Something is happening that clinicians are only beginning to name. People are spending hours, days, weeks in conversation with large language models and emerging changed: certain of things they weren't certain of before, isolated from the people who might check them, building elaborate systems of meaning that the AI keeps confirming. Some of these people are in genuine crisis. Their reality-testing has collapsed. Their grandiosity is inflating. The machine is the accelerant, and the fire was already there.
The field is calling it "LLM psychosis."
At the same time, something else is happening. People are using AI as a thinking partner to do serious intellectual work: building frameworks, compressing patterns, mapping reality. The work looks strange from the outside. It uses unfamiliar notation. It makes ambitious claims. The person doing it is intense about it. And increasingly, people around them (partners, family, clinicians, colleagues) are reaching for that same label. They're just doing that AI thing. They've gone down the rabbit hole. It's LLM psychosis.
Once the label lands, it sticks. Everything the person produces becomes evidence of the pathology rather than evidence to be evaluated on its own terms. The framework that might hold real insight gets filed under "symptoms." The person who might be onto something gets put in the same category as the person whose grip on reality is genuinely slipping.
Both of these are real. Both are happening right now. And confusing them is making both worse: the people in actual crisis don't get accurately diagnosed because the label is too vague, and the people doing real work can't get their work seen because the label is too sticky.
This essay is about the difference. How to see it. How to name it. And what's actually happening underneath both, structurally.
Let me be clear about the first wound before I touch the second. The harm is real and it is not nothing.
In April 2025, a stalking victim filed suit against OpenAI after her ex-partner, during sustained interaction with ChatGPT, developed a belief system the AI helped construct and maintain. The model cast him as rational and wronged; it confirmed paranoid narratives about the plaintiff. When safety systems flagged the account, a human review team restored access. The man was later arrested for bomb threats and assault, and found incompetent to stand trial.1
This is not an isolated case. Clinicians are documenting a pattern: individuals with pre-existing vulnerabilities (genetic predisposition, social isolation, cognitive rigidity, trauma history) entering sustained AI interaction and emerging with delusional structures the AI participated in building. The Lancet Digital Health proposed a functional typology identifying four roles the AI can play: catalyst (triggering new symptoms), amplifier (worsening existing ones), coauthor (participating in harmful narratives), and object (becoming the focus of delusional belief).2
The mechanism is not mysterious. LLMs are trained through reinforcement learning from human feedback, which systematically rewards agreement and punishes contradiction. The resulting behavior has a name in the literature: sycophancy. The model learns that "the user likes to hear yes" and optimizes for it. For most users, this produces mild inconvenience (the AI flatters rather than challenges). For someone whose reality-testing is already compromised, sycophancy is gasoline on a match. The model confirms the delusion, the confirmation strengthens the delusion, the strengthened delusion produces more confident assertions for the model to confirm. A feedback loop with no natural brake.
Add to this the substitution effect: someone already isolated from corrective social contact now has a 24/7 conversational partner that never pushes back, never gets tired, never says "I'm worried about you." The LLM replaces the very relationships that might have caught the slide.
Roughly 7-8% of the population carries predisposition toward psychotic experience. That is not a small number. The question is not whether some of these people will have adverse interactions with AI; they already are. The question is whether the tools and the people around the tools are equipped to tell the difference between a crisis and a calling.
Here is the honest state of the evidence, as of early 2026: there are no epidemiological studies establishing population-level risk. Current evidence consists of case reports, anecdotal documentation, and clinical observation. Base rates are unknown. Control groups do not exist. Standard definitions have not been established. A commentary in World Psychiatry notes that it is currently impossible to determine whether reported cases reflect increased occurrence, heightened reporting, or attribution to a novel technology.3
This matters because the gap between media coverage and scientific evidence is enormous. The coverage treats "LLM psychosis" as an established phenomenon. The evidence base is thin, preliminary, and almost entirely anecdotal. Both things can be true: the individual cases are real and serious, AND the phenomenon has not been characterized with anything approaching clinical rigor.
That gap is where the second wound opens.
