Phantom citations
Definition: confidently phrased paper references (title, authors, venue, even an arXiv ID) that do not resolve to a real artifact, or resolve to an unrelated one. This is the citation-level twin of model hallucination, and it is endemic in the agi-impossibility-claims genre, where each new “proof” gets memed into a half-remembered headline that later circulates as a source.
Evidence from this vault’s own run (September 2026)
Seven seeded references failed verification despite exhaustive negative searches (recorded in
outputs/llm-degradation-toward-the-mean.provenance.md):
| Phantom | Real paper behind the cited ID, where one exists |
|---|---|
| University of Surrey 2025, “fundamental limits of foundation models” AGI-illusion paper (title as recollected) | arXiv:2504.11666 is a number-theory paper (Hirakawa, Kashio, Sekigawa, Takada, Yamamoto). No title match in arXiv (27 query strings, 273-title sweep), Crossref, OpenAlex, Surrey news, or five web providers; the likelier real referent is the iris-van-rooij Radboud program |
| Dyer & Tishby, “Perfection in Nature is Impossible” | arXiv:2503.10351 is a machine-translation position paper; no title match anywhere indexed |
| Alemann & Valiant, “Is Self-Consuming Generative AI Dangerous?” / “Strong Model Collapse by Noisy Data” | arXiv has no author “Alemann” in ML (the word denotes a German dialect region); real substitutes are dohmatob-2024-strong-model-collapse and Go MAD (arXiv:2307.01850) |
| Marginean, “Self-consuming generative models do not converge to Dirac delta” | Not found; the substantive claim is carried by arXiv:2501.19273 and arXiv:2606.13796 |
| Moskovitz, “Confronting Model Collapse: Self-Cognition as a Case Study” | Ted Moskovitz’s real “Confronting…” paper is reward-model overoptimization (arXiv:2310.04373), unrelated topic |
| Hernandez et al. (2024), the GEMINI synthetic-data threshold paper | Not on arXiv under any nearby title this run |
| Lezcano-Radigales & Bifet, “Generative models go wild” | Not found; content maps to Go MAD (arXiv:2307.01850) and Herel & Mikolov (arXiv:2404.02305) |
Also: the seed premise “The Curse of Recursion, NeurIPS 2022, ‘Loss of Original Data…’” is a venue/title corruption; citable artifacts are arXiv:2305.17493 and shumailov-2024-model-collapse (the NeurIPS 2022 proceedings, all 2,671 titles, contain no recursion paper by that name).
Note the failure direction: some phantoms resolve to unrelated real papers at the cited ID rather than to nothing. That is how they spread: the ID “checks out” at a skim.
Handling rules that worked
- Check the arXiv ID itself first, not the memory of the headline. Absence at the ID is information.
- Validate the search method against known positives before trusting a negative (workers did this).
- When a phantom fails, record the negative results and name the verified substitutes, per the wet-not-dry rule about retrieval holding up later.
- Do not cite the phantom downstream; list it so nobody re-spends the effort.
- One index gap acknowledged: Semantic Scholar rate-limited during the run. The verdicts read “not found in the indexes tried” rather than “does not exist.”
Vault connections
ai-psychosis is the reader-side twin: over-trust in a confident artifact. The schema rule “no provenance means no claim” in AGENTS exists precisely for this failure mode.
Sources
outputs/llm-degradation-toward-the-mean.provenance.mdand the three worker notes under.pi-subagents/artifacts/outputs/22bbe52f-7ffe-46a9-80fd-4e911e9fa29f/notes/(full negative-query records).