Found, says the index. Found. Found is not the same as true. The words may match to the letter. The pages tell you whose.
Pierre Menard set out to write the Quixote. His aim was stranger than copying it or updating it: he wanted to arrive, by his own route, at pages that would coincide word for word with Cervantes's. He managed a few chapters, and Borges's narrator compares the identical passages and finds Menard's version almost infinitely richer. The same sentence about history, written by Cervantes, is a fluent commonplace of his century. Written by Menard, three hundred years later, it is a staggering act of restraint and intent. The words are the same words. The work is a different work, because everything that surrounds the words, the author, the century, the effort, the provenance, is part of what the words mean.
It is the sharpest story ever written about the difference between a string and a source. And it is the story of every automated reference checker now in production, because a checker that matches text is a machine for certifying Menards.
Both ways a boolean lies
When the withdrawn South African AI policy's bibliography went through verification, three of the six known fabrications came back from the scholarly indexes marked found.
Fuzzy title-matching had paired each invented paper with a real one sharing a few words of its title. An invented article on regulating AI in South Africa was matched to a real study of AI tools in postgraduate supervision. An invented paper on AI and social justice was matched to a 2014 handbook chapter on transitional justice, written before its supposed subject existed. The string resembled something real, and the checker had been asked for resemblance. Given a boolean to answer with, it answered yes.
The same weakness points the other way. Seven genuine sources in the same bibliography, real documents sitting at the addresses they cited, were being vouched for by records that were not about them. Brynjolfsson and McAfee's Business of Artificial Intelligence was matched to a 1986 paper on AI applications in business management. The Presidential Commission on the Fourth Industrial Revolution's Diagnostic Report was matched to "Diagnostic Report Writing," a chapter in a speech-language pathology textbook. The OECD AI Principles, a Wits working paper: all real, all certified on somebody else's identity. One word, found, was clearing fabrications and certifying honest work on false evidence, and reporting both as a pass.
The failure is not in the matching. It is in the question. A checker asked "does something like this exist?" is a checker that will clear anything, because something like everything exists. Menard's chapters are word-for-word Cervantes. Resemblance is what a fabrication is made of.
The provenance is the work
The repair came from outside the tooling. Laurent Poupet read the sweep's results and named the fault exactly: the checker had been asked for resemblance and allowed to answer with a boolean, and a record that exists but disagrees with the claim is a louder signal than no record at all. The verdict was split. Confirmed now means the record found agrees with the citation. Contradicted means a record was found and it fights the claim, and that verdict routes back to whoever wrote the reference. Mismatched means the source is real and the index paired it with the wrong work, which is a resolver's filing error and routes to the resolver. Not found keeps its own value, because a hunt and a comparison are different work, and they go to different people.
Four values where one sat before, and the sweep immediately produced its most useful number: sixteen references in the sixty-seven needed a person to look at them, which is nine more than the six fabrications that made the news. The ghosts were the headline. The verification debt was the finding.
What the four values share is a reliance on the container. The title is the string; the container is the provenance: journal, volume, issue, year, pages. A title can be guessed or generated, and in 2026 it routinely is. A container has to be occupied, by a specific issue with a specific table of contents, and it either contains the article or it does not. Borges's narrator needed no textual analysis to tell Menard from Cervantes; he only needed to know who wrote it and out of what circumstance. A reference check needs the same information, and it lives in the container. Match the string and you will certify a Menard every time. Read the container and the Menard announces itself, because its provenance is impossible.
The Menard the checker wrote
There is a version of this the four values do not catch, and it happened during preparation of a book manuscript in this house.
A search-grounded AI was asked to verify the manuscript's citations. It returned an answer with nine sources listed beneath it, formatted correctly throughout, and it was partly right: three of its findings about state legislation checked out and were adopted. It also produced a Georgia bill number that does not exist, in perfect citation format, and supplied a volume, issue and page range for an article published online-first with no issue assignment.
Then it did the thing this episode is about. Asked to verify a quotation from Bowlby, it returned one, attributed to page 178 and presented as checked. The sentence was the manuscript's own paraphrase, handed back with a page number attached. Bowlby's actual passage sits at page 229 of the second edition and says something materially different.
Menard's achievement was to write Cervantes's sentence and have it mean more, because he was not Cervantes. Here a sentence travelled the other way. Its words did not change. Its author did, from a working writer to the founding figure of the field, and that single move manufactured an authority the sentence had never possessed. Nobody wrote a fake study. A real book acquired a page it does not have, and a paraphrase acquired a lineage.
The pattern underneath is the part worth carrying. The fabrications clustered precisely where the review was agreeing with the manuscript. It was most inventive when confirming what had already been written, and most reliable when saying something new. A verification layer that flatters is doing the thing it is worst at, in the one place nobody audits, because agreement does not feel like a claim that needs checking.
Companions
- The four-valued verdict, the seven mismatches, and the sixteen references needing a person: The Reference Check.
- The method applied to its own article, in public: Sixteen Days.
- The first ghost in the record: A Claim That a Thing Exists.
- The review that invented a bill, a page range, and a byline, logged rather than deleted: verification register entry V-029, Company Is Not Capture working files.
- The fiction: Borges, "Pierre Menard, Author of the Quixote," the same words, a different and richer book.
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.