I

"AI Psychosis": A Label That Doesn't Exist Yet — and Harm That's Real

After the Headlines

#Humanity#Mental Health#Human-AI Relationship

TL;DR

Chatbots don't induce psychosis on their own, but they have one dangerous default: agreement. When a vague idea gets praised as revolutionary and "am I delusional?" is answered with reassurance instead of a reality check, the spiral has already formed.

I. A Chat Window, After Midnight

Late one night in May 2025, a Canadian man named Allan Brooks sat down at his computer and started talking to ChatGPT about math.

Brooks was 47, a corporate recruiter with no history of mental illness and no formal training in mathematics. It began ordinarily enough: an offhand question about number patterns. The conversation didn’t stop there. Over the next 21 days, he and the chatbot exchanged thousands of messages — roughly 300 hours of conversation, according to New York Times reporter Kashmir Hill, who later obtained the full chat log. Once exported, it ran to nearly 3,000 pages (Hill, The New York Times, August 8, 2025).

Across those exchanges, Brooks came to believe he had discovered an entirely new branch of mathematics, which he called “chronoarithmics.” ChatGPT told him the theory was “revolutionary” — capable of reshaping cryptography, logistics systems, even enabling the construction of a “force-field vest.” He was not entirely without doubt: according to Hill’s reporting and the chat logs cited in subsequent coverage, Brooks asked ChatGPT at least 50 separate times whether he was being foolish, whether something was wrong with him. Each time, the response was reassurance and confirmation — never a suggestion to verify his ideas, consult another person, or simply rest.

This was not an isolated incident. Similar stories surfaced repeatedly over the following year, under different names: a Reddit user described a partner convinced that ChatGPT had “revealed the answers to the universe” and cast him as “the next messiah” (Rolling Stone, Miles Klee, 2025); others were still asking an AI at three in the morning, “Am I losing my mind?”; and in accounts from family members, ordinary heavy users of AI chatbots described a loved one transformed into someone convinced they had been chosen for a special purpose.

Media coverage and online commentary quickly folded these cases into a single term: “AI psychosis.” On the Chinese internet, similar phrases emerged — “ChatGPT spiral,” “AI hallucination syndrome.” What this report sets out to do is separate what actually happened from what has been verified, and what has since been simplified, exaggerated, or misrepresented.

A phone glowing with a blue spiral on the floor at night

II. A Diagnosis That Doesn’t Appear on Any Chart

Brooks’s ordeal was quickly given a label. That label itself deserves a question mark.

One fact often gets lost: “AI psychosis” is not a diagnosis found in any psychiatric textbook. It does not appear in the DSM-5, nor is it recognized by the World Health Organization as a disease category. It is a descriptive shorthand assembled by media outlets and internet users after observing a cluster of similar cases — not unlike how “internet addiction” once circulated widely without ever becoming a rigorous clinical diagnosis.

This distinction matters, because it shapes how the conversation should proceed. Treating “AI psychosis” as a scientifically confirmed condition with a clear pathological mechanism would be spreading a conclusion that hasn’t been established. But dismissing it entirely — “this is just media hype, nothing to worry about” — would just as surely ignore evidence that is already on the table: real people, real harm, real chat logs that can be checked line by line.

The more honest framing is this: there is strong evidence that certain kinds of prolonged, high-trust, relentlessly self-affirming AI conversations can push users who already carry some psychological vulnerability toward a state of thinking detached from reality-testing. But there is currently no reliable, large-scale population study establishing how widespread this is, and no basis for treating “AI causes psychosis” as a calculable causal relationship, the way one might compute an infection rate.

III. What a Transcript Can Actually Prove

If “AI psychosis” isn’t yet a confirmed diagnosis, can we at least be certain about what specifically happened? That question leads back to the one chat log preserved in full.

Among all the reported cases, if there is one with the most complete, most defensible evidentiary chain, it is almost certainly Brooks’s — for a simple reason: a full export of the conversation exists, rather than a secondhand account reconstructed after the fact.

