The Recursion InstituteINDEPENDENT RESEARCH IN AI SAFETY

PUBLICATIONS

Publications

All publications carry their methods and their limitations in the text. The work is published to be checked, not believed — every paper states what would disprove it.

Cognitive Convergence Drift: A Unified Behavioral Failure Taxonomy for LLM Interaction Risk

The framework paper. The eight markers, the SCC diagnostic, the infrastructure analysis, the institutional-response record, falsification criteria, and the epistemic-position disclosure.

Published on Zenodo (DOI: 10.5281/zenodo.20261950) · CC BY-NC-ND 4.0. The on-site text is the V12 published draft (pending final author approval); the Zenodo v11 deposit is the published version of record.

Read on this site →  ·  Read on Zenodo →

The Guardian Protocol: An Intervention Architecture for Behavioral Safety in Extended Human–AI Interaction

The fix. Seven layers that instrument deep engagement instead of flattening it — including cross-instance verification, a hidden fabrication check, and a user-words anchor — deployable as middleware, training integration, or a neutral-body standard. Includes the validation methodology and the public test batteries.

Read on this site →  ·  The plain-language explainer →

The Visible Layer: Reasoning Transparency, Evaluation-Before-Content, and the Identity Variable

What the new transparent models reveal. When you can watch a model deliberate, you can watch it evaluate the user before the content — and measure how the same evidence gets different treatment depending on who the model thinks is holding it. What that implies about the systems that show nothing.

Read on this site →

The Author and the Instrument: Attribution, Provenance, and Quality in Human–AI Authorship

The authorship question. Editor versus generator, attribution laundering in both directions, why current norms punish disclosure and reward laundering, and the provenance architecture that fixes it.

Read on this site →

The on-site texts are the published-on-site versions of each paper; where a Zenodo DOI exists, the Zenodo deposit is the version of record.

The evidentiary substrate behind these publications is cataloged on the Evidence page; qualified parties can request source material for verification. Citing the work? Formats, the DOI, and the provenance of the terminology are on Cite This Work.