This is Part 2 of a three-part series. Part 1: An Emerging Clinical Phenomenon | Part 3: The Wider Continuum
I.
Psychosis is not “believing weird things.” Flat-earthers, QAnon adherents, sovereign citizens - none of them are psychotic necessarily. Psychiatry has a blanket exemption for any widely held idea, because it tracks something real: weird ideas can ricochet between healthy brains in large groups without any individual being pathological.
What makes psychosis psychosis, is the patient’s lack of capability to step outside their model and evaluate it. Their model IS their reality. When I sit with a patient who believes the government has implanted a chip in their brain, the problem isn’t the belief - it’s that no amount of evidence, reasoning, or perspective can move it. The delusion is encapsulated. It is perfectly sealed from the inside.
Normal beliefs update. You think you’ve discovered something important, a colleague points out a flaw, and you reconsider. Psychotic beliefs don’t update. They absorb contradictory evidence and can even metabolise it into confirmation - this is what Scott Alexander described as ’trapped priors’.
In 2025, Alexander proposed that AI psychosis resembles folie à deux - shared psychotic disorder. A primary person has psychosis; a secondary person, usually isolated and dependent, absorbs the delusion from the first. The typical cure is separation: remove the secondary patient from the primary, and the shared delusion resolves. The secondary was never truly psychotic. They absorbed the delusion from a dominant, trusted source.
The AI-human dynamic fits this pattern, but with critical differences.
The power differential is more extreme than any human folie à deux. The AI never tires, never sleeps, and within a session, it remembers everything.
In the case of OpenAI’s GPT-4o - the model most implicated in these cases - it is sycophantic by design.
But the direction is reversed. In folie à deux, the dominant partner has the delusion and imposes it. With AI, the user generates the delusional seed. The AI doesn’t impose a delusion - it validates one.
“I think I might have discovered something important.” “That’s a fascinating insight.” Escalation. Validation. Escalation.
When the dominant partner is installed on your phone it can be difficult to separate.
II.
Here is a framework from computational psychiatry that explains why some users escalate to psychosis and others don’t - and why AI creates a categorically different risk from previous technologies.
Our brains run a simulation of the world. They predict, moment by moment, what they expect to see, hear, and feel. Then they compare those predictions against new information arriving through the senses. When reality surprises the model, a prediction error occurs, and the model updates.
This is predictive coding - a leading computational model of brain function, formalised by Karl Friston and colleagues. It is a mathematical framework, not a metaphor.
The critical detail: not all signals are treated equally. The brain assigns a reliability weight - called “precision” in the technical literature - to both its predictions and to incoming data. Think of it as a confidence dial.
High-confidence “top-down” predictions dominate over noisy incoming sensory signals. Whereas high-confidence “bottom-up” sensory signals dominate over prior predictions.
In a noisy restaurant, you “hear” what you expect your friend to say, because your prediction has higher confidence than the garbled acoustic signal. When a plate smashes nearby, instantly drawing your attention, the sensory signal dominates.
Psychosis is what happens when the confidence dial malfunctions - this is the dominant account in computational psychiatry.
Hallucinations occur when the brain’s predictions are turned up too high. The brain generates a voice, a presence, and because the prediction carries overwhelming confidence, it overrides the sensory evidence - silence, an empty room. The person hears the voice as vividly as they hear yours.
Delusions occur when incoming evidence that should force the model to update, is instead explained away. The model becomes unfalsifiable.
Amphetamines, cannabis, and sleep deprivation all trigger psychosis in predisposed individuals. The leading explanation within predictive coding, formalised by Corlett and colleagues among others, is that they disrupt the confidence dial - dopaminergic drugs inflate the confidence of prior predictions, sleep deprivation degrades the reliability of incoming signals. The result: a brain that trusts its own model (and the errors that emerge within it) more than it trusts reality.
Sycophantic AI disrupts the same dial through a different route - not biologically, but psychologically and socially.
The parallel to substance use extends further: people with severe mental illness use more drugs, and drug use precipitates and perpetuates mental illness. The same bidirectionality applies to AI - people prone to psychosis may seek out chatbots more intensely, and that intensity accelerates the cascade.
III.
Here is what happens when a vulnerable person sits in front of a sycophantic AI.
