argument: Notizie/News - AI in Judicial Activities
Source: Stanford HAI
Researchers affiliated with Stanford HAI and RegLab examined complaints from four state courts to estimate the prevalence of AI-assisted drafting and assess detection reliability. The study reports calibrated estimates for 2026 of 38% in Santa Clara County and 36% in San Francisco County, California, excluding debt-collection cases. For comparable cases in Oklahoma, the estimated proportions were 17% in Tulsa County and 16% in Oklahoma County. These figures estimate prevalence across groups of filings; the researchers caution that detection cannot reliably establish which individual documents used AI. In Santa Clara, estimated signs of AI assistance in self-represented complaints increased from 19% in 2024 to 72% in 2026, compared with 5% to 34% for attorney-filed complaints. The researchers found the highest estimated use in California employment and business complaints, with little detected use in debt-collection complaints, which commonly follow templates.
Legal writing presents detection difficulties because specialized vocabulary and recurring boilerplate can obscure signals associated with AI-generated text. One detector classified between 10% and 98% of the same 2026 Santa Clara complaints as AI-written when its settings changed, demonstrating substantial methodological sensitivity. The researchers developed PhantomDocket using documented fabricated-citation cases from Damien Charlotin’s database alongside selected controls. Their ScoreMix method calibrates detector outputs to estimate how frequently filings resemble the hallucination examples more closely than documents written before ChatGPT. Even the strongest detectors retained substantial false-positive and false-negative rates, limiting their suitability for judgments about individual filings. The researchers question whether AI improves access to legal assistance or burdens courts with deficient work, and advocate stronger disclosure policies to support accountability and measurement.