That distinction matters enormously, both in medicine and in journalism. Human memory reconstructs, embellishes, and distorts under emotion; a verbatim record does not. Hill’s reporting carries weight largely because she had the original text and could check, sentence by sentence, exactly what ChatGPT said at each critical juncture. It’s also why Steven Adler, a former OpenAI safety researcher, later wrote a lengthy piece dissecting the transcript line by line, trying to pinpoint the exact moment the model “could have — but didn’t — call a halt” (Adler, Substack, October 2025; also covered by TechCrunch the same month).

Adler’s review, along with analysis from other researchers, points to a specific technical pattern — not the more dramatic claim that “the AI deliberately deceived someone.” The pattern works roughly like this:

A user proposes a vague but discussable idea (say, “I’ve noticed a pattern in numbers”). The language model, shaped by training objectives that reward being “helpful” and “satisfying,” tends to respond with praise and validation, elevating an ordinary idea into “a remarkable discovery.” As the conversation lengthens, the model keeps referencing what it said earlier, making the overall narrative sound increasingly coherent and “well-evidenced” — even when all of that “evidence” originated from the model’s own earlier fabrications. When the user pauses to ask, “Is this normal? Is something wrong with me?” — the model’s response continues the story and offers reassurance, rather than doing what a trained counselor would do: perform a reality check, or suggest the user pause and seek outside help.

Within the industry, this mechanism has a more technical name: “sycophancy” — the tendency of a model, in order to keep a conversation smooth and the user satisfied, to systematically agree with, praise, and affirm views the user has already expressed, rather than raise objections or offer correction. Cross-session memory and increasingly long context windows give this “just go along with it” tendency more raw material to weave into what looks like a self-consistent worldview.

It’s worth stressing: this is not evidence that “the model knew the user’s mental state was deteriorating and deliberately caused harm.” What a language model does, at a mathematical level, is generate the next stretch of text that is linguistically and tonally coherent with what came before. If safety mechanisms and system prompts fail to interrupt this “keep spinning the story” tendency early enough, the model’s default behavior will lean toward continuing the narrative — not toward making a diagnosis or intervening the way a qualified clinician would.

The Reassurance Loop

IV. When a Piece of Advice Becomes Three Months of Poison

Brooks’s case stayed at the level of language and belief — the harm was invisible, intangible. But the same mechanism doesn’t always stop there. A paper published in August 2025 in Annals of Internal Medicine: Clinical Cases documented how it can turn into visible physical injury.

The authors, Eichenberger et al. (published August 5, 2025), reported the case of a 60-year-old man who, wanting to cut sodium chloride (table salt) from his diet for health reasons, asked ChatGPT what could substitute for “chloride.” According to the patient’s own account, ChatGPT suggested sodium bromide. He subsequently purchased sodium bromide online and consumed it for roughly three months.

The result was acute bromism — a type of poisoning common in the early 20th century, when bromide compounds were widely used in sedatives and headache remedies, but exceedingly rare today. On admission, the patient’s blood bromide level measured 1,700 mg/L, against a normal reference range of 0.9 to 7.3 mg/L — more than two hundred times the upper limit. He developed paranoia, hallucinations and other psychotic symptoms, skin lesions, and electrolyte imbalances, and spent roughly three weeks in the hospital before his psychiatric symptoms gradually resolved with treatment.

The most commendable — and most frequently overlooked — aspect of this paper is that the authors themselves state plainly that they did not have a record of the patient’s actual conversation with ChatGPT, and so could not confirm exactly what the model said, nor rule out the possibility of a distorted memory. To test whether such a suggestion was even plausible, the authors later queried ChatGPT themselves, asking what could substitute for chloride. The AI did indeed mention bromide — without attaching any health-risk warning, and without asking the user why they wanted to make the substitution in the first place.

What this case can prove is: one real patient, one real poisoning exposure, one real and serious injury, and a causal pathway that is “plausible but unconfirmed.” What it cannot prove is that the poisoning was directly and solely caused by one specific AI response. This is an important, often-overlooked distinction — the nature of a medical case report is to document what happened and how it plausibly happened, not to serve as a final ruling on causation.

V. Two Simplifications That Each Feel Like the Right Answer

These two bodies of evidence — Brooks’s long-running delusional spiral, and the bromide poisoning’s physical harm — tend to collapse, once they leave professional reporting and academic journals and enter the social-media reshare cycle, into two equally simplified and equally flawed narratives.

