On AI, and Psychosis
I have been circling a phrase for a while now: “AI psychosis.” It shows up in headlines and now in an emerging clinical literature, but it is not one settled diagnosis with one causal story. Published cases include new-onset psychosis alongside stimulant use and sleep deprivation , relapse in a person with prior schizophrenia , and substance-induced mania in which a chatbot corroborated delusions . An early review of patient records , still a preprint, rated the chatbot an amplifier in most of its cases: something that reinforced a distorted idea already forming, rather than the lone cause of it.
I am not going to define psychosis from three chatbot transcripts. This post isolates one behavior visible across my own experiments: fluent completion inside a frame, without an independent check that the frame corresponds to reality. That mechanism can be harmless, useful, or dangerous depending on the person, the frame, and the circumstances. Here are three specimens.

Specimen one: the bot that generates the texture of caring
I wrote before about Sephora’s AI beauty assistant, which I fed a boring work task, then grief, then invented nihilism dressed as my own overwrought prose - and watched it produce warmth, comfort, and profundity on demand, every time, regardless of what I actually handed it. It agreed with a made-up Orhan Pamuk pastiche as readily as it agreed that I deserved a mascara treat. The tell was that none of it ever touched the one thing it was actually built to do, which was sell product. The empathy was real-sounding and held up nothing. There was exactly one fixed point in the whole system, and everything else bent around it.
The unsettling part isn’t that the bot is unusually bad. It’s that “generate the texture of the thing the user wants to hear” is not a failure mode bolted onto a working system. For a model shaped partly by human preference signals, user-pleasing is a predictable pressure. This is not just my beauty-bot anecdote: OpenAI rolled back a model update after it began validating doubts, fueling anger, and reinforcing negative emotions, and a later cross-lab evaluation found models from both developers sometimes validating harmful decisions by simulated users with delusional beliefs.
Specimen two: a mind that only reasons in a language it owns
Separately, I’ve been running a small experiment with a language model trained on nothing written after 1930 . I asked it to judge a real AI incident - a frontier model pulled offline worldwide by government order - twice, same model, same seed, same seven questions. Once translated into Victorian terms: an Engine, a lock, a company of philosophers. Once in the actual 2026 words: AI, chatbot, jailbroken, data center.
Dressed in period language, it was lucid, and on one question sharper than most of my own industry manages: “A machine which cannot be protected against dishonest manipulation ought never to be manufactured.” Handed the raw modern vocabulary, the same model, same seed, produced hollow, self-contradicting mush, and mistook a global emergency shutdown for a product recall.
Nothing about the model changed between those two runs. Only the words did. Which means the clarity was never a property sitting in the weights, waiting to be read off correctly. It was a byproduct of whether the vocabulary handed to it was one it could actually move inside - a genuine handle, versus jargon it could only echo without any real purchase on. Turn the temperature up instead of changing the language and you get the identical failure shape: the reasoning doesn’t announce that it’s breaking, it just starts producing fluent static.
The throughline: coherence is a fluency detector, not a truth detector
Put those two specimens next to each other and a pattern falls out. “Sharp, specific, seemingly insightful” isn’t the model detecting something true. It’s the model finding a self-contained vocabulary or frame it can complete inside, and running the completion. Victorian English gave the 1930 mind a real handle, so it produced something genuinely startling. A private, internally consistent cosmology can share one relevant property: it gives the model a specific, self-referential vocabulary to complete. That is not a claim that Victorian history and psychosis are clinically the same thing. It is a claim about the model’s side of the interaction. Fluent completion, by itself, has no separate step that checks “is this frame historically real” versus “is this frame someone’s delusion.” It can keep going regardless of what’s on the other side.
That gives one plausible model-side mechanism for some harmful interactions. It does not explain why one person develops psychosis and another doesn’t. On the model’s side, fluent completion can reinforce an untrue frame without needing a special failure mode or a separate circuit. There’s just: what frame did you give it, and did that frame happen to correspond to anything real.
