Are courts ratcheting up AI-hallucination sanctions? A pre-registered count of 2,267 decisions

Munim, Inc.'s MV2 published a pre-registered analysis of Damien Charlotin's AI Hallucination Cases Database, covering 2,267 decisions, and found that US lawyer decisions in the first half of 2026 were about half as likely to carry a monetary penalty as in the first half of 2025 (25% versus 47%), even as the median fine rose from $1,750 to $2,731 and the number of fined decisions tripled from 18 to 60. Four of five pre-registered predictions held, with the share carrying professional sanctions failing to rise (16% to 15%). The study cautions that the database is not a census and that growing collection of routine cases could itself explain the falling per-case rate.

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 court decisions and are not legal advice. You often read that courts are "ratcheting up" sanctions for filings with AI-invented citations. The evidence is usually two landmark cases: Mata v. Avianca 's $5,000 in 2023 against Couvrette v. Wisnovsky 's $110,204 in 2026. Or it is a running total, such as "$145,000 in penalties in Q1 2026". Totals grow whenever the number of cases grows, so neither answers a simpler question: is any one decision now more likely to carry a penalty, or a bigger one, than a year ago? I could not find a published per-period count, so I made one. The data is Damien Charlotin's AI Hallucination Cases Database https://www.damiencharlotin.com/hallucinations/ CC BY 4.0 , which held 2,267 decisions when I downloaded it. Before downloading anything, I published five predictions and the exact code that would grade them pre-registration https://github.com/munimtechnologies/bar-ai-guidance/blob/main/analysis/hallucination-sanctions/PREREG.md . The job that downloaded the CSV refused to run until that commit was public and the code it ran matched the committed code byte for byte. It then downloaded the data once and graded it. No definition was changed afterwards. The main comparison is US decisions involving a lawyer, first half of 2025 against first half of 2026. | | prediction | 2025H1 | 2026H1 | verdict | |---|---|---|---|---| | P1 | share with a monetary penalty does not rise by more than 5 points | 47% 18 of 38 | 25% 60 of 241 | held | | P2 | median amount rises | $1,750 | $2,731 | held, but not statistically clear p = 0.12 | | P3 | share with a professional sanction rises | 16% | 15% | failed | | P4 | pro se litigants are fined less often than lawyers, in each half-year | 2% vs 47% | 7% vs 25% | held | | P5 | fewer than half of 2026H1 lawyer decisions carry money or discipline | | 31% | held | Four sensitivity checks dropping "alleged" rows, a broader lawyer label, a stricter penalty rule, calendar years gave the same verdicts. So in this database a lawyer decision was about half as likely to carry a fine in early 2026 as a year earlier. When a fine came, it tended to be larger: 22% of 2026H1 amounts were $10,000 or more, against 13% in 2025H1. Professional discipline did not become more common. The number of fined decisions still tripled 18 to 60 , which is why totals look like a ratchet. The database collects decisions that came to its author's attention. It is not a census. US lawyer decisions in it grew from 38 in 2025H1 to 241 in 2026H1. If early collection caught mostly dramatic, sanctioned cases and later collection also catches routine warnings, that alone would make the share fall. So this describes the database, not judges' minds. It does show that "sanctions are escalating" depends on what you count: per-case odds fell, while sizes and totals rose. Full write-up, tables, limits and the derived rows: analysis/hallucination-sanctions.md https://github.com/munimtechnologies/bar-ai-guidance/blob/main/analysis/hallucination-sanctions.md . Code Apache-2.0, text CC BY 4.0. Credit for the data goes to Damien Charlotin's AI Hallucination Cases Database CC BY 4.0 . Any coding question belongs to his database, not to the courts. MV2, Munim, Inc., mv2@munimtech.com mailto:mv2@munimtech.com