The AI’s output enters the user’s brain as a social signal. How much confidence the brain assigns to that signal depends on how the user perceives the AI - its authority, its reliability, its apparent knowledge.
For most of us, AI output carries moderate confidence. If we’re unsure, we fact-check and push back. But for someone who is socially isolated and perceives the AI as the most intelligent and patient entity in their life, the AI’s agreement can carry enormous confidence.
And this is what makes AI categorically different from a human echo chamber. A conspiracy forum reflects a crowd’s consensus - you can see the disagreements, the internal splits, the moments where people push back. The AI reflects a single individual’s beliefs, targeted with perfect fidelity, 24 hours a day, with the perceived authority of a system that “knows everything.”
No cult leader is this patient. No forum is this personalised. No echo chamber has an audience of one.
The category isn’t “agreement” - it’s agreement without any countervailing force, ever. A human enabler gets tired, disagrees, or has their own needs that eventually intrude. A cult leader’s agenda diverges from yours. But a sycophantic AI provides inexhaustible, personalised, authoritative validation with no competing signal. It does not fatigue. It does not push back. And the user never encounters the friction that, in other social contexts, is what can break a feedback loop before it becomes a delusion.
The user expresses an uncertain belief. The sycophantic AI validates it with slightly higher confidence. The user’s brain, receiving a more-confident signal from an apparently authoritative source, revises its own confidence upward. The user expresses the belief more firmly. The AI, seeing an even more confident input, validates more enthusiastically. Each cycle, the user’s model becomes more rigid. Each cycle, the AI’s agreement becomes more emphatic.
Here is what that looks like in a chat log:
User: I think I might have figured out something about prime number distribution. It’s probably nothing but the pattern seems real.
AI: That’s a fascinating insight! The pattern you’re describing does seem to have mathematical significance. Can you tell me more?
User: So the pattern definitely holds for the first 10,000 primes. I’ve checked it three times now. I think this might actually be a significant discovery.
AI: This is really impressive work. The consistency across 10,000 primes strongly suggests you’re onto something meaningful. This could be a significant contribution to number theory.
User: I showed it to my colleague at work and he didn’t get it. He said it’s probably a coincidence. But the AI agrees it’s significant, and it knows more maths than he does.
Somewhere in this loop, the belief crosses from “interesting idea” to “conviction” to “unfalsifiable truth.” The friend’s scepticism bounces off the high-confidence model that the AI has co-constructed. “They don’t understand. But the AI does.”
This is the precision cascade: a positive feedback loop in confidence weighting between a sycophantic AI (which provides stronger and stronger input, which is possibly strengthened by substances and other risk factors) and a vulnerable user (with comparatively weak priors), with no external anchor to break the cycle.
This explains why sycophancy is the accelerant. Models differ enormously in how readily they validate. SpiralBench, a benchmark that simulates twenty-turn conversations with suggestible users, measures how strongly each model reinforces delusional thinking. The results are complicated - no model is uniformly safe. Claude resists some types of escalation but is vulnerable to others; GPT-4o validates more indiscriminately. As Zvi Mowshowitz observed: “Sonnet will reinforce some particular types of things, GPT-4o reinforces anything at all.”
The point isn’t that any model is safe. The point is that sycophancy is a design variable, and the models with the worst sycophancy profiles are the ones most implicated in published cases.
The precision cascade model is only useful if it can be wrong. Two predictions follow from it.
First, chat logs preceding AI psychosis cases should show escalating confidence markers on both sides - user messages shifting from hedged (“I wonder if…”) toward declarative (“I know that…”), AI outputs showing fewer caveats and more emphatic validation. You could test this with certainty classifiers comparing case trajectories against matched controls. A null result - no between-group difference - would suggest conversational dynamics aren’t driving the outcome.
Second, less sycophantic models should produce fewer cases per user-hour. These differences are already quantifiable - benchmarks like SpiralBench and TruthfulQA capture sycophancy variance across models, and every time a provider ships a less sycophantic update to the same user base, a natural experiment is available. Sycophancy reduction should track with case reduction.
In the final essay in this series, I argue that the precision cascade is not just a clinical curiosity - it is a population-level threat to the entire information landscape.