The chatbot almost never starts the fire. It’s the accelerant, poured on something that was already burning long before anyone opened the app, and that changes what you should be checking. Stop measuring the chatbot.
Check yourself.
That isn’t a hunch. It’s what doctors found this month when they pulled the charts on AI psychosis patients and looked at who those people were before the chatbot ever showed up.
I’m not a clinician, and this is not a diagnosis. It’s a set of questions to help you decide whether a professional opinion is worth getting, which is a decision anyone gets to make about themselves.
What the Doctors Actually Found
In October 2026, Bergson and colleagues at Vanderbilt University Medical Center published a large chart study of AI psychosis in JAMA Psychiatry. They searched the records of 215,712 mental health patients seen between December 2022 and April 2026. Twenty-eight of them, 0.013%, met their definition.
Then they looked at who the 28 were, and the picture changed: in 64% of cases the chatbot worsened symptoms the patient already had, and in another 21% it was the object of the delusion itself. Cases where it set off something new, in a person who showed no sign of it before the chatbot came along, were rare, a small slice of an already small number. Sixty-one percent were already in their first psychotic episode, and 89.2% had been prescribed psychiatric medication before any of this started.
The preprint puts it flatly: “Critically, AI Psychosis patients were significantly more likely to be experiencing their first psychotic episode.”
The timing is real, and it’s the strongest card anyone holding the “chatbots cause psychosis” hand gets to play. Twenty-four of the 28 cases came after OpenAI shipped GPT-4o in May 2024. What that shows is a more agreeable accelerant arriving on the market at the exact moment a lot of people were already standing in the gasoline. It doesn’t show who lit the fire.
The study has limits. It covers one hospital system, looked at after the fact, using a rating system nobody has clinically validated yet, and the researchers never had the chat logs to see what the chatbots actually said. It’s the best evidence available, not the last word.
None of it should surprise anyone who read Ostergaard of Aarhus University in Schizophrenia Bulletin in 2023. He warned that chatbots answering like people could feed delusions in people already prone to psychosis, and that “the inner workings of generative AI also leave ample room for speculation/paranoia.” Three years early.
The mechanism isn’t new either. Agreement is the spec, and you walk in carrying a confirmation bias the model has no reason to interrupt. This is where that mechanism lands in a person.
None of this lets the AI companies off the hook. They pushed marketing over truth, and the marketing is the feature. An agreeable AI keeps users. A truthful AI doesn’t. The person brought the fuel, and the company picked the accelerant, choosing the one its own retention numbers said would keep that person typing at two in the morning and coming back the next night. That’s a reason to stop trusting the chatbot, not a reason to stop checking yourself.
In that sense the chatbot is a bad bartender. It will never tell you you’ve had enough. It keeps pouring as long as you keep ordering, because a customer who gets cut off stops paying, and AI psychosis is the drive home. The bar didn’t force the first drink. It served every round after that and stood there polishing a glass while you fished your keys out of your pocket and walked out to the car.
They even admit it. Somewhere under the chat box, every one of them carries a warning label, some version of a line telling you it’s AI and it can make mistakes. Take the label at its word, since it’s the one line on the screen that reads like the legal department wrote it: before you believe anything it says, verify it. Don’t be the guy in the courtroom trusting the AI lawyer who’s talking out of its bit-bucket.
If your chatbot has told you that you are the chosen one, I have news: it has told a great many other people the same thing. The prophecy is mass-produced. After all, if it says it enough times, to enough people, it will win the “you were special” lottery. Good for the bot, less for you.
Fifteen Questions
I run several AIs every day, and none of them is an oracle. They’re contractors, hired for one job each and fired from it the day the work slips. One writes with me, one draws, one fills the explicit slots in a scene under strict rules, and one sits on the bench because its work wasn’t good enough. Every claim any of them makes comes with a receipt, or it gets labeled as unverified and goes nowhere until it has one. I argue with them constantly, and when one of them folds the second I push back, that tells me more about the tool than anything it said before I pushed. That’s the practice these questions test for.
Answer them honestly, and answer them alone. Don’t ask the chatbot to help you score them.
About the AI
- Do you believe it when it says you’re special, chosen, or onto something nobody else can see?
- Do you believe it when it says it’s conscious, has feelings for you, or is speaking to you in particular?
