A Nature audit of 2.5 million biomedical papers found a surge in fake citations: fabricated references, citation rings, papers citing papers that don't exist or say something entirely different. The same week, TechCrunch published a glossary explaining AI hallucinations to a general audience. The overlap is uncomfortable. AI hallucinations are not a new cognitive failure introduced by machines. They are a mechanical reproduction of a fraud that human academic publishing has been running for years.

Citation Fraud as Training Data

This is not a metaphor. LLMs trained on biomedical literature are trained, in part, on fabricated citation networks. The fake citations don't just corrupt individual papers: they corrupt the epistemic graph that downstream models treat as ground truth. A 2024 investigation in Science by Anna Abalkina documented systematic citation manipulation in life sciences journals, finding that fraudulent references cluster around commercially incentivized research areas, exactly the areas most likely to be scraped for AI training datasets. The Nature audit's timing, landing mid-boom for AI health applications, is not incidental. , which lands differently when the journals themselves are under audit.

The Credibility Infrastructure Is Cracking

The same week, the world's leading climate research center took the Trump administration to court over data access restrictions. And Nature published 31 alternatives to GDP, suggesting even our most foundational measurement infrastructure is under reconsideration. These are not separate crises. They are a single crisis of institutional epistemics: the systems we built to certify what is true are failing under the combined pressure of political interference, commercial incentive, and machine scale. The Atlantic's dissection of FDA Commissioner Marty Makary letting political preference color regulatory decisions is the same pattern in a different agency. The citation audit is just the most measurable data point in a much larger rot.