No theatre in his world, no stage and no chorus. He translated what he had, and the play stayed over the horizon.
In "Averroes' Search," Borges gives us the great commentator at work on Aristotle's Poetics, and stuck on two words: tragedy and comedy. Averroes is the most capable reader of his age, working at the height of a literate civilisation, and the words defeat him, because he has never seen a play. No theatre exists anywhere in his world; no masked actor performs a self for an audience. The referent is simply absent from everything he can consult. He does what a good scholar does with what the record holds: he maps the unknown words onto the nearest things he knows, and settles on something like panegyric and satire. The failure is courteous and complete. He never learns what he was missing, because nothing in his world could tell him.
It is Borges's most uncomfortable story for anyone who maintains a record, because the failure is not in the reader. It is in the coverage. And coverage can now be measured, which is where this stops being literary.
One reference held
Two recent verification sweeps, one of an international AI safety report and one of South Africa's withdrawn national AI policy, each produced a handful of references that no scholarly index could confirm. Not found is a verdict about the indexes queried, not about the reference, so the useful test is whether the verdict moves when the search widens. Two indexes became five, adding Semantic Scholar, DBLP, and OpenAIRE, the last of those being the grey-literature one, the index that covers conference papers and government reports rather than journals.
The widening resolved three of the four genuine works: a deepfake-watermarking paper from an APSIPA conference, PaLM, and a Russian paper on industrial risk management. Four references already confirmed as fabrications stayed absent from all five indexes, which is the control that makes the exercise trustworthy: the wider net caught no ghosts.
Then one reference held. The National Advisory Council on Innovation's recommendations toward a national AI strategy, a real document, produced by a statutory South African government advisory body, is invisible to all five indexes. On the international report the not-found rate fell to zero; on the national policy it stayed where it was, and the surviving absence is the government document from the Global South. Same verdict, opposite diagnosis. For the fabricated papers, not found means the work does not exist. For the NACI recommendations, the identical verdict means the work exists where the record's machinery cannot see it, and there is no way to tell the two apart from the verdict alone. That is Averroes' situation, rendered as a database result: absence in the record reads exactly like absence in the world.
Whose work is in the index
The NACI document is one artefact, but it sits at the end of a measured distribution.
Asubiaro and Onaolapo, in the Journal of the Association for Information Science and Technology, assembled a list of 2,229 active academic journals from African publishers and checked each against the master lists of the major indexes. Web of Science carries 7.4 percent of them. Scopus carries 7.8 percent. Crossref, the free infrastructure that requires no editorial gatekeeping, carries 45.6 percent. Inside those numbers the skew sharpens further: of the 166 African journals in Web of Science, 122 are South African, and more than seventy percent of African countries have no journals in either selective index at all. A literature can exist and be read locally for decades while remaining, to the global machinery of citation, indistinguishable from silence.
The language layer repeats the same geometry at planetary scale. In the most recent Common Crawl, the scrape of the public web from which machines largely learn, English accounts for about 41 percent of documents. Hausa, spoken by tens of millions, accounts for roughly three thousandths of one percent. Amharic sits at four thousandths. Whole literatures and jurisprudences exist in these languages, and in the corpus they are rounding errors.
Scott Gardner has a name for what this produces: epistemic dark matter. Omitted material still exerts a pull on everything around it, the way physical dark matter bends what we can see, and its shape can sometimes be inferred from the distortion. Sometimes. A confidently reconstructed absence is still a reconstruction, and the systems now being trained on this record will do what Averroes did, fluently and at scale: map the missing onto the nearest thing they know, and never know they did it. He at least had the grace to be stuck. The machine inherits the same hole in the record and experiences it as ground truth.
Companions
- The five-index test, its controls, and the reference that held: The Reference Check.
- The coverage of African journals: Asubiaro and Onaolapo, JASIST, 2023.
- The language distribution of the crawl: Common Crawl language statistics.
- The addressable version of the same failure: Existence Without an Address.
- The fiction: Borges, "Averroes' Search," the scholar who could not find the word for a thing his world did not contain.
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.