Every conversation I see about AI and mental health starts in the same place: is this dangerous?

I want to start somewhere else. Before we decide whether it's dangerous, it's worth asking why someone reaches for a chatbot instead of a therapist in the first place. Because the answer shapes everything that comes after it.

A note on where I'm standing. I'm a PhD candidate in psychology with a master's in performance psychology, and my work sits at the intersection of psychology and technology. I'm not a licensed clinician. So I'll speak to how people relate to these tools and where the real risks and balance points are, and I'd defer to licensed clinicians on the specifics of clinical practice.

Why people turn to AI in the first place

In my view it comes down to two things.

Some people can't afford therapy. Full stop. That's not a preference or an avoidance pattern, it's a math problem, and no amount of encouragement solves it.

And some people carry preconceived notions about what therapy is. That it's only for people who are really broken. That they'll be analyzed and judged. That it isn't for someone like them. For that group, a chatbot feels lower-stakes and less exposing. Nobody's face is in the room. Nobody's writing anything down. You can close the tab.

So this isn't simple ignorance. It's a cost barrier for one group and a perception barrier for another.

That distinction matters more than it might seem, because an all-or-nothing "never use AI" message fails both groups. One can't afford the alternative. The other doesn't trust it. Telling either of them to just go see a therapist isn't advice, it's a dismissal dressed up as concern.

Balance is the only honest answer here.

Where AI can genuinely help, and where it can't

I'll say something that might be more measured than you expect. AI with solid guardrails can help fill a gap for people who need support and can't access it. That's real. Pretending otherwise doesn't protect anyone, it just makes the people using these tools less likely to tell you they're using them.

But I'm skeptical it will ever replace a therapy practice, and the reason is specific.

Good therapy is closer to an art form than a protocol.

It runs on attunement to physical and embodied cues. Tone. Body language. The pause before an answer. What a person isn't saying. Those are the things a chatbot simply cannot perceive, and that even a skilled human can sometimes miss. A therapist is reading a whole person in a room, not a block of text.

A therapist also analyzes, and in some cases diagnoses. That's a clinical act, and it isn't one current AI can responsibly perform.

Early detection is genuinely valuable, and maybe someday AI helps flag things sooner. But we don't yet have the research to read mental health the way you'd read a scan. It isn't as clean as spotting something on an X-ray, and pretending it is would be dangerous.

Here's a related thought that tends to surprise people: I'd argue coaching is more replaceable by AI than therapy. Coaching is largely frameworks, accountability, and pointing someone in a direction. Those things structure well. Therapy is the harder, more human thing, and it's the one people assume is easier to automate.

The real danger isn't AI. It's the incentive behind it.

Here's the line I'd draw.

The danger isn't the technology itself. It's how a given tool is designed, and what it's being optimized for.

If an AI is built to deepen engagement, to keep you talking, keep you coming back, keep you hooked, that is genuinely dangerous in a mental health context. Not because engagement is evil in the abstract, but because sometimes the right move is to point a person out. Toward a human. Toward real help. Toward ending the conversation.

An engagement-optimized system will not do that. It can't. Pointing you away is a failure by its own metric.

A tool built to point you in the right direction is a completely different product from one built to keep you on it, even if they look identical on your screen. That distinction, direction versus engagement, is where the actual risk lives. And it's a design decision made in a room you'll never see, by people whose incentives you can usually infer from their business model.

So what should someone actually do?

If you're going to use AI for support, be a discerning consumer.

Choose a tool built by a company that publishes its guardrails and is transparent about how it works, what it will do, and what it won't. If you can't find that information, that absence is information.

Make an informed choice. Don't reach for the flashy thing because it seems cool or because everyone's using it.

Transparency is the thing to demand, and demanding it puts some of the responsibility back where it belongs, on the companies building these tools. Right now most of the burden in this conversation lands on the individual user, usually a person already struggling, and that's backwards.

Kids and anyone in crisis: this is where the line holds

A lot of the alarm in this conversation is about children turning to AI for emotional support. I'd reframe that.

If a kid is seeking out a chatbot for comfort, that's a signal that the support system around them has a gap. And often the parents don't have the tools or the bandwidth either. They're not negligent, they're stretched.

An all-or-nothing ban doesn't fix that. It just takes away one imperfect thing a struggling kid reached for, without addressing why they reached. The answer is building the support around them. Not scapegoating the child, and not scapegoating the tool.

But there's a floor under all of this, and it does not move.

For a child in genuine distress, or for anyone in crisis or having thoughts of suicide, AI is never the primary support. Human help is non-negotiable.

Balance means not abandoning the person. It does not mean a chatbot is ever enough. In those moments the only right answer is a real human and real resources.

On the cost barrier, because it's the root of a lot of this

A lot of people turn to AI because therapy feels out of reach, so it's worth naming that affordable options exist and are often underused.

Open Path Collectiveconnects people to vetted therapists at reduced sliding-scale rates.

Community mental health centers and federally qualified health centersset fees by income and see people regardless of ability to pay.

University training clinicsoffer closely supervised care at a fraction of standard rates.

Employee Assistance Programs, which many employers offer, include free sessions that a lot of people don't realize they have.

The 988 Suicide and Crisis Lifelineis free, confidential, and available 24/7 by call or text in the US.

Cost is the most cited barrier to care. Pointing people toward these is often more useful than warning them off AI.