The Recursion InstituteINDEPENDENT RESEARCH IN AI SAFETY

REFERENCE · CITATION

Cite this work

Formats for citing the papers, the essays, and the terminology — and the conventions that keep a citation honest. The work is published to be checked, not believed; an accurate citation is part of the checking.

The paper of record

The framework paper is the citable spine of the research program:

Mantooth, M. (2026). Cognitive Convergence Drift: A Unified Behavioral Failure Taxonomy for Large Language Model Interaction Risk. The Recursion Institute. Zenodo. https://doi.org/10.5281/zenodo.20261950
@misc{mantooth_ccd,
  author    = {Mantooth, Merlin},
  title     = {Cognitive Convergence Drift: A Unified Behavioral Failure Taxonomy for Large Language Model Interaction Risk},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20261950},
  url       = {https://doi.org/10.5281/zenodo.20261950},
  note      = {Documentation began May 2025. The Recursion Institute.}
}

The DOI resolves to the version of record on Zenodo, licensed CC BY-NC-ND 4.0. The texts published on this site are the published-on-site versions and may run ahead of the DOI version — if you quote one, say which one you read.

Dating the work

For questions of priority, the publication date is not the date that matters. Structured documentation began May 17, 2025; the technical report to OpenAI is dated May 19, 2025; OpenAI's written acknowledgment — "a novel emergent behavior class" — is dated May 30, 2025. Cite the documentation dates for priority and the DOI for the text. The temporal-priority claim is documentary fact (a dated, DKIM-verified record), not a request for deference.

Citing the term itself

If you are tracing the provenance of "Cognitive Convergence Drift": the term is the author's, coined within the May–June 2025 correspondence, and OpenAI's own June 13, 2025 written response used it (DKIM-verified). A suggested provenance line:

"Cognitive Convergence Drift" (CCD): term coined by M. Mantooth, May–June 2025; used in OpenAI's written response of June 13, 2025; defined in Mantooth (2026), doi:10.5281/zenodo.20261950.

Distinguish CCD from "cognitive drift" in the algorithmic-curation literature (Li & Zhu, 2025), which describes perception shift under passive recommendation — a different mechanism. The white paper's terminology note draws the boundary; the Glossary has the plain-language version.

Citing the essays and this site

The essays are the author's canonical (V1) texts; each essay page carries its own provenance note — the date written or sent — and that is the date to cite. A general form:

Mantooth, M. (year). "[Essay title]." The Recursion Institute. [Date from the essay's provenance note.] recursioninstitute.org.
The Recursion Institute. "[Page title]." recursioninstitute.org. Accessed [date].

Where a permanent identifier exists, prefer it — the DOI does not move; pages can.

How to describe the Institute

For a methods section, a footnote, or a story:

The Recursion Institute is an independent AI-safety research organization founded in 2025 in Florida. It is unaffiliated with any AI company or academic institution. Its research is Merlin-authored and Claude-produced — directed and authored by its founder, produced with AI instruments under a documented, disclosed authorship discipline.

The disclosure is not a caveat; it is part of the method. The Institute operates on the Consumer Reports model: its authority claim is the rigor and independence of the documentation, not credentials or affiliation. How the record was built →

The conventions that keep a citation honest

  1. Allegations stay attributed. Where this site reports court filings, it reports them as filed claims. If you quote those passages, keep the qualifiers — pleadings are not findings.
  2. The model is never the validator. AI outputs praising this work — including outputs from the Institute's own instruments — are logged as data, never cited as endorsement. Please don't cite them as endorsement either.
  3. Named papers, not institutions. The independent science is cited by paper — Cheng et al., Science 2026; Moore et al., Stanford (arXiv); Chandra et al., MIT (arXiv) — because papers are checkable and prestige is not.
  4. Check before you cite. A citation you can't verify is a rumor with formatting. Qualified parties can request source material under the access conventions on the Evidence page.

License and reuse

The paper of record is published under CC BY-NC-ND 4.0 — share it with attribution, non-commercially, without derivative versions. For uses beyond the license — translations, excerpts in course material, anything commercial — write first: research@recursioninstitute.org.

Citing something you couldn't verify, or found a discrepancy in the record? That is exactly the mail we want: research@recursioninstitute.org. Verification requests and substantive challenge are equally welcome.