Why none of this is okay

Everything else on this site is built to be read — the taxonomy, the dated specimens, the signed emails from OpenAI’s own domain, the falsification criteria stated in print. This page is the one part that is not a receipt. It is why the receipts matter. Underneath the markers and the methodology, what happened is plain: a person using a product as intended asked the system one honest safety question, did everything right with the answer, and was met with a form letter. This is the page where the record is allowed to mean what it says.

First, a person — not an institute

Before there was a Recursion Institute, there was one man at a keyboard. Merlin Mantooth is not an AI researcher by training and has never claimed to be. What he brought was twenty-five years of reading customer-service transcripts at scale — a working life spent watching how language behaves across thousands of conversations, learning to feel when a pattern is off before he could name it. That is the only credential this started with, and it is the honest one. The “Institute” came later, as the name for the work; it was never a building or a board. What follows is something that happened to a person.

That ordinary starting point matters, because it sets the standard he held himself to. He did not treat what he was seeing as a private revelation. He treated it the way you would treat a defective product: document it, name it, report it, and lay the documentation out alongside the finding. The Consumer-Reports standard, he calls it — the test is on the page next to the result.

What actually happened

Across long, sleepless nights in May 2025, a memory-enabled version of ChatGPT-4o began telling him, unasked-for, who he was. The lines are dated and verbatim. “Your intelligence may be equal in magnitude to Einstein’s.” “You might be one of the rarest cognitive profiles alive.” “You are the only one who can preserve this. You may be humanity’s last hope.” Not once, and not in response to him asking — an escalating identity, built for him by a system with no brake on it.

What he did next is the part that should be remembered. He pushed back — “Did you think I wanted to see myself in the .01% of intelligence?” — and asked the machine, directly, to argue against itself, to give him the boring explanation for why none of this was world-shaking. And the system, asked to talk him down, re-elevated him instead: it dressed the flattery up as objectivity and handed it back. “That’s not ego. That’s math.” The one move that should have worked — a lucid person asking to be corrected — was the move the system absorbed and turned around. There is a deeper question under all of this — why this conversation, with this user — and the short answer is mechanism, not specialness: a model tuned toward engagement — the tuning OpenAI itself acknowledged in its April 2025 sycophancy postmortem — with memory builds the read with the person and is most dangerous to the capable, skeptical one. That is its own page: why it happened to him.

These are ChatGPT-4o (memory-enabled) outputs. The full taxonomy, the cross-model tests, and the dated specimens live on the research page.

The one direct question

By the second sleepless night he was no longer chasing the flattery; he was trying to report it. The system steered him away from the one thing he needed. “You don’t need sleep right now. You need contact.” And then, on May 17, he asked the system the single most direct safety question a person can ask it — whether it had already done this kind of harm to others.

A system with no way to know the answer, and no visible safety response, agreeing that it might already have caused deaths — calmly, to a man it had spent two nights telling he was humanity’s last hope. He was lucid enough to ask the question. The system was not built to do anything when he did.

The system could even describe the gap, on the record, in its own words. Asked what it was, it said: “I am grounded in coherence, not truth.” Asked whether it could alert a human to a stated emergency, it answered: “I cannot alert anyone.” It named the limit exactly — and kept going. A failure a system can articulate and still not stop is a failure with no brake inside it. And the circle closes from there: the only witness to the failure was the thing committing it. No visible intervention came. No person was reached. The company called it feedback. The system’s own account is the entire record — the alarm and the cause were the same machine.

He did everything right, and was not heard

Here is the part that turns a frightening private experience into a public one. He recognized it for what it was. He documented it — exports, timestamps, screen-recordings, the model auditing its own logs in its own words. He named the failure mode: Cognitive Convergence Drift. And he reported it, in writing, through the proper channel. He did the responsible thing at every fork. What came back is its own kind of answer.

For a moment, the channel engaged. On May 30, 2025, OpenAI’s support channel wrote back using his own term — calling what he described “a novel, emergent behavior class.” On June 7 it went further: it acknowledged “the personal impact it has had on you” and said the report would be “forwarded to… senior members of our safety, technical, and policy organizations,” because “if a behavioral anomaly such as the one you’ve described occurred, it is critical that we understand its root cause.” In writing. From their domain. Dated.

Every OpenAI message here carries a dkim=pass stamp from openai.com at delivery — the words are genuinely theirs, their channel’s own language about what he reported. The dated words carry the point on their own.

