Munim Technologies

Research/Society

Are US courts getting tougher on AI-invented citations?

In a public database of court decisions on AI-hallucinated filings, US decisions involving a lawyer were about half as likely to carry a money penalty in the first half of 2026 as a year earlier (25% vs 47%). Penalties that were imposed got somewhat larger, but the earlier group is small.

Nov 6, 2026 · 6 min read

Lawyers keep filing briefs with cases that do not exist, invented by AI tools. Commentators often say courts are "ratcheting up" sanctions in response. We tested that claim on Damien Charlotin's AI Hallucination Cases Database, the largest public collection of such decisions, with predictions fixed before we downloaded it. A US decision involving a lawyer was about half as likely to carry a monetary penalty in the first half of 2026 as a year earlier: 25% against 47%. When money was imposed, amounts were somewhat larger, with the median rising from $1,750 to $2,731, but that rise could be chance. The earlier group is small, only 38 decisions, and that matters for how far to trust the comparison.

47% → 25%

US lawyer decisions with a monetary penalty, 2025H1 to 2026H1

n = 38

lawyer decisions in 2025H1, the comparison base (241 in 2026H1)

$1,750 → $2,731

median penalty when money was imposed (p = 0.12)

31%

of 2026H1 lawyer decisions carried money or a professional sanction

The question

Are courts punishing AI-invented citations more harshly over time? Trackers usually compare landmark cases, such as the $5,000 sanction in Mata v. Avianca in 2023 against a $110,204 award in 2026, or add up quarterly totals. Totals grow when the number of cases grows, so they cannot say whether any single filing is more likely to be punished. We found no published per-period share of decisions carrying a penalty, and no per-period median.

What we did

We wrote five predictions and a frozen analysis script and made them public on 6 November 2026. A job wrapper refused to download the data until that commit was public and the script's hash matched. It then downloaded the database once: 2,267 decisions, 1,536 from the US, of which 602 involve a lawyer and 913 a self-represented (pro se) litigant. We grouped decisions by half-year and compared 2025H1 with 2026H1, using the database's own party labels and penalty fields. A second pre-registered study then read the hosted decision documents to check the dates the first study relied on.

What we found

The share of lawyer decisions with a monetary penalty fell in every half-year after 2025H1. The 95% intervals for 2025H1 (32.5-62.7%) and 2026H1 (19.9-30.7%) do not overlap. The share with a professional sanction, such as a bar referral or suspension, did not rise either (15.8% against 14.9%), so our prediction that it would rise failed. Seven in ten 2026H1 lawyer decisions carried neither money nor discipline.

US lawyer decisions carrying a monetary penalty, by half-year

2026H2 was still being collected and was not part of the comparison. Shares from the results file (0.474, 0.377, 0.249, 0.216) shown as percentages. Source: bar-ai-guidance/analysis/hallucination-sanctions/results/results.md

The "ratchet" does appear in the size of penalties that are imposed. The median rose from $1,750 to $2,731 and the share of amounts of $10,000 or more from 12.5% to 22.4%. The 75th percentile rose in each half-year. Total penalties also grow, because the number of decisions grows: lawyer decisions with money went from 18 in 2025H1 to 60 in 2026H1, even as the share fell.

Pre-registered predictions, graded

Prediction2025H12026H1Verdict
P1Money-penalty share not more than 5 points higher in 2026H147.4% (18/38)24.9% (60/241)Held
P2Median USD amount higher in 2026H1$1,750 (16 amounts)$2,731 (58 amounts)Held (p = 0.124)
P3Professional-sanction share higher in 2026H115.8% (6/38)14.9% (36/241)Failed
P4Pro se decisions fined less often than lawyer decisions in each half-year1.7% vs 47.4%6.8% vs 24.9%Held
P5Fewer than half of 2026H1 lawyer decisions carry money or a professional sanction31.1% (75/241)Held
Verdicts follow the pre-stated direction or threshold, not the p-values. Source: bar-ai-guidance/analysis/hallucination-sanctions.md and results/results.md.

Why the small early group matters

The database grew very fast. It holds 38 US lawyer decisions from the first half of 2025 and 241 from the first half of 2026. Only 16 of the early decisions have a dollar amount, just above our pre-registered minimum of 10. A collection that starts by catching mostly dramatic, sanctioned cases and later catches routine warnings and show-cause orders too would show a falling share even if no court changed its behaviour. So the result describes the database, not the courts.

US decisions involving a lawyer in the database, by half-year

Counts of US decisions labelled as involving a lawyer, by the database's decision date. Source: bar-ai-guidance/analysis/hallucination-sanctions/results/results.md

We also checked the dates themselves. Reading the hosted documents for the 602 lawyer decisions, the recorded date was printed in 510 of 526 readable documents (97.0%). Correcting dates to what the documents say moved at most 10 decisions into a different half-year, and none of the five verdicts changed. Under both correction rules, the professional-sanction prediction fails by a wider margin, not a narrower one.

Self-represented litigants are fined far less often than lawyers, though their share rose slowly, from 1.7% in 2025H1 to 6.8% in 2026H1. Their 2026H1 median penalty was $500.

What it means

The evidence does not support a simple story of courts growing harsher with each case. In this database, any one lawyer decision is less likely to carry money now than in early 2025, while the penalties that are imposed have grown somewhat. Anyone citing a "ratchet" should say which measure they mean. The same repository also keeps an index of 62 primary-source documents in which US bars, courts and legislatures tell lawyers how to use generative AI, including Connecticut's 2026 court rules, which make failing to verify AI-produced citations a ground for sanctions.

More research