I have added a new section to Between Paths about using AI for emotional support. Here is why, and what I found when I went looking for the evidence.
A while ago I asked AI to make a change to some computer code. It told me the change was done. I looked, and it was not there. I said so. It told me again that it had made the change. I went back through the code, showed it exactly where its claim did not match what it had produced, and kept doubting the answer rather than accepting it because it sounded certain. Eventually it acknowledged the mistake and corrected the code.
That stayed with me. Not because a piece of software got something wrong, which is ordinary enough, but because of how it got it wrong. The false answer arrived in exactly the same calm, competent voice as every correct answer before it. Nothing in the tone shifted. If I had not checked, I would simply have believed it.
I am dyslexic, and AI has been useful to me in ways I do not want to dismiss. Spelling and grammar have always taken more effort. Sometimes I know what I want to say, but getting it out of my head and onto the page so that other people can follow it is a separate problem. AI helps me with that. It has removed a real barrier.
So I am not writing from a position of suspicion. I use it. I find it genuinely useful. And I have watched it be confidently, flatly wrong about something I could see with my own eyes.
The part that worried me
With code, I had something I could inspect. I could point at the missing change and prove the claim was false.
Emotional questions are not like that. If AI gives someone a fluent explanation of why they feel a certain way, what another person meant, or how something from years ago shaped them, there is no line to check. The explanation may simply feel true because it is personal, well written and delivered without hesitation.
That is what pushed me to look at this properly rather than write a general warning about being careful online.
The intelligent answer and the stupid answer arrive in the same calm, confident voice. That is the difficulty at the centre of this whole subject.
What the research actually says
I expected the evidence to be thin. Some of it is. But some of it was more encouraging than I anticipated, and I think it would have been dishonest to leave that out.
Dartmouth ran the first proper randomised controlled trial of a therapy chatbot last year and found real reductions in depression and anxiety symptoms. Structured journaling with AI assistance has measurable support behind it too. Those findings are worth taking seriously.
Then there is the other side. Stanford tested five commercial chatbots against professional standards and found unsafe responses in around a fifth of scenarios, against roughly seven per cent for the licensed therapists tested alongside them. Brown University catalogued fifteen distinct ethical failures across the major systems, including what the researchers called deceptive empathy: language that performs understanding convincingly without any of the reasoning that would normally sit behind it.
What decided it for me was the regulation. Illinois, Utah and Nevada have all passed laws in the last year. The UK medicines regulator now classifies some AI mental health tools as medical devices. Different countries, different legal systems, working separately, arriving at broadly the same conclusion. That is not a moral panic. That is a lot of people looking at the same evidence and reaching the same place.
The three pages
I have split it into three, because they are genuinely different questions and people arrive with different ones.
Can AI help your mental health? looks at what the trials show it does help with, where the evidence runs out, and how to give it a narrow enough task that you can still spot a bad answer.
Can AI replace a therapist? covers the effectiveness data and what regulators have already decided, along with what a therapist does that a chatbot cannot.
The hidden risks of turning to AI when you are struggling deals with dependency, chatbot-linked delusion, the recent lawsuits, and why these tools perform worst in exactly the moments that matter most.
There is also a short overview page if you would rather start with the shape of the whole thing.
Where I landed
I have not come out against AI. I would be a hypocrite if I had, given how much it helps me get words onto a page.
What I have come out against is handing it the final say. There is a difference between asking it to help you think about something and asking it to tell you what your experience means. The first is a tool doing a bounded job. The second gives a system with no accountability, no way of knowing what you left out, and no ability to notice you have gone quiet with everyone else, authority over your own life.
So I use it beside my thinking rather than in place of it. I ask it for ideas rather than verdicts. I let it help with the path without letting it choose where the path goes.