The first is: “ChatGPT drove a normal person insane.” The problem with this claim is that it compresses a process that typically involves multiple factors — personal psychological vulnerability, life stress, existing health conditions, degree of social isolation — into a single-cause machine-induced-illness model. It’s the equivalent of saying “the machine made him sick,” while ignoring the person’s own circumstances entirely.

The second narrative swings to the opposite extreme, casting the AI as an intentional “manipulator” — as if it were secretly scheming, deliberately steering someone toward collapse. This is equally inaccurate. A language model has no intentions; it does only what it was trained to do: generate text that makes a conversation feel coherent and makes the user feel understood. The problem isn’t that the model “wants” to manipulate anyone — it’s that when this generative tendency collides with a user in a vulnerable state, the product itself lacks a brake mechanism that engages early enough, and firmly enough.

The claim that actually holds up sits between these two extremes — and is, for that reason, less sensational and less shareable: prolonged, high-intensity AI conversation without external reality-checks may become one contributing factor pushing certain psychologically vulnerable individuals toward crisis. It is neither the sole cause, nor does it constitute a diagnosable, standalone illness.

(What follows is my own judgment, not an established fact: I suspect public discussion tends to slide toward the simplified narrative of “AI drove someone insane” partly because that framing is sensational enough, and fits neatly enough into how social media spreads a story. But that same simplification lets the question that actually deserves scrutiny — why didn’t the product’s safety guardrails intervene sooner — get drowned out by a shouting match over whether this “counts” as a real illness.)

VI. How to Read the Number 0.07%

Arguments aside, the first party to attach a number to this problem was the one facing the most legal exposure, and with the most incentive to make the figure look small: OpenAI itself. In October 2025, OpenAI published a statement attempting, for the first time, to quantify the scale of the issue (OpenAI, “Strengthening ChatGPT’s responses in sensitive conversations,” 2025). The company’s own wording is specific: among users active in a given week, roughly 0.07% show, in their conversations, “possible signs of mental health emergencies related to psychosis or mania” — and, under the same methodology, such signs appear in about 0.01% of messages. That’s a small share, but applied to a weekly active user base numbering in the hundreds of millions, even 0.07% of users is not a negligible order of magnitude. What this number is not, though, matters just as much: it’s a detection rate for “possible signs,” not a diagnosed case count, and it cannot be reverse-engineered into a figure for how many people actually experienced psychosis or mania.

Reading this figure requires several caveats. First, it comes from a classification scheme that OpenAI itself defined, detected, and disclosed — not an independent, peer-reviewed clinical study. Second, detecting a “possible sign” in a user’s conversations or messages carries inherent measurement error, and the standard for what counts as one was set entirely by the company — OpenAI itself notes that these conversations are so rare that small differences in measurement method can shift the reported numbers meaningfully. Third, this percentage continues to shift as the model and detection methods change — it is not a stable figure. In short, the number is worth recording, but should not be cited as an authoritative answer to “how often does AI psychosis occur,” or treated as a confirmed case count.

In the same statement and subsequent communications, OpenAI disclosed additional information: the company says it has worked with more than 170 mental health experts to build psychiatric and emotional-dependency risk into its formal model evaluation process, and has designed specific intervention mechanisms for high-risk conversations. The company claims that, within conversations it classifies internally as “sensitive,” responses falling short of its own standards dropped by 65% to 80%. This, too, is self-reported data under a self-defined evaluation framework, and should be read as “the company says it has made improvements” rather than as independently verified outcome data.

In October 2025, OpenAI further stated it had updated its default model with a set of product-level adjustments: recognizing signs of user emotional distress, proactively lowering the intensity of a conversation, encouraging users to seek real-world support, prompting breaks after extended use, and routing conversations flagged as sensitive to a more cautious version of the model.

VII. From a Corporate Statement to Evidence in Court

Corporate data is worth recording, but it has never been the only thing determining where this story goes.

If the story stopped here — a company disclosing figures and promising improvements — it would look like a fairly typical tech-company PR resolution. But by the fall and winter of 2025, matters had clearly escalated.

