AI 可能正大量進入美國州法院,但要證明並不容易
AI Might Be Flooding State Court Houses, but Proving It Is Another Matter
Stanford HAI 與 RegLab 研究人員分析四個美國縣法院的申訴文件,發現 2026 年加州 Santa Clara 和 San Francisco 的非債務追收案件中,AI 使用比例分別升至 38% 和 36%。
AI Might Be Flooding State Court Houses, but Proving It Is Another Matter
Date
November 19, 2026
Topics
Regulation, Policy, GovernanceGenerative AI

Stanford researchers use a novel benchmark built from documented AI hallucination cases to calibrate detectors and track generative AI’s rise across four U.S. state courts.
Since the release of ChatGPT, AI-generated text has been seeping through our society: Its telltale signs have been spotted in news articles, political speeches, and even UN press releases. Many observers have been pointing to the courts as an obvious – and vulnerable – institution: Indeed, over the past three years, fake citations have been showing up in court filings – by lawyers, district attorneys, and even judges. A wave of AI-generated court cases could overwhelm courts, waste court time and resources, and ultimately deny or delay vulnerable individuals’ access to justice.
But how ubiquitous is AI assistance in the courts? And how can we rigorously detect AI in court filings? Researchers with the Stanford Institute for Human-Centered AI and the Regulation, Evaluation, and Governance Lab (RegLab) tackled this question by collecting and analyzing complaints from county courts. We found signs of AI-assisted drafting rose from near zero in 2022 to more than a third of examined Santa Clara complaints in 2026. We also showed how the same detectors can give starkly different estimates with different settings.
Why Detecting AI Is Hard
The science of AI detection is far from perfect, especially in law. In our new paper, Detection Objection: Documenting, Calibrating, and Detecting the Rise of AI-Assisted Text in U.S. State Courts, we show the unique challenges.
AI detectors work by looking for AI’s telltale signs (e.g., “too many em-dashes”): In different domains where language is used differently, these signs may simply show up differently. Indeed, legal writing is distinct: Law uses a narrower vocabulary than general speech and repeated boilerplate patterns, which leave less signal for an AI detector. The training data for commercial detectors also may not have included legal writing.
Put all of these together, and we find that detection rates (1) are especially poor in professional domains, and (2) can be highly sensitive to modeling choices. For example, we find that an AI detector flagged anywhere from 10% to 98% of the same set of 2026 Santa Clara complaints as AI-written, depending on the settings used. This is a cautionary tale for all AI-detection work in different subdomains.
Building Ground Truth
Our breakthrough came when we looked at an area of particularly egregious AI violations: hallucinations and fabricated case lookups. These are cases where we know AI must have been used, because the mistakes are simply so bad. And in fact, this is already an area of intense scrutiny: the Charlotin AI Hallucinations Cases Database wasestablished in April 2025 by legal researcher Damien Charlotin specifically to track cases where judges accuse or admonish parties for using fictitious citations.
These cases, we noticed, were rife with novel telltale signs. Existing AI detectors like Pangram or EditLens both failed to automatically detect AI or needed extensive tuning to function. We put Charlotin’s database together with a carefully crafted set of controls into a new benchmark for AI detection in the courts, called PhantomDocket.
In addition, we developed a method, called ScoreMix, to calibrate the outputs of AI detectors and report the prevalence of filings that are more similar to the hallucinated filings than to those written before ChatGPT. Together, these gave us calibrated, robust, and conservative estimates of AI usage in court filings.
Key Findings: A Third of Cases
Using these techniques, we were able to paint a rich portrait of how generative AI has entered four state courts.
By 2026, the prevalence of AI usage for Santa Clara and San Francisco counties (California) has risen to 38% and 36%, respectively (in non-debt-collection cases).
The prevalence varies across courts. Tulsa and Oklahoma counties (Oklahoma) had noticeably lower AI utilization at 17% and 16%, respectively, for comparable cases.
Pro se litigants, or people who represent themselves at court, show higher adoption rates. In Santa Clara, the share of self-represented complaints showing signs of AI use went from 19% in 2024 to 72% in 2026. In contrast, the share of attorney-filed complaints went from 5% to 34%.
