Real journal. Real volume. Real issue. Wrong world.
Borges opens "Tlön, Uqbar, Orbis Tertius" with a mirror and an encyclopedia. The mirror is unsettling because it multiplies what is real; the encyclopedia is worse, because it invents what is not. His narrator finds an entry for a country that does not exist, written in the flat, confident tone of reference works, complete with geography and a literary tradition. The entry has no source. Then objects from the invented country start turning up in the real world: a compass whose markings belong to Tlön, a small cone of impossible weight. Ideas, it turns out, can leave artefacts. Describe a thing thoroughly enough, in a place people trust, and the description starts to acquire mass.
The story was published in 1940, and for decades it read as a fable about totalitarianism, about how a disciplined fiction can colonise a reality that stops checking. It reads differently now, because the mechanism it describes has become a routine production process. The encyclopedia entry that has no source is being written constantly, at scale, in the exact format that makes it trustworthy. And the objects are turning up.
The entry that got into the encyclopedia
Reference 33 of the bibliography attached to South Africa's withdrawn draft National AI Policy cites Karr and Smith, 2023, "Digital Rights and AI Governance in Africa," in the Journal of African Law, volume 67, issue 1, pages 95 to 108.
Everything in that string is real except the article. The journal exists. The volume and issue exist, dated February 2023. The page range exists, and in the actual issue it is occupied by two entirely different studies: a study of presidential vetoes in Kenya that ends on page 96, and a study of sentencing reform in Tanzania that begins on page 97. The journal's own table of contents settles the matter in about a minute. The article, and its authors, are not there.
This was the seventh fabrication in a sixty-seven item bibliography. Six had already been reported when the policy was withdrawn, sixteen days after gazetting. The seventh surfaced only when the reference list was run through a verification sweep months later, and it is the instructive one, because it shows how little an invented citation needs in order to pass. A plausible title built from the field's own vocabulary. Two plausible author names. A real journal, chosen because the topic belongs in it. Pagination inside a real issue. A fabricated reference is convincing for the same reason a forged banknote is: every checked feature is genuine except the thing itself. A language model produces the shape of a source with perfect fidelity, because the shape is what it has learned. Existence is not part of the shape.
And the entry is now in the encyclopedia. The gazette is a real object: Government Gazette No. 54477, General Notice 3880 of 2026. The fabricated references inside it have been published by the state, indexed by commenters, scraped, summarised, and fed back into the systems that produce the next round of documents. For sixteen days, national AI policy in Africa's most industrialised economy stood on a chain of evidence with seven ghosts in it. Then the document was withdrawn, and the ghosts stayed in the record, because withdrawal removes the host and leaves the artefacts. Tlön's objects were heavy and unsettling, stamped with alien script. Ours are well-formatted.
A population event, measured
It would be comfortable to file this as one department's bad month. The measurement says otherwise.
In May, an audit published as a correspondence in The Lancet reported on 97.1 million references drawn from nearly two and a half million PubMed-indexed biomedical papers. The researchers, Maxim Topaz's group at Columbia, used a trained classifier to separate genuine fabrications from formatting noise, and found 4,046 references to papers that do not exist, spread across 2,810 published articles. The rate is the finding. In 2023, roughly one paper in 2,828 carried at least one citation of a nonexistent work. In 2025, one in 458. In the first weeks of 2026, one in 277. A twelve-fold rise in two years, with the steepest climb in mid-2024, when AI writing tools went mainstream.
At one in 277, the problem stops being a scandal of careless authors and becomes a background rate, a level of contamination the record now carries the way seawater carries salt. At that density, fabricated sources stop being isolated hoaxes and become an environmental condition: every literature review and every training crawl is drawing from water that has a measurable proportion of ghosts in it. The South African bibliography was early, and it was caught because a journalist looked. The audit suggests how much of the same material is entering the record without anyone looking at all.
What a citation claims
A citation looks like an administrative courtesy, a bit of scholarly housekeeping. It is actually a claim: that a specific work exists and says roughly what the citing document says it says, offered by someone who stood in some relation to it. The whole apparatus of evidence-based policy, and of any argument that offers to be checked, runs on the assumption that these claims are mostly true and that the false ones are aberrations somebody will catch.
Both halves of that assumption are now under strain, and the strain is asymmetric. Producing a plausible citation costs nothing; the machine will generate fifty in the time a person reads one. Confirming a citation costs about a minute of human attention: open the journal's table of contents, read down the issue, find the paper or fail to. The verification cost has stayed flat and human while the fabrication cost has fallen to zero, which is the same asymmetry Voice After Abundance found in persuasion: when producing claims becomes free, the cost of the whole exchange moves to whoever has to check.
Tlön's real subject was never the hoax. It was the speed at which an invented reference becomes a real object once something downstream cites it and files it. A fabricated paper that has been gazetted and summarised has an existence in the record that is independent of its existence in the world, and the two are drifting further apart as generation gets cheaper. One question runs under everything that follows: what is a reference list actually for, if the entries in it are claims nobody is assigned to test?
Companions
- The seventh fabrication, reference 33, and the full sweep of the withdrawn policy's bibliography: The Reference Check.
- The measured contamination rate, 97.1 million references audited: Topaz et al., The Lancet, May 2026, reported by Retraction Watch.
- An article that runs this check on its own sources while you read it: Sixteen Days.
- The episode that sets the week's question: The Place to Check.
- The fiction: Borges, "Tlön, Uqbar, Orbis Tertius," an invented country whose objects begin turning up in the real world.
These notes come out of Sociable Systems, a practice that reads AI-shaped documents the way a hostile reviewer will, before a lender or a court finds the gap. The argument has an operational form: the Interim Protocol sets out four rules for AI use in environmental and social deliverables, covering disclosure at touch-point grain, evidence custody, the phrases no automated screening may settle, and a hostile read before anything ships. Free, and written to be cited or retired once institutional guidance arrives.