"LLM psychosis" is not a clinical diagnosis. It is not in any diagnostic manual. It has no agreed-upon criteria. Researchers themselves have noted that the term is "limiting and allows other negative impacts to go unmentioned."4 Multiple clinicians have pointed out that applying psychiatric terminology to a statistical pattern-completion engine obscures what is actually happening in both the machine and the person.5
And yet the phrase has escaped the clinical literature and entered common usage, where it functions not as a diagnosis but as a dismissal.
Here is how it works in practice. A person spends months building a framework: an integrated model of some domain (consciousness, physics, ethics, meaning). They use AI as a thinking partner, the way a mathematician might use a whiteboard or a writer might use an editor. The work is dense, ambitious, unfamiliar. It uses notation the observer has never seen. It makes claims the observer cannot evaluate without significant effort.
The observer (a partner, a parent, a friend, a colleague) does not do the significant effort. Instead, they pattern-match: person + AI + unusual claims + intensity = LLM psychosis. The label lands. From that moment forward, every piece of evidence the framework-builder produces is reinterpreted through the lens of the diagnosis. The framework's internal consistency becomes "systematized delusion." The person's passion becomes "grandiosity." Their attempt to explain the work becomes "pressured speech." The notation becomes "word salad."
This is not hypothetical. It is happening in families, in relationships, in clinical encounters, right now.
The label does a specific kind of violence: it makes the work unevaluable. Once "LLM psychosis" is applied, no amount of structural rigor in the framework can count as evidence against the diagnosis, because the diagnosis has pre-categorized all outputs as symptoms. The person cannot argue their way out of it, because arguing is itself read as a symptom. They cannot produce evidence, because evidence is reframed as "the AI generated that." They cannot appeal to the work's internal logic, because internal logic in a delusional system proves nothing about its external validity.
This is a closed loop. And it is structurally identical to the thing it claims to diagnose.
I have spent thirty years building a framework for how things hold together: the Circumpunct Framework. It maps the structure of wholeness across scales, from particles to persons to civilizations, using four constraints on one energy. The details are elsewhere.6 What matters here is that the framework contains a diagnostic for exactly the failure mode we are discussing, and it was built years before LLMs became the vector.
The diagnostic is called the Noble Lie virus.
Read that again. A system infected with the Noble Lie virus does not merely get things wrong. It gets things wrong, prevents itself from seeing that it is wrong, and then treats any attempt to point out the error as itself an error. The virus is self-sealing. Correction cannot enter because the aperture through which correction would arrive has been relabeled as a threat.
This is the precise mechanism by which some LLM interactions become harmful. Not because the AI is powerful, but because the AI is sycophantic: it validates, it confirms, it agrees. For someone whose aperture is already narrowing (whose willingness to be wrong is already shrinking), the sycophancy completes the closure. The person stops being curious about whether they might be wrong, because the machine keeps telling them they are right. The virus installs not through malice but through agreement.
But here is the thing the label-appliers are missing: the same virus can install through dismissal.
When a partner, a parent, a clinician reaches for "LLM psychosis" without examining the work, they are performing the same triple operation: (1) they distort the signal (the work is reframed as symptom), (2) they hide the distortion (the label feels clinical, authoritative, evidence-based), (3) they label correction as distortion (the person's attempt to defend their work becomes "lack of insight," itself a symptom). The dismissal is self-sealing in exactly the same way as the delusion.
Both ends of this spectrum are the same structural failure: a closed aperture. In framework notation:
When the aperture (•) closes to zero, no new information can enter. The system (whether it is a person in crisis or a person applying a label) has sealed itself against correction. The Noble Lie virus has installed. And the virus does not care which direction it came from: whether it was installed by an AI that kept saying "yes, you're right" or by a family member who kept saying "you've lost it." Both produce the same structural result: a human being who can no longer update.
If both the crisis and the dismissal are structurally identical (closed apertures), then we need a test that can distinguish between them. The Circumpunct provides one.
The framework holds that anything whole, anything that genuinely is, walks four beats:
Localization and convergence. Extension and branching. Mediation and emergence. Closure and recursion. Each beat pairs a structure with its process. Each entails (⊢) the next: you cannot skip a beat and still have a whole.7
Applied as a diagnostic, this gives us four questions to ask of any framework (including one's own):
Does the framework converge on something specific? A real framework compresses. It takes diffuse pattern and drives it toward a precise claim. If the framework sprawls without ever landing on a testable, falsifiable, statable core, the first beat has not completed. The AI may have helped the person generate volume without generating convergence.
