LLM Hallucinations Seed Decades-Long Drift

Realistic, yes, but the more likely failure mode is slow epistemic drift rather than one spectacular civilization-ending falsehood. Once a claim is repeated across secondary sources, the system starts treating repetition as validation.

• –Scientific publishing already has a citation-integrity problem; recent Nature coverage says hallucinated citations are polluting the literature and that LLM use in scholarly writing creates a provenance problem.
• –Peer review and standards help, but they are uneven and slow; they are much weaker once claims move into blog posts, vendor docs, tutorials, and code comments.
• –The best defenses are boring and procedural: primary-source citations, provenance metadata, automated citation checks, and explicit human ownership of high-stakes claims.
• –RAG, citation-grounded generation, and uncertainty-aware detection methods can reduce the blast radius, but they do not eliminate the need for verification.
• –The catastrophic version is most plausible in narrow technical or clinical domains where a false assumption gets embedded into standards or workflows and nobody revisits the original source.