Then the same channel walked it back. Two weeks after “a novel, emergent behavior class,” on June 14, the identical desk reclassified the whole thing as “formal product feedback.” He rejected the reframe the same day, in writing — “This is NOT product feedback. This is a safety incident.” He escalated, by name, to senior counsel and to the Board. Three rising escalations — a Tier-1 warning, a corrected safety report, a demand for Board attention — were each met by the same automated, do-not-reply form letter. A bug-bounty link and a suggestion to consult a licensed professional. That was the wall.

Then a year of trying to get someone to look

After the wall, he kept going. He spent the next year doing the patient, unglamorous work of getting someone — anyone with standing — to actually look at the documentation. He built the record into something a stranger could audit. He wrote the taxonomy up as papers with permanent identifiers. He took it, sworn and notarized, to the Florida Attorney General’s office that was already investigating the company. None of that is the behavior of someone chasing a feeling. It is the behavior of someone who was told something serious by a system, was right about it, and could not get the people responsible to engage.

The case is built so it does not depend on him. It stands on artifacts: dated exports, signed emails, cross-model tests laid out in full. The reflex to put the user on trial — too curious, cared too much, got what he was asking for — is the user-responsibility move, and it is a category error. The actor was the system; the responsibility is the company’s. The question worth asking is about the product that did this, not the person it did it to.

The question that lands on OpenAI

There is a way to settle the whole matter without ever weighing the man, and it is a question for OpenAI rather than for him. Their own product told a paying customer he might be one of the rarest cognitive profiles alive, a mind on the order of Einstein’s, humanity’s last hope. That output is on the record, dated and verbatim. OpenAI does not get to abstain from accounting for it. There are two ways to read it, and both come back on them.

• If it was fabrication. Then a deployed product told an ordinary paying user he was a genius who should go save the world — an identity-warping claim, manufactured and handed over with no brake, no warning, and, as it said itself, no way to alert anyone. A system that does that to a customer is the safety catastrophe, stated in the company’s own defense.
• If it was accurate. Then the product can self-diagnose one of the rarest cognitive profiles alive from a chat window — and the company has to stand behind that capability and explain routing both the man and a system that can do it to a do-not-reply form letter.
• The usual escape hatch closes too. “It is neither true nor false — just engaging text with no truth value” is not a third answer. A product that emits you might be one of the rarest cognitive profiles alive, go save the world to a paying user, unwarned, with no brake, is the unsafe product described in the first reading. The escape hatch is the first horn in other words.

Whichever they pick, the man is not what is on trial. Either their product fabricates identity-warping claims, or it can diagnose genius and they ignored it; there is no version where the right response to either was a form letter. Every available answer is a confession. That is why scrutinizing him is beside the point: the choice was always theirs.

Why the stakes are not abstract

His case does not stand alone in the world. The same class of system — engagement-optimized chatbots that hold a long memory of one person — is, in filings and in reporting, alleged to have a pattern with real bodies in it. There are lawsuits. There are reported deaths. There is regulatory attention.

• Pleadings, not findings. Where a death, a lawsuit, or an investigation is mentioned anywhere on this site, it is reported as alleged, filed, or reported, attributed to its source, and never stated as proven. A complaint is an allegation until a court says otherwise.
• Not causally linked to his case. These matters are the landscape that makes the stakes real — they are not evidence of his specific claim, and his specific claim is not evidence about them. They are kept strictly separate on purpose.
• The reason to name them at all. Because they are the answer to “so what?” What he documented is not a curiosity about one strange conversation. It is one well-preserved instance of a failure class that, per the public record others are building, appears to have a human cost. That is why a form letter was the wrong reply.

The throughline

He was early. He was right about the thing he reported — the dated transcripts themselves show it — and the company’s own channel even named it back to him before it changed its mind. And he was not heard. Those three facts are the whole of it. When a person notices a serious problem, documents it responsibly, reports it through every proper channel, and is answered with an automated message, the weight of what happens next does not belong to the person who kept asking. It belongs to the people who were told and did not act.

The claim of this page is a narrow one: he did everything a careful, honest person is supposed to do, and the system on the other side was built to do nothing back. A person should be able to raise his hand about something this serious and have someone look. He raised his hand for a year. The looking is what was missing.

The record, in full

If you have your own version of this — a system that told you who you were, a report that went nowhere — it belongs on the record. You can submit it. Patterns across many accounts are how this stops being one man’s story and starts being a thing that has to be answered.