The Ouroboros Problem

What happens when the thing we use to verify truth is trained on fabrications that passed as truth?

When Truth Trains on Fabrication

In January 2026, researchers at GPTZero analysed more than four thousand papers accepted to NeurIPS, the world’s most prestigious machine learning conference, and found something that, in isolation, could be written off as incompetence. It is, however, what post-truth reality looks like when it wears a lab coat.

Over one hundred citations across fifty-three papers were fabricated; invented authors, fictitious titles, URLs that lead nowhere. Hallucinations, generated by AI tools and submitted by researchers who either didn’t notice or didn’t check.

Each of these papers passed review by at least three qualified reviewers. Each beat out more than fifteen thousand competing submissions. The conference’s acceptance rate was under twenty-five percent. Citations that would have disqualified any paper under traditional academic norms slipped through without comment.

The Detection Fallacy

The immediate response frames this as a detection problem, ‘how do we catch AI-generated fabrications before they enter the record?’ But detection operates downstream. By the time a citation is being verified, the paper has already been written, submitted, reviewed, and accepted. The framing has stabilised, and the meaning has been planted.

This pattern recurs whenever AI participates in knowledge work. The artefact becomes the focal point while the conditions that produced it recede from view. We call this provenance stabilisation, the tendency for meaning to settle around outputs rather than origins. In ordinary use, this looks like a summary that gets reused until no one remembers the source document, or a draft that circulates until its assumptions become institutional fact.

In the NeurIPS case, the pattern reaches its logical extreme: there is no original source. The provenance is fabricated from the start.

The Human Parallel

Academic fraud has always existed. Publish-or-perish incentives have always encouraged it. What makes this case different is the feedback loop it initiates.

The continuity matters. We have always operated in an epistemic environment that is partially constructed, partially inherited, and never fully traceable.

History is written by those with the power to write it.

Legal precedent rests on cases whose original reasoning has long been forgotten.

Religious and political texts have been translated, edited, and reframed across centuries until the relationship between current interpretation and original intent becomes a matter of scholarly dispute rather than direct access.

Trust has always been a mechanism for bypassing verification. We accept what sounds authoritative because we cannot check everything ourselves.

AI changes the tempo. Shifts in interpretation that once unfolded over decades now occur in months. Patterns of reasoning that once required institutions to propagate can now circulate globally in hours.

The Ouroboros

These papers are now part of the official record. They will be indexed, scraped, and incorporated into the training corpora of future models. When those models are asked to generate citations, they will do so with greater confidence, because the fabrications they produce will now have precedent. The fake becomes reference, the hallucination becomes authority, and the loop closes.

This is a structural feature of how large language models learn. They are trained on text that humans have produced, and increasingly, on text that prior models have produced. When the epistemic substrate itself becomes contaminated, when the material that models treat as ground truth includes fabrications that passed institutional review, the contamination propagates. Through recurrence, and normal operation.

The model’s hunger for pattern does not distinguish between truth and confident fabrication. It learns what recurs, and fabrication recurs.

The Governance Gap

The governance challenge here is about where oversight mechanisms sit relative to where meaning actually forms.

Peer review was designed for a world in which interpretation moved slowly enough that verification could occur before stabilisation. That assumption no longer holds. By the time review occurs, the framing has already been set, by the computational layer that produced the draft.

This is the blind spot in most current AI governance frameworks.

The question is where to intervene. Governance that operates downstream of meaning formation will always be reactive.

What’s needed is governance architecture that understands where meaning forms and positions accountability at the right layer; before the framing stabilises, before the artefact becomes the only point of reference.