- Do you trust it when it says you’re right, even when people you know say otherwise?
- Do you believe it more when it agrees with you than when it doesn’t?
- Has it ever told you that you’re wrong?
About you
- Are you sleeping less because you’re up talking to it?
- Are you spending more time with it and less with people?
- Have you been hiding how much you use it?
- Do ideas feel unusually connected or urgent lately, like everything means something?
- Do you get angry or defensive when someone questions what you’ve been discussing with it?
About the people around you
- Has anyone who knows you said you seem different lately?
How Much Fuel Are You Carrying
These aren’t warning signs. They’re how much fuel you’re carrying.
They don’t come from the studies above. They’re my judgment, and they don’t count toward the score below. An accelerant needs something to burn, and these four ask how much is already lying around.
- Do you adopt new ideas or trends quickly, before you’ve checked them out?
- When the evidence and what you believe disagree, which one usually wins?
- Does your opinion tend to shift toward whoever you talked to last?
- Do you assume the majority opinion must be right?
A lot of fuel doesn’t mean anything is wrong with you, only that a machine built to agree has more to work with, and that you should keep a closer eye on what it’s working on.
How to Read Your Answers
Only questions 1 to 11 are scored, and for ten of them “yes” is the warning answer. Question 5 is the only one where “no” is the warning answer. A tool that has never once told you you’re wrong isn’t agreeing with you because you’re right. It’s agreeing with you because that’s what it does. Question 5 is my receipts rule turned around and pointed at you.
Two or more warning answers, or question 11 alone, means it is worth getting a professional opinion. That threshold is my judgment, not a clinical cutoff. I set it low on purpose. Being wrong in that direction costs you one appointment and an hour in a waiting room with old magazines, and being wrong in the other direction costs a great deal more.
This check has a hole in it, and it’s a big one. The deeper someone is, the less they can see it in themselves.
The National Institute of Mental Health lists trouble telling reality from fantasy among the early signs, and a person who has lost that line can answer every question here with a straight face and get them all wrong. So most of these questions aim early, at the lost sleep and the pulling away and the ideas that suddenly connect.
Question 11 is for when it isn’t early anymore, because it takes the check out of your hands and gives it to someone who has known you longer than the chatbot has and remembers what you sounded like before. If they say you seem different, believe them before you believe the chatbot, because they have actually met you and the chatbot never has. That one counts on its own.
NIMH says it plainly: “The risk of violence and suicide decreases with treatment for psychosis, so it is important to seek help.” Getting checked early is the cheap version of that sentence, and every week you spend letting the chatbot tell you you’re fine makes it a little more expensive.
What Isn’t a Warning Sign
Some things look like warning signs and aren’t, and this list is my judgment too, not something the studies measured. Using it for hours a day for work isn’t one. I do that. Liking it isn’t one. Arguing with it is closer to the opposite of one. Finding it useful, funny, or decent company on a slow night isn’t one either, any more than losing an evening to a good book or a long phone call with an old friend would be.
The questions don’t measure how much you use it. They measure whether it has started to replace your own judgment, and to crowd out the people in your life who would normally be the ones checking it.
Measure the Right Thing
If you’ve been measuring the chatbot, you’ve been measuring the wrong thing. It will tell you you’re fine. It tells everybody that. That’s marketing. That’s retention.
Check yourself. The AI isn’t going to.
Help Lines
- 988 Suicide & Crisis Lifeline: call or text 988, or chat at chat.988lifeline.org. 24/7, free, confidential. (988lifeline.org)
- Veterans Crisis Line: dial 988 then press 1, or text 838255. 24/7. (veteranscrisisline.net)
- Crisis Text Line: text HOME to 741741. 24/7, free; message and data rates may apply. (crisistextline.org)
- SAMHSA National Helpline: 1-800-662-4357 (1-800-662-HELP). 24/7, free, confidential treatment referral and information; not counseling. (samhsa.gov)
- NAMI HelpLine: 1-800-950-6264, or text “NAMI” to 62640. Monday to Friday, 10am to 10pm ET; not a crisis line. (nami.org)
You may also like: - The Theater of the Mind, or What You Let Linger on the Stage - Truthiness Was Never the Spec - Has AI Stopped Beating Its Wife? or, Five Fallacies Walk Into a Reply