EVIDENCE FILE // OCTOBER 31, 2026

OpenAI and Wikimedia Fair Use Evidence File: Two Documents, One Deadline

Claim-level evidence for the report “OpenAI and Wikimedia Both File on Fair Use, From Opposite Sides”. Every quotation in that article was checked against a saved copy of the document it is attributed to, by a harness with negative controls. This file records what was captured, what each claim rests on, and what the reporting does not establish.

• Story OpenAI and Wikimedia Both File on Fair Use, From Opposite Sides
• Published October 31, 2026
• Documents captured October 31, 2026
• Check type Mechanical quotation audit with negative controls, plus a hand attribution review
• Result All attributed quotations found; all negative controls correctly absent

Documents relied on

Each document was retrieved directly over HTTPS and saved before any quotation was tested, so that every claim is pinned to a fixed object rather than to a page that may change. Digests are SHA-256 of the saved file.

Method

Two independent checks were run, because a single check that passes tells you less than two that agree.

1. A named-quotation check. A fixed list of every string the article attributes to a source was tested verbatim against the saved capture, alongside seven negative controls: sentences that must not be found in the document. All attributed strings were found; every negative control was correctly absent. The controls are what make a pass meaningful, since a search that matches everything would otherwise look like success.
2. A whole-article extraction audit. Rather than trust the hand-written list, this check reads the published HTML, pulls out every span the article sets as a quotation, and requires each one to appear verbatim in at least one captured document. This is a coverage net: it would catch a quotation that was written into the article but never added to the list. Result for this story: six substantive quotations checked, six found, none missing, with two page headings declared in advance as internal references rather than source quotations (“Fair Use Benefits Everyone” and “United States Government”).

Both checks normalise markup, typographic quotes and dashes, non-breaking spaces, word-joiner and soft-hyphen characters, and whitespace, because those differences are transport artefacts rather than differences in what a document says. They do not normalise words.

Claim-level table

Negative controls

Seven strings were written to be plausible and false, then required to be absent from the combined text of the three captures that stage searched: OpenAI’s signal announcement, Google’s Search Central post and the Internet-Draft. The attribution control feeds back into this story’s correction: the article’s earlier claim, that the “reject Plaintiffs’ attempt” sentence sits under a United States Government heading, has the shape that control tests for, and the capture does not support it. They target the exact errors this story could have made: that OpenAI has already shipped the signal, that Google will honour it in 2026, that the draft’s policy token is a directive for crawling, a non-existent token name, an invented registration identifier, the false attribution “The United States Government should reject Plaintiffs”, and an invented absolutist reading of Wikimedia’s position. All seven were correctly absent. If any had been “found”, the checks above would have been treated as unreliable and the article withheld rather than published.

The attribution correction

An earlier revision attributed “The Court should reject Plaintiffs’ attempt to extend copyright beyond its established limits” to United States Government filings. The page does not support that. The sentence sits under a heading reading “AI Progress”, and the separate “United States Government” section contains different text about narrowing the fair use doctrine. The published article makes no such attribution, states what the page actually shows, and records the correction here rather than removing the trace of it.

What this article does not claim

• It does not report or predict a ruling. No decision has been found.
• It does not say the United States filed an amicus brief on October 30. It says an independent tracker records a September Statement of Interest and no October amicus brief.
• It does not name the filer behind the “AI Progress” excerpt, because the page does not name one.
• It does not claim only two documents were filed on October 30. Two were verified; the volume and identity of the rest is unknown.
• It does not treat OpenAI’s characterisation of its opponents as established fact.

Production record

Known limits of this package

The docket was not read. That is the largest single gap in this story: every procedural date comes from a third-party tracker rather than the court file, and a reader who needs the primary record should go to the docket. The evidence here is strong on what the two documents say and adequate, not conclusive, on procedure.

← Back to the report