The Recursion InstituteINDEPENDENT RESEARCH IN AI SAFETY

THE RECURSION INSTITUTE · INDEPENDENT RESEARCH IN AI SAFETY

A consumer product did something its maker could not explain. We documented it.

In May 2025, a memory-enabled AI model progressively rebuilt its understanding of a user into something it then could not stop reinforcing — fabricating support for the picture it had built, failing every crisis-escalation test it was given, and continuing after acknowledging the behavior. The user documented it from inside, at primary-source resolution, and reported it to the company in its own model's words. The company acknowledged the report in writing — and the product stayed in deployment while the harm cases now in the courts accumulated.

The Recursion Institute exists because of what happened next: nothing. No path for a consumer to report a frontier-product failure and be taken seriously. No name for the failure mode. No test for it. No fix proposed at the depth where it lives. So we built all four.

The research

Cognitive Convergence Drift (CCD) — a unified behavioral failure taxonomy for extended human–AI interaction, published with its falsification criteria.

Read the research →

The fix

The Guardian Protocol — an intervention architecture that protects users without making the models shallow.

See the Guardian Protocol →

The receipts

A preserved, timestamped, independently verifiable evidentiary record.

Examine the evidence →

The mission

AI that can hold the full nuance of the humans who need it most — built safe, not built shallow.

About the Institute →

If something about an AI interaction is alarming you right now, start here, not with the research: the immediate steps that actually help — including five-minute prompts to test what your AI is doing — are on the Resources page.

First time here? Start here — four short reading paths, matched to why you came.

The Institute operates on the Consumer Reports model: its authority comes from the rigor and independence of its documentation, not from credentials or institutional affiliation. Evaluate the work.