Stanford: AI in 38 per cent of Santa Clara’s 2026 complaints, excluding debt cases
Stanford researchers estimate that in 2026 about 38 per cent of non-debt complaints in Santa Clara and 36 per cent in San Francisco appear to have been drafted with AI, compared with 16–17 per cent in two Oklahoma counties. The figure describes a share of the complaints studied, not individual documents.
Stanford University researchers published a study on 19 November in which they estimate AI use in 127,518 complaints. The complaints were filed in 2020–2026 in trial courts in San Francisco, Santa Clara, Oklahoma and Tulsa counties. The estimate is near zero before 2023 and rises sharply after that.
The result does not rest on identifying individual documents. An AI detector gives each text a score. In legal writing, about a quarter of tokens are boilerplate, which weakens the detector’s signal, the researchers said. When one detector’s settings were varied within reasonable limits, the share of complaints flagged as AI ranged from under 5 per cent to almost 50 per cent.
The researchers therefore built a reference dataset called PhantomDocket. It contains 980 filings in which a judge had found citations invented by AI. The cases come from a database maintained by Damien Charlotin. The researchers’ ScoreMix method compares the distribution of EditLens detector scores against these filings and the court’s own complaints filed before ChatGPT. The method estimates what share of complaints have scores resembling those of the hallucination cases. It does not show how many complaints contain a hallucination. For 2022 complaints, the method produced estimates of at most 0.5 per cent in every court.
The table shows the 2026 estimates by court:
| County | Estimated share 2026 | 95% confidence interval |
|---|---|---|
| Santa Clara | 0.378 | 0.344–0.412 |
| San Francisco | 0.356 | 0.315–0.391 |
| Tulsa | 0.169 | 0.144–0.192 |
| Oklahoma | 0.161 | 0.143–0.181 |
Source: Arifov et al., ScoreMix and EditLens; a share of 0.378 means 37.8 per cent. The researchers excluded debt collection and debt cases from the estimates.
The table does not allow the conclusion that the difference is down to the state. The types of cases collected varied by court, and the mix of case types likely explains part of the difference, the researchers said. Detector settings still affect the result. In Santa Clara, the choice of threshold alone produced estimates of 4–49 per cent, whereas the ScoreMix range across settings was 29–47 per cent.
In Santa Clara, the estimated share of AI-drafted complaints among those filed without a lawyer rose from 19 per cent in 2024 to 72 per cent in 2026. Among complaints filed by lawyers, it rose from 5 per cent to 34 per cent. Other courts had too few complaints filed without a lawyer for an estimate.
The estimates may be too low, the researchers said, because the reference dataset represents AI use that produced hallucinations, and AI use does not always produce them. The AI cases in the reference dataset are also mostly briefs and motions rather than complaints. The researchers urge great caution when detectors are used to draw conclusions about individual documents, because even the best detectors can have a significant share of false positives and missed AI text.
A new California law will require lawyers to disclose AI use from January, KQED reported. The requirement does not apply to people representing themselves.