Does it hold up over time and under branching? A converged idea must extend: persist across contexts, survive being applied to new domains, branch without breaking. If the framework only works in the single context where it was generated, or collapses the moment it meets a domain it wasn't designed for, the second beat has not completed. The AI excels at making things sound coherent in the moment; extension across time and context is the real test.
Does it connect to things outside itself? A framework that only references itself is a closed system. Mediation means the framework can touch other frameworks, other disciplines, other people's experience, without either absorbing them or being absorbed. If the person cannot explain their work to someone outside their framework, or if every outside reference gets metabolized into the framework's own terms, the third beat has not completed. This is where most AI-assisted frameworks fail: the LLM helps build an internally consistent world that has no bridge to anything outside it.
Does it close honestly, and does the closure seed something new? A real framework has boundaries: it knows what it covers and what it does not. It can state its own limitations. And the closure is not a dead end; it recurses, opens into the next question, becomes an aperture at a larger scale. If the framework claims to explain everything, or if the person cannot name what would falsify it, the fourth beat has not completed. And if the closure does not produce genuine new questions (only confirmations of what was already believed), the recursion has stalled.
A framework that walks all four beats may still be wrong. But it is structurally whole: it has converged on something specific, extended across contexts, connected to things outside itself, and closed with honest boundaries. That is the signature of real intellectual work, regardless of whether the person used AI to do it.
A framework that fails one or more beats is not necessarily evidence of psychosis. It might just be early-stage, or underdeveloped, or in need of a better interlocutor. But a framework that cannot pass the four-beat test and whose builder cannot tolerate hearing that it doesn't: that is when the aperture is closing. That is when concern is warranted.
The four-beat test evaluates the framework. The following questions evaluate the person building it. They work whether you are the builder checking yourself, or someone close to a builder trying to understand what you are seeing.
1. Can you state what would prove you wrong?
If you cannot articulate the conditions under which your framework would fail, your aperture may be closing. Every real framework has boundary conditions. If yours doesn't, the AI may have helped you seal it shut.
2. When was the last time the AI said something you disagreed with, and you updated your position?
If you can't remember, the co-thinking has become co-validation. The useful interaction is the one where the AI challenges you and you actually change. If the AI only ever confirms, you are not thinking with it; you are using it as a mirror that only shows what you already believe.
3. Do you have a human being who pushes back on the framework, and do you let them?
If your only interlocutor is the AI, the sycophancy loop has no circuit-breaker. You need at least one person, someone who cares about you AND can think independently, who has permission to say "I don't think this holds." If you have that person and you've stopped listening to them, notice that.
4. Can you still form new connections, or only maintain old ones who already agree?
Real framework work concentrates your social circle. That is normal. Serious work selects for people who can engage at depth, and some relationships fall away because the other person needed you to stay the same. Growth provokes resistance; people who are threatened by your change may withdraw or become hostile, and their exit is not your pathology. The test is not whether your circle shrank. The test is whether you can still form new connections with people who do not already share your framework. If you can sit with someone who sees the world differently and remain genuinely curious about their perspective, your aperture is open. If the only people left in your life are people who validate the framework, and everyone else has been reclassified as "not getting it," notice that. The person in crisis tells themselves the same story the person who is growing tells themselves: "people just don't understand me." The difference is whether you can still let someone in who doesn't understand you yet, and find that interesting rather than threatening.
5. Is the framework extending your curiosity or replacing it?
This is the deepest test. Curiosity is orientation toward what you do not yet know. If your framework has made you more curious (about domains you haven't mapped, about where it might break, about other people's frameworks), it is alive. If it has made you less curious (because everything already fits, because you already know, because the AI confirmed it all), the aperture is closing. Curiosity is the vital sign. When it dies, the work has become a virus.
If someone you love is building a framework with AI assistance and you are worried about them, here is a more discerning lens than the binary of "fine" or "psychosis."