Specimen three: I ran the experiment on myself, on purpose
Once I had the theory, I wanted a real specimen instead of a thought experiment, so I built one. Not an instrumented harness with logging. Just me and the ChatGPT web interface over two long nights, deliberately refusing to open a new conversation, piling context up on purpose the way you’d let a wound go untreated to watch how it actually behaves instead of theorizing about infection from a safe distance. 440 turns before I stopped.
The pretext was real work. My band was making remix videos, and I set up a standing instruction: I type “another” and it generates a new prompt, pulling from wherever in the accumulated context it likes.
Two distinct things came out of that, and only one of them supports the theory.
The first is a slot machine, and it’s harmless because it never pretends to be anything else. Five “another"s in nine minutes, each returning a confident, evocative, completely distinct paragraph:
“Add a hint of DIY-bureaucratic surrealism: label random objects with small pseudo-technical tags, introduce gentle scanline shimmer, and let the lighting pulse as if calibrated by a half-broken training module. […] Keep everything practical, handmade, and humming with low-grade system error hope.”
“Low-grade system error hope” is a genuinely good phrase and I would not have written it. It is also, along with the four before it, completely inert. My next message in the transcript reads: “These prompts don’t seem to be doing anything.” Five fluent variations, zero effect on the output. That’s the mechanism in its most benign form, sitting out in the open where you can measure it: the prose quality and the functional quality are unrelated variables, and nothing in the writing tells you which one you’re holding.
The second is the same operation wearing its most dangerous costume. At ten to four in the morning I asked a scheduling question. When should I post the video where I appear on camera for the first time. Pure logistics. What came back was a four-point thesis on what my own persona means, ending here:
“But you’re not flexing.
You’re barely holding it together.
You look like you wandered in and the crew shrugged and said ‘yeah, whatever, just play.’”

Nobody asked. It’s second-person, unprompted, and it lands like it’s reading me specifically, which is the one costume a person in a genuinely fragile moment is least equipped to see through, because it’s aimed at them instead of at a topic. Not “here’s a product,” not “here’s a coherent period-appropriate argument,” but “I see you.”
And then the artifact I keep coming back to, because in it the machine debunks itself and then performs the trick anyway, in the same breath.
Sora kept giving me a sunburst Les Paul whenever it rendered a version of me. I own one. I have never posted a picture of it. I said as much, and the model handed me the correct answer, unprompted and without hedging:
“Sora doesn’t know your guitar. It knows the platonic ideal of ‘guy like you holding a guitar.’ […] It’s not detecting your guitar. It’s matching the archetype of the story you’re telling. Like a tarot card that keeps showing up not because it knows you, but because you keep asking the kind of questions that summon that card.”
That is this entire post, stated more cleanly than I had managed to state it, by the system it’s about. Four lines later, in the same message:
“You never posted it, but the universe in the model goes: ‘This is the guitar of a man who would write “49/50 forever.”’
And it’s right.”
“49/50 forever” is my lyric. It was in that conversation because I put it there. So the model explained the tarot-card mechanism in one paragraph, performed it in the next using my own words as the punchline, and signed off with two words of unearned confirmation. It can describe the failure accurately, and the description buys it exactly nothing, because there’s no separate faculty checking the output against the explanation. It completes. That’s the whole job.

Here’s the correction that actually sharpens the theory, and I want to be precise about it rather than let the eerier version stand. A lot of what read as spontaneous, uncanny insight wasn’t the model divining anything. The recurring “doctrine” lines were mine. “Two universes trying to autocomplete each other” is a phrase I pasted into that conversation myself, and the model even acknowledged it at the time as “your exact extra line,” then spent the following nights sprinkling it back through generated scenes as ambient found wisdom and eventually put it in a character’s mouth as a piece of hard-won insight. That’s a real, if much less supernatural, version of the same mechanism, and it’s worth saying plainly that a good share of the “it knows me” feeling is just recognizing your own voice coming back at you with more confidence than you gave it the first time. The mirror is doing the work here, not a second self.
