Stanford study warns against using AI detectors to police individual court filings

Stanford researchers reported on 19 November 2026 that signs of AI-assisted drafting appeared in more than a third of the non-debt-collection complaints they examined in two California county courts—but warned that detection tools remain unreliable for identifying individual offenders.

Their calibrated estimates put AI use at 38% in Santa Clara and 36% in San Francisco, compared with 17% and 16% in two Oklahoma counties. In Santa Clara, estimated use was substantially higher among self-represented litigants than lawyers. These are statistical estimates, not verified disclosures by the people who prepared each filing. 

The researchers developed a benchmark using documented AI-hallucination cases and controls. They found that detector settings could produce dramatically different results: one detector flagged between 10% and 98% of the same Santa Clara complaints. False positives and false negatives remained significant even with stronger methods. 

The consequential finding is therefore not simply that AI use is growing. A tool suitable for estimating population trends may be unsuitable evidence that a particular lawyer or litigant used AI.

For Australian courts and practitioners, the implication is procedural: scrutiny should focus on demonstrable defects, source verification and reliable records of preparation—not stylistic suspicion or an unexplained detector score. The researchers favour better disclosure policies to improve accountability and measurement.

Source: Stanford HAI research announcement⁠

Leave a Reply

Your email address will not be published. Required fields are marked *