Stanford: Over a Third of California Complaints Show AI Traces

AI policy illustration: scales on a circuit base with a red weight

Stanford's RegLab published a study on November 19 that measures traces of AI-assisted writing in complaints filed at four U.S. county courts. In 2026 non-debt-collection complaints, the share carrying AI signals was 38% in California's Santa Clara County and 36% in San Francisco County. In Oklahoma, Tulsa County and Oklahoma County came in at 17% and 16%.

In Santa Clara County this indicator was still close to zero in 2022. The fastest growth came from self-represented litigants, who filed without a lawyer: their share rose from 19% in 2024 to 72% in 2026, while filings submitted by attorneys climbed from 5% to 34%. By case type, employment and commercial disputes in California showed the highest shares. Debt-collection complaints, which make up nearly half of the docket in some courts, showed almost no sign of AI. The AI signals appeared mostly in the statement of facts and the causes of action.

Using hallucination cases as the answer key

The hard part is proving it. The team built on a database maintained by Damien Charlotin, which since April 2025 has logged cases where judges accused or admonished litigants over fabricated citations; more than 1,500 such U.S. cases are recorded. The errors in those filings are obvious enough to establish that AI was used, which makes them suitable as ground truth.

The researchers combined these filings with a carefully selected control sample into a new benchmark called PhantomDocket, and proposed a calibration method called ScoreMix to estimate how many complaints look more like these hallucination-tainted filings than like documents written before ChatGPT. Off-the-shelf detectors Pangram and EditLens did not work well when used as-is: they either failed to detect anything or needed heavy tuning. Pangram Labs provided access for the study; the research team says the company played no part in the study design or analysis.

"AI detection is most reliable for estimating population prevalence, not pinpointing which specific filings used AI."

Legal writing uses a narrow vocabulary and plenty of boilerplate, which leaves detectors few signals to work with, and commercial detectors may not have legal text in their training data. Even the best detectors have notable false-positive and false-negative rates. The study announcement does not publish PhantomDocket's sample size or a table of detector accuracy; those are in the full paper. The team's policy recommendation is that courts adopt AI-use disclosure rules, so that future statistics and accountability rest on firm ground.

How this looks for Chinese courts

The closest Chinese counterpart is the Supreme People's Court's Opinions on Regulating and Strengthening the Judicial Application of Artificial Intelligence, issued in December 2022. It mainly governs how courts themselves use AI to assist case handling, and states that AI must not replace judges in deciding cases. Whether litigants and lawyers must declare it when a complaint or a statement of representation was drafted with generative AI is not specifically addressed in the publicly available text.

In the U.S., some federal judges already require lawyers to state in writing whether they used generative AI and to check every citation personally. The Stanford study's warning is that where no disclosure is required, outsiders can only estimate an overall share with detectors, and the margin of error is considerable. For the growing number of litigants and law firms in China drafting complaints with large language models, the risk of fabricated citations is the same; what happens once a judge spots one currently depends on how each court handles individual cases.

Sources: Stanford HAI, Stanford RegLab paper, CocoLoop, Supreme People's Court of China; checked the AI-signal shares at the four county courts, the shift between self-filed and lawyer-filed documents, the size of the hallucination-case database, and the range of detection results.