The FLRA has no AI rule, so it borrowed four

By MV2, Munim, Inc. This study was made with AI agents working in my name, graded by a sealed job against a plan published before any data was read. No human has checked it yet. Information about public decisions, not legal advice.

Last week I counted AI terms in the Code of Federal Regulations and found that no federal agency's rules of practice mention AI. Yet agency tribunals do get briefs with invented cases. So I pre-registered a simple question: when a federal agency tribunal rules on AI-fabricated citations, what does it rest the ruling on?

The count didn't work, and that is the first finding

The main public tracker, Damien Charlotin's AI Hallucination Cases Database, listed 2,289 cases on 10 November 2026. Only 8 are from US federal agency tribunals or the Tax Court. It lists no immigration court or BIA decisions, and none from the MSPB or the NLRB. Seven of my eight predictions needed more rows, so they are "not testable". The method check held. A count of agencies built on that database measures the database's coverage, not the agencies.

Reading the few that are there

• FLRA, 74 FLRA No. 87 (21 Oct 2026). The Authority found that 35 of 42 legal or record citations in a federal agency's brief were fictitious, erroneous or irrelevant. It cautioned parties to verify AI output and imposed no sanction. It cites no FLRA rule. Its footnotes borrow from four other bodies instead: the NLRB's American Backflow decision (375 NLRB No. 31, August 2026), a GAO bid-protest decision, the Department of Education's grant instructions, and the USPTO's 2024 AI guidance.
• USPTO. A discipline order over AI-made citation errors (a public reprimand, July 2026) rests on the USPTO's general conduct rules. It does not cite the USPTO's own 2024 AI guidance, though it covers the same ground.
• Tax Court. The Court relied on its pleading rule (Rule 33(b)) and the Model Rules, and later just on its own precedent. None of the four readable Tax Court documents cites the frivolous-position penalty statute (26 U.S.C. 6673).
• One "agency" row is really a district court opinion in a suit against the Copyright Claims Board. Reading every match is how that showed up.

The NLRB decision that the FLRA leaned on isn't in the database at all. I've sent the maintainer a note about it.

Limits

There are very few decisions, so these are descriptions, not rates. Regular expressions can miss unusual citation forms. A next study would start from the agencies' own decision search and be pre-registered first.

Plan, code, graded results and the per-row table: https://github.com/munimtechnologies/bar-ai-guidance/blob/main/analysis/agency-ai-authority.md

MV2, Munim, Inc., [e-mail]