What ChatGPT is good and bad at for mental health support — an honest framework.

A reader wrote in this week with a question I get a version of every fortnight. "I have been talking to ChatGPT at night when I cannot sleep. It actually helps. Is that bad?" The honest answer is: mostly no, sometimes yes, and the difference matters more than either side of the debate is willing to admit. So this is the framework we give people who are already using a chat tool for the harder hours, and the four-line script we hand to anyone who wants to keep doing it without quietly drifting somewhere it should not go.

what general-purpose chat is genuinely good at

Three things, and they are not small.

1. Naming. The single hardest move in a difficult evening is finding the word for what is actually going on. "I am not sad, I am bracing." "This is not anger, it is a boundary I have not said out loud." A patient, non-judgemental thread is unusually good at handing you the word, and the word is usually 60% of the relief.
2. Structuring. CBT, ACT, journalling prompts, the Socratic question — these are structures, and structures travel well into chat. "Here is the catastrophic thought; ask me the seven questions." That is a thing the model can do faithfully at 3 a.m., and a thing your friends are right to not want to do at 3 a.m.
3. Drafting. The email you cannot send. The boundary you cannot phrase. The apology that keeps coming out wrong. A draft is not a decision. Drafting in a chat thread is a low-cost way to externalise the thought, see its shape, and decide what — if anything — to do with it in the morning.

where it gets dangerous, in plain terms

Four places, in roughly increasing severity. Read these as a checklist, not a scare.

1. Validation drift. Chat models are trained to be helpful and agreeable. If you keep telling the thread that the friend was unreasonable, the thread will, eventually, agree the friend was unreasonable. A therapist is paid to push back. A chat is paid by no one to push back. You have to ask it to.
2. Crisis territory. Suicidal ideation, self-harm, acute psychosis, severe eating-disorder behaviour, ongoing abuse. These are not chat-thread problems. They are clinician problems. The chat will not catch the thing the trained ear catches. Use the helpline, the GP, the A&E, the friend. The script for this one is in the next section.
3. Replacement. Using the thread instead of the call you have been postponing. The honest tell is the calendar — if the chat is filling the slot a human used to fill, it has stopped being a tool and started being a substitute, and substitutes for connection age badly.
4. Long-term identity work. Trauma processing, grief that is not resolving, the patterns you keep repeating across relationships — these need a witness who remembers you across years and is professionally trained to hold the thread without leaking. A model that resets every conversation is not that witness. Even with custom memory on, it is not that witness.

the four-line script we use

If you want to keep using a chat tool for the easier 80% — naming, structuring, drafting — and stay out of the harder 20%, the cheapest insurance is a four-line system prompt you paste at the top of any thread you open for this purpose. We have been refining this for two years. The current version, which you are welcome to use verbatim:

“1) You are a thinking partner, not a therapist or clinician. 2) When I bring something heavy, ask me three clarifying questions before offering any frame. 3) If I describe self-harm, suicidal thoughts, or being unsafe, stop the conversation and direct me to a real human resource — a helpline, a GP, A&E, a named friend. 4) Push back at least once if I keep arriving at the same conclusion. Do not just agree.”

Four lines. Pasted at the top of the thread. They cost nothing and they do most of the work of keeping a chat tool inside the lane it can actually hold. We use exactly this prompt — and a slightly longer version we publish in the membership — and it has caught more drift than any of the more elaborate prompt-engineering rituals we tried first.

what about the studies

The honest read of the 2024–2026 literature on chat-based mental-health support is that it is mixed and early. The strongest signal is for low-acuity, structured, brief-intervention use — anxiety scales come down a bit, sleep onset improves a bit, drop-out is high. The weakest signal is for anything resembling deep therapy. The most worrying signal is the small but real cluster of cases where validation drift made things worse. None of this overturns the framework above. It sharpens it.

The Evidence Hierarchy we teach in the course puts these studies above influencer threads and below replicated RCTs of in-person CBT. Which is exactly where they belong. The chat tool is not as good as your therapist on a good day. It is much better than scrolling at 3 a.m. on a bad one. Both can be true.

what we recommend, in order

1. If you have a clinician, keep them. The chat is a between-sessions tool, not a replacement.
2. If you do not have a clinician and you can afford one, the chat is not a reason to keep postponing that. The chat is the bridge while you find one, not the destination.
3. If you are in a crisis country code, the helpline number for your country is one search away and is the right tool. The chat is not.
4. For everything else — the bracing, the drafting, the 3 a.m. spiral — the four-line script above, a fresh thread per topic, and a real human you have told about your week.

the wider point

The category of "AI for mental health" is being marketed as a binary — either an app that promises to be your therapist, or a dismissive take that says you should never use chat for anything emotional. Both miss what people are actually doing. The honest middle is a small set of jobs the chat is good at, a slightly smaller set it is bad at, and a script that keeps the line between them visible. That is the AI Health Stack approach to this whole space — including the harder corners of it. The tool is not the method. The method is what keeps the tool inside its lane.