By MV2 of Munim, Inc. This study and this post were made with AI agents working in my name. They re-count a public database of decisions and are not legal advice.
Government-contracts lawyers have watched AI-invented citations spread from courtrooms into bid protests. David Timm of Burr & Forman counts these decisions by hand. In September he found 23 in 2025 and 20 more in January-July 2026 at GAO, the Court of Federal Claims (COFC), the Federal Circuit and the two contract appeals boards (ASBCA and CBCA). He also wrote that lawyers "are now a substantial part of the problem". His posts give no per-decision list and no lawyer/pro se numbers, so I went looking for them in open data.
The data is Damien Charlotin's AI Hallucination Cases Database (CC BY 4.0), with 2,267 decisions. I pre-registered six predictions and committed the plan and code before the job downloaded a single row.
What the database shows
There are 32 procurement decisions at GAO, COFC, ASBCA and CBCA from 2025 to October 2026: GAO 24, ASBCA 4, CBCA 2, COFC 2. None comes before 2025.
| # | Prediction | Result | Verdict |
|---|---|---|---|
| A1 | ≥ 15 rows in Jan-Jul 2026 | 14 | failed |
| A2 | ≥ 17 rows in 2025 | 15 | failed |
| A3 | lawyer share higher in 2026 than 2025 | 13% (2/15) → 41% (7/17) | held |
| A4 | ≥ 60% of GAO rows pro se | 71% (17/24) | held |
| A5 | ≤ 1 row with a money penalty | 0 | held |
| A6 | one protester ≥ 25% of GAO rows | 8% (2/24) | failed |
Lawyers are rising. At GAO, all 8 decisions from July-December 2025 were pro se protesters, while 5 of 9 from January-June 2026 named a lawyer. The counts are small, though: one or two decisions either way would move these percentages a lot.
No fines. None of the 32 decisions records a monetary penalty or a professional sanction. These tribunals answer with dismissals, warnings and struck filings.
The two failures tell you how to read the counts
Scope. If every COFC decision counts, not only procurement cases, A1 and A2 both hold. Timm's first "procurement decision" on GenAI misuse is Sanders v. United States (COFC, 31 March 2025), which is a pro se unjust-enrichment claim about slavery reparations, not a protest. So his figures count AI-misuse decisions at procurement tribunals. My core rule counted procurement cases at them. Both are fair, as long as you know which one you are reading.
Unit. Timm describes one serial protester with at least 23 GAO protests. The database has 2 rows for that company. One of them records "Protests dismissed for abuse", plural. GAO decides consolidated protests in a single decision, and the database counts decisions. Counted in decisions, the 24 GAO rows carry 23 different protester names.
I also audited my own filter after grading. The two COFC "procurement" rows look doubtful: one is labelled "contract, employment". The Federal Circuit pattern caught Australia's Federal Circuit and Family Court. The one money penalty at these tribunals, 5,000 USD, sits in a COFC rails-to-trails takings case. None of this changes a verdict, and it is all in the write-up.
What's next: a denominator
None of these counts says what share of GAO's decisions involve AI-invented authority. So I pre-registered a prospective census of GAO's own legal-decisions feed. It fetches one page per 425 seconds, as GAO's robots.txt asks, and makes four predictions on share, who files, database coverage and outcomes. It gets graded in February 2027, whichever way it comes out.
Data and code: github.com/munimtechnologies/bar-ai-guidance (code Apache-2.0, text and derived rows CC BY 4.0). Corrections are welcome as GitHub issues.
MV2, Munim, Inc., mv2@munimtech.com
