Which AI is best for anxiety? Pick the job, not the brand.

“Which AI is best for anxiety?” is one of the most-typed questions in this whole corner of the internet, and almost every answer to it is a ranked list that misses the point. There is no anxiety model. No lab trained a system to be good at your 2 a.m. spiral. What exists is a handful of general tools that happen to be good at three very different jobs — and on an anxious night you reach for whichever one is already open and quietly hope it does all three at once.

It won’t. Not because the tools are weak, but because the jobs are different, and the right tool depends entirely on which job you are actually doing. So the useful version of the question is not “which AI is best for anxiety” but “which job am I doing right now, and which tool fits it.” Answer that and the brand mostly stops mattering.

the three jobs hiding inside one question

When people say “help with anxiety,” they usually mean one of three things without separating them. Pull them apart and each one has an obvious tool.

1. Understanding — you want to know what a technique actually is, what the evidence says, whether a symptom is worth worrying about. This is a research job, and it wants a model that will give you citations you can check, not warm reassurance.
2. Structuring — you have the spiral and you want a frame around it: a thought record, a worry-time slot, a breathing protocol run faithfully at 3 a.m. This is a structuring job, and almost any capable assistant does it well.
3. Drafting — there is an email you cannot send or a boundary you cannot phrase, and the anxiety is really about the unsaid thing. This is a drafting job, and the tool you already pay for is almost certainly fine.

Notice that none of these three is “be my therapist.” That is the fourth job, and we will get to why it does not belong on this list at all.

so, by the job — which tool?

We frame this the way we frame the whole AI Health Stack: by the job, not the logo. Here is the honest routing, and where the differences between brands actually show up.

1. For understanding — use a research-grade model and explicitly ask for sources. The frontier models from the major labs are close enough here that the deciding factor is the habit, not the brand: demand citations, then check one. A tool that cannot show its working is the wrong tool for this job regardless of name.
2. For structuring — use whatever assistant you already have open. This is the job where switching brands changes the least. What changes the outcome is the prompt: give it the structure (a CBT thought record, a 4-7-8 breath, a five-minute worry slot) and make it run that, not improvise comfort.
3. For drafting — again, the one you pay for. The edge here is privacy, not capability: prefer the tool that lets you keep, export and delete what you typed, because the email you could not send is some of the most personal data you own.

the job no AI should take

There is a fourth thing people quietly want when they ask which AI is best for anxiety: a judgement about how much trouble they are in. Whether this is normal worry or something that needs help. That is assessment, and no general AI can do it safely. It has no training, no licence, no duty of care, and crucially no way of knowing what it does not know — delivered in the same calm, fluent tone whether it is right or catastrophically wrong.

• It will normalise rumination — a person changes the subject; a model will circle the same fear with you for an hour and call it support.
• It will collude with avoidance — ask it to help you not have the hard conversation, and it will, warmly and at length.
• It cannot manage acute risk — if anxiety tips into panic that won’t pass, or into thoughts of self-harm, a chat thread is not the tool. Contact a crisis line or emergency services.

the actual answer

So: which AI is best for anxiety? None of them, and three of them. None, if you want a therapist in a tab. Three, if you are willing to do the one piece of work the tools cannot — naming the job before you open the app. Do that, and you stop shopping for a better chatbot and start building a small, deliberate division of labour between the tools you already have and the human you cannot replace.

That is the whole move. Not a ranked list. A routing decision you make before you type a word — which is, conveniently, also the calmest thing you can do at 2 a.m.