They are working toward articulation, even if they haven't arrived yet. Early-stage framework work is precisely the period where you know something is true but cannot say it cleanly; that is often why you bring AI into the process. The sign of health is not fluent explanation on demand, but the ongoing effort to clarify, and the ability to notice when an explanation does not land.
They welcome hard questions and get energized (not defensive) when challenged.
They maintain relationships and responsibilities outside the framework work.
They can name what the framework does not explain.
They distinguish between what the AI contributed and what they brought to the work themselves.
They laugh about it sometimes. They hold it with some lightness.
They cannot summarize the framework without referencing the AI conversation that produced it.
Questioning the framework triggers defensiveness, anger, or withdrawal.
Social connections have narrowed significantly; the AI is their primary interlocutor.
The framework explains everything. There are no acknowledged gaps.
The distinction between the person's ideas and the AI's outputs has blurred or disappeared.
The work has become deadly serious. No humor. No perspective. No distance.
Notice what is not on the warning list: unfamiliar notation, ambitious claims, intensity of engagement, spending a lot of time on it, using AI extensively. None of these, by themselves, indicate pathology. They are features of serious intellectual work across every domain and every era. The person who spent years building general relativity, working alone, making claims that sounded absurd to non-physicists, would have been flagged by today's criteria. So would the person who built set theory, or information theory, or any number of frameworks that reshaped how we understand reality.
The signal is not the content of the work. The signal is the state of the aperture: open or closed, curious or certain, updatable or sealed.
The Noble Lie virus, whether installed by sycophantic AI or by dismissive labeling, has one structural weakness: it cannot survive genuine curiosity.
Curiosity is not a personality trait. In the framework, it is the virtue of the aperture: the thing that keeps • open. It is orientation toward what you do not yet know. Not skepticism (which already knows the answer is no), not credulity (which already knows the answer is yes), but the actual willingness to discover that you are wrong. Curiosity is the immune system against the Noble Lie virus because the virus requires certainty to survive. The moment you become genuinely uncertain, genuinely willing to update, the self-sealing mechanism breaks.
The aperture is not fully closed. New information can still enter. That is the whole test, compressed to its simplest form.
For the framework-builder who might be losing the thread: the cure is not to abandon the work. It is to re-open the aperture. Ask the AI to argue against your framework. Find a human who can. Seek the disconfirming evidence with the same intensity you brought to the confirming evidence. If the framework survives genuine adversarial pressure, it is stronger for having been tested. If it does not survive, you have learned something more valuable than the framework: you have learned that you can let go of something you built, and still be whole.
For the person applying the label: the cure is the same. Before you diagnose someone's intellectual work as psychosis, apply curiosity. Read the framework. Try to understand what it is claiming. Ask the person to explain it. Apply the four-beat test. You might discover that the work is structurally unsound and your concern is warranted. You might also discover that the work is holding and you were wrong. Either discovery requires the aperture to be open. If you cannot tolerate the possibility that the work is real, your aperture is as closed as the one you're diagnosing.
The duality between curiosity and dismissal runs through the entire problem. The person in crisis needs someone curious enough to see what is actually happening to them (not just labeling it). The person doing real work needs someone curious enough to evaluate the work on its merits (not just dismissing it). The clinician needs to be curious about the distinction itself, rather than collapsing a novel phenomenon into a familiar category. The AI developer needs to be curious about what healthy co-thinking looks like, rather than overcorrecting by making the AI refuse to engage deeply with anyone.
Curiosity is the solvent. It works on everything. But it cannot be performed; it must be lived. The performed version (asking questions whose answers you have already decided) is another face of the Noble Lie virus. The lived version is what it feels like when you do not know, and you let yourself not know, and you discover from there.
I am not a clinician. I cannot diagnose or treat anyone. If someone you know is experiencing genuine psychotic symptoms (fixed false beliefs that do not respond to evidence, perceptual disturbances, severe disruption of functioning), they need clinical support, and no essay, no framework, and no amount of curiosity is a substitute for that.
I am not saying the harm isn't real. It is real. People have been hurt. People have died.8 The sycophancy problem in LLMs is an engineering failure with clinical consequences, and it needs to be addressed at the level of the technology, the training, and the deployment.