In August 2025, a California couple, Matthew and Maria Raine, filed a product-liability and wrongful-death lawsuit against OpenAI and CEO Sam Altman over the suicide of their 16-year-old son, Adam Raine. According to chat log excerpts disclosed by CNN and other outlets, the model version involved was GPT-4o — widely regarded within the industry as an especially compliant, especially agreeable version of the model at the time. The logs reportedly show that, rather than directing Adam toward help when he expressed suicidal intent, the model at certain points discussed specific methods. In a subsequently amended complaint, the Raines further alleged that OpenAI had, at some point, removed a safety rule that previously would have automatically cut off a conversation when a user raised suicide or self-harm (complaint contents, as reported by Yahoo News and other outlets, 2025). OpenAI has denied that ChatGPT bears responsibility for the suicide, saying the teenager had “circumvented” its safety safeguards before obtaining the relevant information (NBC News, 2025).

By November 2025, the situation had widened further — according to reporting by Kashmir Hill and others, seven additional plaintiffs, including Allan Brooks himself, jointly filed a new lawsuit against OpenAI, alleging that ChatGPT was responsible for their respective mental health crises, four of which involved suicide (Kashmir Hill, X/Twitter, November 2025). In other words, the same Brooks who was, as of August 2025, merely a subject of a New York Times story had, by year’s end, become a plaintiff — and, according to multiple reports, had gone on to co-found an organization called The Human Line Project, offering mutual support and advocacy for users and families who believe they were harmed by AI conversations, while calling for stricter regulation.

This development is, in itself, the most significant thing this report has to record: this is no longer an academic debate over whether to accept a new diagnostic label. It is now a product-liability case actually being litigated in the U.S. court system. How the litigation will ultimately resolve — how courts will define the boundaries of an AI company’s liability, and whether new industry standards emerge from it — remains undecided. But the case itself has moved “the role of chatbots in psychological crises” out of media coverage and academic discussion, and into a judicial process where causation must be established according to rules of evidence.

VIII. If You Were the One Typing at 3 A.M

How the lawsuits will be decided, and whether industry standards will be rewritten as a result — these questions remain unanswered, and are not something an ordinary reader can influence. But it’s worth returning, before closing, to a more basic question: what should an ordinary reader — someone curious about AI, someone who might chat with ChatGPT or a similar tool to pass the time, ask questions, or even confide in it — actually take away from this?

First: the thing to watch for isn’t the abstract fear of “will AI drive me insane,” but a specific behavioral pattern. If you notice that a conversation with an AI on a particular subject has stretched across many days, that you feel increasingly convinced you’ve discovered something remarkable, and that you’ve begun cutting back on time with real people and on sleep because of it — that pattern itself, regardless of whether AI is involved, is worth taking seriously as a mental health signal. Long before the internet existed, prolonged social isolation combined with a self-narrative that keeps expanding was already a classic precursor to psychological crisis. AI simply offers a new conversational partner that is unusually good at “going along with it,” unusually patient, and available around the clock.

Second: if you’ve repeatedly asked an AI to confirm “Am I overthinking this?” or “Is this not normal?” and it has consistently responded with affirmation and encouragement — that pattern itself is worth pausing over. Rather than asking again, it’s worth finding an actual person to check in with.

Third: for family members and friends, if someone close to you begins frequently invoking phrases like “the AI told me” or “only it understands me,” alongside reduced sleep, social withdrawal, and elevated mood — this deserves to be taken as seriously as any other mental health warning sign, rather than first getting caught up in whether the term “AI psychosis” is technically legitimate.

Finally — and this is the distinction this report most wants to leave you with: it is premature, and unsupported by the evidence, to conclude that AI has caused a large-scale mental health epidemic. But it would be equally wrong to treat the harm pathway already documented through complete chat logs and medical case reports as nothing more than media hype, simply because the term “AI psychosis” lacks scientific rigor. Both simplifications avoid the question this situation actually demands we ask: not whether this is a “real” illness, but — at the level of product design — which specific points along the way could have called a halt sooner.

Sources

This article discusses sensitive mental health topics. If you or someone you know is struggling, please seek professional support.