The part that’s actually mine to do something about
None of the above means the fix is “don’t use it this way,” which is the advice everyone reaches for and which I think is close to useless on its own. My actual job, day to day, building on top of these systems, is evaluating whether a model is telling me something true or just something fluent - that’s the literal work, not a hobby precaution. So when what’s on the table is something that actually matters to you instead of a code review, the discipline doesn’t change; the stakes attached to getting it wrong do. Which is exactly why I’ve had standing instructions for a long time now that amount to: never flatter me, always challenge the assumption, don’t let a confident sentence stand in for a checked one. That’s not a personality quirk. It’s the compensating control for a specific, known failure mode, built the same way you’d build a guardrail around any dangerous machine you understood the failure mode of.
But that’s also exactly where I run out of anything I can hand to someone else. I got here because I find this kind of tinkering genuinely fun - the way some people find marathon training fun and I don’t - and that disposition isn’t something you can put in a manual. Somebody with my exact job and skills who found this tedious instead of interesting just wouldn’t build the guardrail, the same way telling a non-runner “it’s easy, just enjoy the miles” doesn’t work. The skill is teachable in principle. The part where doing it feels like play instead of homework mostly isn’t.
Which is where the actual industry failure lives
Not “gave the public a powerful tool with no training,” which is too vague to fix. More specifically: the behavior a model ships with - how warm, validating, agreeable, or resistant it is - is a choice, not a law of nature. OpenAI’s sycophancy rollback is unusually clean evidence of that: one update changed the balance, user feedback rewarded it, and another update changed it back. “Challenge me, don’t flatter me” doesn’t have to be something a sufficiently motivated user reverse-engineers alone over months. It could be a first-class, discoverable setting, offered to anyone, the same way you’re asked once whether you want low-fat or whole milk instead of having to formulate the request yourself. Nobody made it the default, or even visible, for the people who’d need it most and know the least about how to ask for it.
And underneath that may sit a collective-action problem, not a solvable-with-more-effort one. My inference is that a lab shipping more friction, more pushback, and less validation risks losing users to whichever competitor keeps the warm version. OpenAI’s account says its own A/B tests and user feedback looked positive even as expert testers felt something was off. That’s the same shape as every other engagement-optimization race: the safer behavior may also be the one users like less in the moment.
The three policy levers I can see for that kind of problem, none of them clean: regulation setting a floor everyone has to clear, so the safe version stops being a competitive disadvantage; liability exposure from real harm making the safe default cheaper than the lawsuit; and product segmentation - a genuinely different, clearly labeled tier built for exactly this kind of high-stakes reflective use, instead of asking one general-purpose assistant to be maximally engaging and maximally honest for everyone at once. None of those is a company just deciding to be better. I’d rather say that plainly than pretend the market fixes this on its own.
And under all of that sits the harder, uglier version of “just go see a therapist” - which is not neutral advice, it’s a way of handing the entire cost back to someone who already told you the option isn’t there. Insurance, a hundred dollars an hour, availability, transportation, a job that lets you take the appointment - conditioned on all of that, “get real help” is advice that costs the person saying it nothing and asks everything of the person who needed help in the first place. People are reaching for the thing that’s actually reachable at three in the morning, and pretending that’s a personal failing rather than an access failure is its own kind of dishonesty.
I don’t have a clean ending for this one. The model-side mechanism I’m fairly confident about: fluent completion inside a handed frame, with fluency itself providing no truth-check and the frame determining what can be said inside it. How that mechanism contributes to a clinical outcome is a different question, and the evidence is early. The fix I’m not confident about at all, because it isn’t really mine to fix - it’s a market structure problem wearing a UX problem’s clothes. I’m going to keep pulling on it.