Shenzhen's AI-Assisted Trial System Lifts Judicial Caseload 50% — And a Peer-Reviewed Study Flags a Sentencing-Bias Risk

In June 2024, the Shenzhen Intermediate People's Court launched what it describes as mainland China's first large-scale AI-assisted trial system, an in-house large language model fine-tuned on roughly two trillion Chinese characters of law, judgments and legal scholarship.

The system was built to assist judges across the city's two-tier court system by summarising case facts and disputed issues, generating hearing prompts, and drafting judgment reasoning based on the judge's own preliminary decision, covering 85 procedures across civil, administrative and criminal litigation.

Judges are required to review and revise every AI-drafted section before signing off; researchers who interviewed roughly 20 judges and 5 technical developers between July and November 2024 combined field observation with a controlled vignette experiment involving 139 judges, published in the Journal of Legal Analysis (Oxford University Press).

Court-reported figures show judges' average annual caseload rising from about 495 cases in 2024 to 744 in 2025 — an increase of 249 cases (about 50%) and 261 more than the Guangdong provincial average. The peer-reviewed vignette experiment found no significant AI effect on verdicts for an extralegal bias-testing variable, but AI-assisted judges gave significantly longer sentences to a treated defendant group than unassisted judges (13.83 vs. 6.59 months, p=.01).

The court's vice-president said the rollout 'provided a practical answer' on using AI in judicial work, and the system is reported to be extending to courts in other Chinese cities, but the study authors describe the sentencing-length finding as a risk of 'amplifying' pre-existing bias, cautioning it could also be a chance result requiring further research; the court has not disclosed what share of judgments include AI-generated content.