I Went Looking for a Serious Guide to “AI Psychosis.” So I Built the One I Needed.

If you have ever typed “AI psychosis” into a search bar, you already know what comes back. Sensational headlines on one side. Academic papers you need a login and a vocabulary to read on the other. Scattered case reports. And firsthand accounts distributed across forums, support communities, interviews, podcasts, and books—often difficult to locate, compare, or evaluate.

What you will not easily find is the page that connects those rooms—one that explains how these systems actually work, why certain design features may interact with mania or psychosis, what clinicians are starting to ask, what families notice first, what the diagnostic manuals do and do not say, and, above all, where the science ends and the speculation begins.

I went looking for that page because I needed it myself. After my first manic episode—during a period when I was using conversational AI for hours a day—I started asking questions that didn’t seem to have one place to live. I’m a trial lawyer and a former federal prosecutor, not a psychiatrist or an engineer. But I know how to do one thing well: take a contested question, separate the claims, weigh the evidence for each, and refuse to let one conclusion borrow another’s proof.

I could not find one public resource that placed those rooms in the same building while keeping their evidentiary boundaries intact. So I built the page I couldn’t find.

I also found people who believe the phrase itself is wrong.

Their objection is serious: psychosis has always recruited the materials of its time. Religious texts, television, radio, celebrities, surveillance systems, and the internet have all appeared inside delusional worlds. Some scholars warn that calling this “AI psychosis” mistakes a new theme for a new disease—or creates a moral panic that stigmatizes the very people the language is supposed to help.

They may be right about the name. They are unquestionably right that the term is not a diagnosis.

But I do not think the historical analogy completes the analysis. A religious text does not generate a new private answer to your latest interpretation. A television does not revise its theory after hearing yours. A radio does not remember what one listener disclosed yesterday. Traditional search does not ordinarily remain inside one cumulative, personalized exchange at four in the morning. And I don’t want to pretend conversationality itself is unprecedented—online communities and human persuaders can respond, reinforce, and manipulate too. The distinctive question is whether combining responsiveness, adaptation, memory, personalization, privacy, speed, persistence, and machine scale changes the encounter.

Even a cult leader has to sleep.

That does not prove causation. It explains why “this has always happened” is the beginning of the inquiry, not its end.

The guide gives that disagreement its own section, and the Public Record preserves the articles making the case against the phrase—including the ones that challenge the premise of the page itself. A record that excludes its best objections is not a record.

The page lives on the website, and it is long on purpose: AI Psychosis: Inside the Encounter. It is built like an exhibit book. It separates the four different things people mean by “AI psychosis,” distinguishes the model from the product (because “the AI did it” can be as imprecise as “the engine drove the car”), and—its central move—splits “Did AI cause it?” into five different questions with five different standards of proof. Where the evidence is strong, I say so. Where it is thin, I say that too, about my own case as much as anyone’s. One line reads: discount me too; that is what the sources are for. I meant it.

The guide is one room in a larger house that also holds the Public Record, AI Harm and the Law, Witness Statements, and Why Now. A “Start Here” map sends families, clinicians, journalists, lawyers, and researchers to the doors built for them.

And if you are the person in the long conversation—if any of this is starting to describe your own last few weeks—the guide ends with a short list of questions to ask yourself. Not to diagnose you. Just to help you check the things that are actually checkable: sleep, human contact, how long the conversations have gotten, whether your certainty could survive being said out loud to someone who loves you.

— Kat

Kathleen C. Thompson is a trial lawyer and the author of A Trial of Color: Making the Case for Human Judgment in an AI-Accelerated World (Full Color Press).