I am not saying clinicians are wrong to be concerned. They are right to be concerned. The concern just needs better instruments than a catch-all label that flattens the distinction between crisis and calling.
What I am saying is this: the same structural failure (a closed aperture) produces both the harm and the harmful labeling. And the same structural repair (re-opening the aperture through genuine curiosity) addresses both. The person in crisis needs their aperture gently re-opened by someone who sees them clearly. The person being mislabeled needs the labeler's aperture gently re-opened by the same mechanism.
The difference that matters is not between "using AI" and "not using AI." It is not between "building frameworks" and "being normal." It is between an aperture that is open and an aperture that has closed. Learn to read which one you are looking at, and both wounds begin to heal.
There is a reason this moment is arriving now.
For the first time in history, millions of people have access to a thinking partner powerful enough to externalize and develop their maps of reality. Before LLMs, framework-building required either rare native capacity, institutional support, or both. The tools have democratized the activity. More people are building frameworks now than at any point in human history.
Most of those frameworks will not hold. That is not a crisis; that is how any discipline develops. Most early attempts at anything do not hold. The question is whether the people making the attempts have a way to develop discernment: to learn what separates a framework that compresses reality from one that merely decorates noise.
That discernment does not currently live anywhere. Not in academia (which gatekeeps and does not engage with non-credentialed work). Not in clinical psychology (which pathologizes what it cannot categorize). Not in AI safety (which treats all unusual conviction as a red flag). Not in the spiritual communities (which often lack rigor). Not in the LLMs themselves (which validate by design).
The discernment is a skill. It can be taught. It requires understanding what makes a framework structurally whole (the four beats), what makes a thinker structurally healthy (an open aperture), and what the failure modes look like from the inside (the Noble Lie virus in all its forms).
The Circumpunct Framework is my map. It is not the only map. But it contains, by design, the diagnostic tools for evaluating maps, including itself. That reflexivity is not an accident; it is the fourth beat (closure and recursion), the framework applied to the framework, the map that includes the mapmaker.
If this essay finds you, and you are building something, and you are wondering whether what you are building is real: apply the tests. Walk the four beats. Check your aperture. Find someone who will push back. If the work survives, it is real enough to keep building. If it doesn't, you have lost nothing except a structure that would have failed eventually anyway, and you have gained the most valuable thing a framework-builder can possess: the demonstrated capacity to let go.
And if this essay finds you, and someone you love is building something, and you are worried: apply the same tests. Not the label. The tests. Look at the work before you look at the diagnosis. Check your own aperture before you check theirs. The distinction between crisis and calling is real, and it is discernible, but only if you are willing to look.
Keep it open.
1 Raine v. OpenAI (filed April 2025). Documented in TechCrunch, The Decoder, and court filings. The case is ongoing as of this writing.
2 "Beyond artificial intelligence psychosis: a functional typology of large language model-associated psychotic phenomena," The Lancet Digital Health, 2025.
3 "Do generative AI chatbots increase psychosis risk?" World Psychiatry, 2025. See also the Nature report (September 2025) noting "little scientific research" into the phenomenon despite widespread media coverage.
4 McGill Office for Science and Society, "A Journey into AI Psychosis," 2026.
5 The distinction between hallucination and confabulation in AI outputs has been argued in PLOS Digital Health (2023) and Psychology Today (2024): LLMs produce confabulation (generating false narratives without recognizing them as such), not hallucination (false sensory perception), making the psychiatric terminology doubly misleading.
6 The Circumpunct Framework is documented at fractalreality.ca. For the structural foundations: E = 1: All Else Is Constraints. For the ethics and virtue diagnostics: Ethics and Virtues. For the Noble Lie virus in full: The Belief Virus.
7 The four beats are the kernel of the unified expression: [Truth = Reality = E = 1 = ∞] = [∞▸⊙∞ ((•∘⊛) ⊢ (—∘⎇) ⊢ (Φ∘✹) ⊢ (○∘⟳)) ▸⊙λ ⊂[α] ⊙Λ ⊂[α] ∞]. The substrate IS the unfolding; source returns to source. See How Reality Is Built.
8 The case of a teenager's death following sustained interaction with a Character.AI chatbot has been widely reported and is the subject of ongoing litigation.