Nobody talks about this one. Not because it’s secret — but because it sounds a little odd when you say it out loud: using ChatGPT or Claude to work through anxious or spiraling thoughts.
Not as therapy. Not as a replacement for talking to someone. But as a place to externalize what’s rattling around in your head at 11pm when you don’t want to wake anyone up, or when you’ve already talked about the same thing enough times that you feel embarrassed bringing it up again.
It works better than most people would guess. And it fails in a specific way that’s worth knowing about before you try it.
Why Externalizing Works At All
There’s a well-established cognitive phenomenon where simply translating an internal worry into words reduces its hold on you. The mechanism isn’t mysterious: when a worry stays inside your head, your brain keeps re-running it as raw sensation and undefined dread. Putting it into a sentence forces you to structure it. Structured things are easier to evaluate.
This is why journaling exists. It’s also why talking to someone helps even when they don’t say anything useful back. The act of articulation is doing real work, independent of whether the listener is any good.
AI fits into that same slot. You type “I’m worried that X will happen and that means Y” and something on the other end responds — which keeps the externalization loop going rather than letting it collapse back into a formless spiral.
Where It’s Actually Useful
The most productive version I’ve found: using the model as a thinking partner to stress-test whether a worry is well-formed or not.
A lot of anxiety is structurally imprecise. You feel dread about a situation but if you had to state the specific outcome you fear, you couldn’t easily do it. Asking Claude or ChatGPT “what am I actually afraid of here?” — after you’ve described the situation — sometimes produces a useful reframe, not because the model is insightful, but because the question itself forces you to answer it.
You can also use it to war-game. Describe the thing you’re worried about, then ask: “What would need to be true for this to actually go badly? What would need to be true for it to be fine?” That conditional structure pulls you out of vague catastrophizing into something you can actually think about.
Another real use: drafting the hard conversation you’re dreading. Not to get a script — scripts break the moment they meet a real person — but to figure out what you actually want to say. Running through it once in a low-stakes environment makes the actual conversation less loaded.
What Doesn’t Work
The model will validate you. Constantly. This is the failure mode.
Describe a situation where you’re probably overreacting, and both Claude and ChatGPT will still find something sympathetic to say about your feelings. They’re not going to tell you that you’re being irrational, even when you are. That’s partly by design — they’re optimized not to say things that feel harsh — and partly a statistical artifact of training on human-generated text where direct challenges are rare.
So if what you need is genuine pushback — someone to tell you that you’re catastrophizing, that you’ve been through this exact spiral before and it resolved fine — the model won’t reliably give you that. It’ll acknowledge your concern and then offer “perspectives from both sides,” which is often worse than useless when you’re already running both sides in your own head at 3x speed.
You can prompt around this, somewhat. Something like: “Don’t try to make me feel better. Just tell me if the reasoning I laid out is sound.” That shifts the behavior noticeably. But there’s a floor of diplomatic softening you can’t fully prompt away.
The other limit: the model doesn’t know you. It doesn’t know your history with this kind of worry, whether this is a pattern, or what you’ve tried before. It’ll offer generic grounding exercises or suggest you “talk to a professional” in situations where what you actually need is someone who knows the context. A real friend with that context beats the model every time. The model just happens to be available at midnight.
A Practical Setup If You Want to Try It
Don’t open a blank chat and start complaining. That tends to produce meandering output that mirrors your own lack of structure back at you.
Start with a short framing message. Something like: “I want to think through something that’s worrying me. I’m not looking for reassurance — I want to understand whether the worry is well-founded and what, if anything, I should actually do about it. I’ll describe the situation, and I’d like you to ask me questions rather than jumping to advice.”
That last part matters. Getting the model to ask questions rather than immediately offer suggestions keeps you doing the cognitive work rather than outsourcing it. If it asks you “what outcome are you actually trying to avoid?” and you have to answer that, you’ve done something useful — regardless of what it says next.
Claude tends to be slightly better than ChatGPT for this kind of reflective conversation, in my experience. Not dramatically — both work — but Claude seems more willing to stay in “probe and question” mode without defaulting to a bulleted action plan after every message.
The Limit Worth Being Clear About
This is a thinking tool, not mental health support. The difference matters.
If what you’re dealing with is chronic, severe, or has been affecting your daily functioning long enough that you’d notice it — a chatbot is not the right tool. Low-friction availability is a feature for occasional “I need to externalize this” moments. It can quietly become a substitute that delays getting real help if you’re not paying attention to that line.
For the ordinary anxiety of decisions, social situations, work stress — things you’ll likely be fine about in two weeks — it’s one of the more underrated uses of these tools. The prompt structure above is the part most people skip. That’s where most of the value actually lives.
