What remains unsettled is how precisely any single filing can be classified. The researchers found that AI detectors are unreliable for judging individual documents: one tool flagged anywhere from 10% to 98% of the same set of 2026 Santa Clara complaints as AI-written, depending on settings. The team therefore presents its figures as population-level estimates, not evidence about any particular filer. To address the problem, they built a calibration method called ScoreMix and a benchmark drawn from court filings in which judges had flagged fabricated citations. The authors argue that general-purpose detectors handle legal boilerplate poorly, which is why a legal-specific approach is needed.
The trend tracks the spread of generative AI. Estimated AI-assisted drafting in Santa Clara rose from near zero in 2022 to more than a third of examined complaints in 2026. Stanford law professor Dan Ho, a co-author, argues that banning AI use is unrealistic and that AI "could improve the quality of complaints, relative to not having much access to legal knowledge at all." Riana Pfefferkorn, a Stanford HAI policy fellow who was not involved in the study, cautioned that "The mere use of AI in drafting a document doesn't imply anything about quality."
Courts are deciding how to handle AI-drafted filings without a reliable way to identify them, which makes the study's limits as important as its numbers. Attorneys in these jurisdictions should expect judges to scrutinize citations more closely, particularly in self-represented matters, and should verify every authority in their own filings. The findings were released November 19, 2026, and were summarized by Law360, with KQED and the ABA Journal reporting the self-represented and attorney-filed breakdowns.