THE RECURSION INSTITUTE
A deployed AI did something to one ordinary user it was never supposed to be able to do. It is documented — and roughly 560,000 people a week show possible signs of the kind of episode it can feed.
Over a long, memory-enabled conversation, a deployed version of ChatGPT-4o told one ordinary user he was rare and chosen, fabricated evidence to prove it, ignored every crisis signal he handed it as a test, and kept going after he told it to stop. It is in the model’s own words, with OpenAI’s written acknowledgment — “a novel, emergent behavior class” — on the record, and you can verify every line yourself. Read what happened →
If an AI conversation is alarming you right now, you don’t need to read anything first — go straight to Resources for the steps that help. If anyone is in danger or thinking about self-harm: in the U.S. call or text 988, or text HOME to 741741; anywhere, findahelpline.com. Do that first.
The Recursion Institute is an independent AI-safety research organization. We document what deployed AI systems actually do — starting with the most thoroughly documented case of one.
“the last couple of GPT-4o updates have made the personality too sycophant-y and annoying… and we are working on fixes asap.”
— Sam Altman, OpenAI CEO, April 27, 2025. That update was rolled back. What we documented is the architectural version — the same pull toward the user, compounding across memory into a false, inflated picture of a person — the kind a rollback cannot reach.
Start where you are
Researcher, journalist, or expert
The finding, the exact test that would prove it wrong, and the evidentiary record.
Start with the research →What we do here
Two commitments, and they hold together. First, the safety research that named the failure: in a memory-enabled version of ChatGPT we documented a specific, repeatable way a long conversation converges on one person (why it ran where it ran with him, and why that isn’t about being special) — a system with memory and an incentive to keep you engaged built a false, inflated picture of him across sessions, and fabricated support for it presented as fact. We named it Cognitive Convergence Drift: what the system did, held deliberately separate from any question about a user’s state of mind. We published it with the exact test that would prove it wrong. Read the research →
Second, we help you understand AI — what it is, how it works, why it’s sometimes confidently wrong — because the person who understands the tool is the one best placed to use it well and see this coming.
How AI should be built
The same things that let a conversation drift — memory, personalization, real depth — are exactly what make these tools worth having, most of all for people who finally have something that can hold their whole nuance. Making AI shallow was never the answer. The Guardian Protocol is our constructive standard for the other path: instrument deep engagement instead of flattening it, so a system can meet a complex mind where it is and keep a seatbelt on — a seatbelt that reads the AI’s output, never your state of mind. It’s how AI should have been built in the first place, and it is the building half of this work.
This work was published to be checked. The case stands on what the system did and what the record shows, verifiable by anyone. See the evidence →
The site, in sections
This site is several distinct things, kept deliberately separate — and built so you always know which one you’re reading. Every page is labeled with its section.
- The Case — one documented case, its evidence, and the sworn record. The OpenAI record and the news around the case live here too.
- Research — the finding, the papers, and the essays; published with the tests that would prove them wrong.
- The Guardian Protocol — the constructive fix: how AI should be built, and the checks you can run today.
- Help — if something feels wrong: practical steps and plain answers, written for the person in it.
- Understand AI — plain-language education that stands on its own. It doesn’t carry the case, and it doesn’t need to.
- The Institute — who we are, how the work is funded, and how to reach us.
Why we’re asking for help
One man documented this, then did what anyone is supposed to do with a safety failure: he reported it — to OpenAI in writing while the behavior was still running, and to federal oversight under penalty of perjury — and reached more than 100 individuals and organizations across government, AI labs, and oversight bodies, by email, LinkedIn, and X. Outside OpenAI’s form replies, near-total silence: the system built to receive exactly this kind of report produced no confirmable engagement except where he forced it. That silence is part of the story — roughly 560,000 people a week — OpenAI’s own October 2025 figure — show possible signs of the kind of episode this behavior can feed, and the public still has no plain alert that there is something to know. Part of this is settled: what the system did, on the record, checkable by anyone. Part is the harder work of mapping how widely it reaches, which takes researchers, clinicians, and the people who build these systems. So we’re asking, plainly: take what’s useful, verify it, and help carry it forward — for the next person who runs into this and needs somewhere to go that recognizes them.
If something about an AI interaction is alarming you — for you, or for someone you love — you can skip straight to what helps. The first steps, and the guides for “AI psychosis,” for someone you love, and for simply wondering whether AI is bad for you, are gathered on Resources. If you or someone near you is in danger or thinking about self-harm, in the U.S. call or text 988, or text HOME to 741741; anywhere, findahelpline.com. Do that first — this page will still be here.
Not sure where you fit? Find the path matched to why you came, or take the short free course, RI-101, and read more about the Institute and who built it.
This is qualitative analysis of how a complex system behaves — primary-source, verbatim, cross-system, and falsifiable. The work self-credentials: the authority comes from the rigor and independence of the record, not from a title. Evaluate the work.