Stanford finds AI text in 38% of Santa Clara court complaints
The same detector marked from 10% to 98% of the same complaints as AI text, with different adjustments.
Stanford researchers studied complaints in 4 US state courts, in a new paper on how AI text enters the courts. They found signs of AI drafts in 38% of 2026 complaints in Santa Clara County, California, up from near zero in 2022. The team says detectors can estimate the share of complaints with AI, but cannot find which one complaint used AI.
How it works
The team made a test set called PhantomDocket from court records in which judges found fake citations, plus control complaints. A method called ScoreMix adjusts detector results. It reports the share of documents that are almost the same as the documents with fake citations, more than documents from before ChatGPT.
The numbers
- Oklahoma counties: 17% and 16%, below the California counties.
- Persons without a lawyer in Santa Clara: 19% in 2024, 72% in 2026.
- Complaints from lawyers there: 34% in 2026.
What is not known
All detectors give some incorrect alarms and miss some AI text. A detector that is not made for law can give incorrect results on legal text. The team does not know if AI helps persons who cannot pay a lawyer, or fills the courts with defective work.
Insight: Invented facts are the one sure sign of AI in court
The Stanford test set uses court records in which judges found fake citations. In October 2026, Artificial Analysis said that more than 60% of correct legal AI results held a material hallucination, an invented fact that can mislead. The one sure sign of AI is an invented fact, and legal AI tools make them too, which points to disclosure rules, not detectors.
Sources
Posted