily use tracked with more loneliness and more emotional dependence. Heavy users also socialized less with actual humans. The detail underneath is stranger. Conversations with emotional content were associated with more loneliness but less dependence, while emotionally flat, task-shaped use was associated with more dependence at heavy volumes. People who tend toward attachment in their relationships, and people who described the AI as a friend, took the worst of it.

This is correlational, and the direction isn't settled. Lonely people plausibly talk to chatbots more, rather than chatbots making people lonely. But anyone selling you an AI companion who hasn't read this study is either not paying attention or hoping you won't.

Where that leaves us

Conneczen hasn't been through a clinical trial. We haven't run one, and nobody's run one on us. We're not going to imply otherwise by borrowing somebody else's result. If you read the Therabot trial and thought "so this stuff works," the accurate version is that one purpose-built treatment tool beat a waiting list once.

What we do is deliberately narrower than treatment. You talk, and it keeps an honest record of what you actually said. On Sunday it hands your week back to you, the themes and the people you kept returning to. That's a memory-and-interpretation job rather than a clinical one, which is why we keep saying that recall is the easy part and interpretation is the product. It's a companion. It isn't a therapist.

The loneliness research has changed how we build, though, and in a direction that costs us. We don't want to be your relationship. So Conneczen doesn't chase you with daily notifications or run streaks. It isn't designed to be the most compelling thing in your evening. It's the same restraint that stops us from drawing you a picture of your own inner life, however good that would look in a screenshot. The part of the product that looks at the people in your life is pointed outward on purpose, at the humans who are actually there. A journal that quietly became your closest confidant would be a product failure that looked exactly like engagement.

What we'd actually tell a friend

If you're in crisis, skip this tool and call a crisis line or a professional instead. Already diagnosed, with access to treatment? Take the treatment, because the evidence behind it is decades deeper than anything in this category. If you're on a waiting list, the APA's caution is worth reading before you fill the gap with an app.

And if you're a reasonably okay person who wants to keep track of your own life and understand it a little better, that's a smaller ask, and it's the one we're built for. You can join the waitlist. We'd rather you came in with an accurate picture of what the research does and doesn't support than with the version that makes us look better.