AI utilization varies by case type. Employment and business complaints in California show the highest rates, but there is almost no AI utilization in debt collection complaints, which are templated and comprise nearly half of the filings in some courts.
AI use also varies by section of a court filing, more often appearing in the statement of facts and causes of action.
What the Data Can – and Can’t – Tell Us
AI-assisted text is creeping into the courts, at higher rates for pro se litigants and at different rates across jurisdictions.
Based on our paper, we caution that AI detection is most reliable for estimating population prevalence, not pinpointing which specific filings used AI. In our benchmark dataset, we show that the false positive and false negative rates remain nontrivial, even for the best detectors, and that using off-the-shelf detectors is risky in professional domains with distinct language usage.
Our research also raises real questions about whether AI is helping litigants who cannot afford attorneys or flooding the courts with inadequate lawyering. To date, over 1,500 U.S. cases have been documented to include legal hallucinations. This evidence underscores the need for the judiciary to grapple with the rise of AI in the legal profession. We argue for better disclosure policies across courts, allowing for better accountability and measurement of generative AI assistance in the future.
Javokhir Arifov is a Research Fellow at Stanford RegLab.
Alexander Spangher is a Postdoctoral Scholar at Stanford University, affiliated with RegLab, Stanford Artificial Intelligence Laboratory (SAIL), and Stanford HAI.
Jonathan Hennessy is a Senior Research Manager at Stanford RegLab.
Diya Karan is a Research Assistant at Stanford RegLab and a student at Stanford University.
Emaan Hariri was a Law Student Fellow at Stanford RegLab and received his J.D. from Stanford Law School.
Daniel E. Ho is the William Benjamin Scott and Luna M. Scott Professor of Law, professor of political science and (by courtesy) of computer science, senior fellow and associate director at Stanford HAI, senior fellow at the Stanford Institute for Economic Policy Research (SIEPR), and director of the Stanford Regulation, Evaluation, and Governance Lab (RegLab).
This work was supported in part by Stanford HAI and Stanford Impact Labs. The authors thank Damien Charlotin for maintaining the AI Hallucination Cases Database, and Pangram Labs for providing research access to their models. Pangram Labs had no role in the study design, analysis, or this work.
Share
Link copied to clipboard!
Contributor(s)
Javokhir Arifov, Alexander Spangher, Jonathan Hennessy, Diya Karan, Emaan Hariri, and Daniel E. Ho
Related News
Stanford HAI Backs Ambitious AI Projects with 2026 Hoffman-Yee Grants
Nikki Goth Itoi
Nov 16, 2026
News
Seven interdisciplinary research teams win a total of $3.5 million to pursue bold human-centered approaches to AI across medicine, science, law and education.
News
Stanford HAI Backs Ambitious AI Projects with 2026 Hoffman-Yee Grants
Nikki Goth Itoi
Economy, MarketsGenerative AISciences (Social, Health, Biological, Physical)Workforce, LaborNov 16
Seven interdisciplinary research teams win a total of $3.5 million to pursue bold human-centered approaches to AI across medicine, science, law and education.
The Jagged Edge of the AI Boom
Nov 15, 2026
News
New findings from the Stanford AI Index reveal a more uneven picture of public sentiment and labor market outcomes than previously thought, as AI investment surpasses $1 trillion.
News
The Jagged Edge of the AI Boom
Economy, MarketsRegulation, Policy, GovernanceNov 15
New findings from the Stanford AI Index reveal a more uneven picture of public sentiment and labor market outcomes than previously thought, as AI investment surpasses $1 trillion.
The World Has Changed: Insights From The New AI100 Report
Nov 10, 2026
News

From battlefield targeting to synthetic media, the third AI100 report finds a technology that has outrun both expectations and governance.
News

The World Has Changed: Insights From The New AI100 Report
Economy, MarketsEducation, SkillsEthics, Equity, InclusionFinance, BusinessGenerative AIInternational Affairs, International Security, International DevelopmentRegulation, Policy, GovernanceNov 10
From battlefield targeting to synthetic media, the third AI100 report finds a technology that has outrun both expectations and governance.
Navigate
Participate
Stay Up To Date
Get the latest news, advances in research, policy work, and education program updates from HAI in your inbox weekly.
© Stanford University. Stanford, California 94305.
Pause Media
AI Might Be Flooding State Court Houses, but Proving It Is Another Matter | Stanford HAI
來源:Stanford HAI News · hai.stanford.edu