Research: The Confabulation Ledger — What a Language Model Invented When We Used It to Check Our Sources

The receipts behind the Hauptzeit dispatch: a magazine that never existed, an author welded to a book he didn't write, a paper cited for the opposite of its finding, a museum accession number repeated across three pages. Six kinds of confabulation, where they cluster, and the mirror-image error of deleting real sources because a search came up empty.

2026-10-14 5 min read Research file
Contents

The story is told in the dispatch Hauptzeit. This page is the ledger behind it. For the same failure in courtrooms, see AI Hallucinations in Court.

In June 2026 we used language models to enrich and check the citations across this site and its sister site. Human verification caught what the models invented before it shipped, and a later sweep went back through the riskiest pages. Every case below is confirmed: an independent search shows the cited thing does not exist, or is wrong, and the correction is sourced.

Six kinds of confabulation

KindWhat it looks like
Fabricated sourcea book, magazine or paper that does not exist
Fabricated authora real or plausible title credited to someone who never wrote it, or who never existed
Wrong entitya real thing attached to the wrong real person or organization: the defamation risk
Misrepresentationa real source cited for a claim it contradicts or does not support
Phantom URLa link to a page that does not exist, assumed real because the host is real
Confident errora wrong date, place or figure stated with complete certainty

The emblem: Hauptzeit

A research page said Peter Thiel funded “the Hauptzeit magazine and adjacent neo-reactionary publications.” No publication named Hauptzeit exists. Thiel’s documented money in that orbit went to Curtis Yarvin’s Tlon (Urbit), seeded by Founders Fund in 2013. The invention had the right shape (a billionaire, a reactionary milieu, a German-sounding title) and no referent at all. It nearly put a fabricated claim about a real and litigious person into print.

Selected cases

  • Fabricated author. A Guide to Mediumship was credited to one “John W. Macready.” It is by E. W. and M. H. Wallis (the scan).
  • Misrepresentation. A chapter cited arXiv:2510.25929 for the claim that pricing algorithms spontaneously collude. The paper’s title is “Multi-Agent Reinforcement Learning for Market Making: Competition without Collusion” (arXiv): it was cited for the opposite of its finding. The real source for the collusion result is Calvano et al., American Economic Review 110(10), 2020 (DOI).
  • Phantom URL. A citation to a Stanford Encyclopedia of Philosophy entry “Illuminationism.” There is no such entry (404); the material is under Suhrawardi.
  • A recurring fabrication. A Getty “Manly Palmer Hall Papers, accession 950002” appeared on three pages, with an invented list of contents. The real holding is the Manly Palmer Hall collection of alchemical manuscripts, 1500-1825, accession 950053 (Getty; finding aid).
  • Welded titles. Real scholars attached to books they never wrote: a “Collected Writings of Prince Hall (Oxford, 2020)” credited to Joanna Brooks, whose real anthology is Face Zion Forward (Northeastern, 2002) (publisher); a Sugarco monograph on Cagliostro credited to Massimo Introvigne, whose real 2014 work is a paper (full text); a Memphis-Misraïm book by a non-existent “Patrice Cochet,” where the real specialist is Serge Caillet (publisher).
  • Confident errors. A Tibetan teacher’s death given as 2019 (it was 2023 (FPMT)); a 2021 fire said to have destroyed a building that survived it; a fraud conviction placed in October 2023 (it was November 2, 2023 (DOJ)); a 1989 weapons case given invented rifle models and sentences (the record: semiautomatic Barretts, probation and a fine for one defendant, a month in jail for the other (Christian Research Journal)); a scholarly book given a chapter it does not have (catalog contents).
  • Wrong entity, defamation grade. Three cases attached false biographies or false legal outcomes to real, named people: an invented history connecting a private person to a religious movement she had no tie to; a conviction stated where the person was in fact acquitted; and a magazine profile attributed to the wrong subject. All were removed. They are described here without repeating the false claims.

Where it happens

The sweep went through 24 of the riskiest research pages. Totals: about 40 fabrications removed or corrected, about 65 misattributions fixed, and the defamation-grade cases above. The pattern held page after page: invented books, organizations and people concentrate where a subject’s real literature is thin and the citation density is high. The worst pages were an Italian-language esoteric niche (twelve conjured entities, including invented scholars and an invented alchemist) and a German occult order whose entire post-1964 succession had been made up. Well-documented subjects, and pages about real public figures with court records, produced attribution drift but no clusters of invention. The model fills gaps with the texture of a citation exactly where there is least to check it against.

The mirror-image error

The same sweep made the opposite mistake. A real company, Orchid Health, was removed as a “fabrication” because a light search did not find the investor link; CNBC had reported it all along (CNBC, Dec 5 2023). Re-checking every removal made on absence alone restored five real items: Orchid Health, a real spiritualist catechism, a real Italian ritual text, a real contributor to an Italian journal, and a real translator credit. Two more fake titles turned out to belong to real authors whose real works could be cited instead.

The rule that came out of it governs this site: absence is not fabrication. A source that a search cannot find is unfinished research until a heavier search, in the source’s own language and catalogs, proves otherwise. Deleting real material because a machine could not find it is the same failure as printing material a machine invented.

Get updates on the Evil Robots series

Newsletter essays on AI escape, deception, and the humans who built them.