--- title: "Shenzhen's AI-Assisted Court System — 35 Million Tasks, Judges Still Decide" date: 2026-10-23 category: AI in Action site: NeuroAI canonical: https://neuroai.site/a/na-app-shenzhen-ai-court language: en --- # Shenzhen's AI-Assisted Court System — 35 Million Tasks, Judges Still Decide > Shenzhen runs the world's first full-process AI-assisted adjudication system, lifting judges out of drudge work while keeping human verdicts at the center. A judge in Shenzhen used to open a new file and face hundreds of pages of filings, scattered evidence, and a ticking clock. Today a model pre-reads the dossier, flags missing proof, and drafts a first-pass judgment — but the signature at the bottom is still human. The question hanging over courtrooms everywhere is no longer whether AI can help, but exactly where the line is drawn. ## What the system actually does The platform is not a robot judge. It is a set of assistants wired into six stages of a case, each one requiring a human click before anything moves forward: - **Filing review (立案审查):** checks materials against 35 filing standards and tells a litigant what to fix in one go. - **Dossier reading (阅卷核验):** auto-catalogs evidence and extracts claims, with every line traceable back to the original page. - **Courtroom support (庭审辅助):** builds a hearing outline and suggests "scripted" questioning on consistency and logic. - **Case analysis (案情研判):** helps frame the facts and dispute points; once the judge decides, the model drafts the reasoning. - **Drafting (文书草拟):** produces a first version of the judgment, then re-reads it from the opponent's angle to poke holes. - **Knowledge retrieval (知识检索):** pushes the relevant statutes and similar cases to the judge's screen. Every node carries a confirm-or-reject button. The system cannot "generate a conclusion with one click" and close a case on its own. ## The numbers behind the rollout Shenzhen courts went live with what they describe as the world's first systematic, full-process AI-assisted adjudication system (人工智能辅助审判系统) in June 2024, covering civil, criminal, and administrative cases — now more than 90% of case types. The white paper *AI in the Judiciary — The Shenzhen Practice (人工智能司法应用的深圳实践)*, released at the China Law Forum (中国法治国际论坛) on September 22–23, 2026, laid out the scale: - As of August 31, 2026, the models and agents had completed **35.451 million** tasks. - They assisted **609,000** dossier reviews, **269,000** court sessions, and **678,000** document drafts. - More than **700,000** cases were touched, and legal knowledge was pushed to judges **22.746 million** times. - In 2025, Shenzhen courts closed **773,828** cases; the average judge closed **744** cases, up **249** year-on-year, with monthly closure per judge **60%** higher** than before the system. - The white paper reports appeal rates and remand/reversal rates fell by **48%** and **65%** respectively. One detail worth noting: the 2026 budget of the Shenzhen Intermediate People's Court lists a software-development line item of **220,200 RMB (≈ US$30,600 / HK$239K)** for the AI-assisted adjudication application project — a hint that the public-facing rollout rides on existing judicial-intranet infrastructure rather than a single big-ticket build. ## Where the human stays in charge The design choices are the real story for other jurisdictions: - **Judge-led, machine-assisted.** The Supreme People's Court has stated plainly that AI must never replace a judge; the model's output is only a reference, and the verdict stays with the adjudicating body. - **Physically isolated.** The model runs on the judicial intranet (政法专网), cut off from the public internet, to blunt data-leak and tampering risk. - **Sourced answers.** Legal content is pulled from an authoritative knowledge base with citations, to fight "hallucinated" statutes. - **A human veto.** Shenzhen assigns a dedicated security officer with an independent review power and a one-vote kill switch. Hong Kong University law scholar Liu Zhuang, who has studied the deployment, calls Shenzhen the earliest court system anywhere to apply large language models to real adjudication work in a systematic way. ## Trend reading The Shenzhen case shows a template that travels: pick the repetitive, rules-bound steps; keep a human at every decision point; isolate the model from the open internet; and make the judge the one who signs. That is a more useful export than any accuracy number, because it answers the liability question first and the efficiency question second. ## Honest limitations - **The quality gains are not AI-only.** The white paper itself notes the improvements came alongside management reforms and specialization — so the 60% efficiency and 48%/65% drop figures should not be read as a pure technology dividend. - **The comparison baseline is unclear.** The white paper states the 48% and 65% declines without specifying the year or period they are measured against, so the exact size of the effect is hard to pin down. - **Coverage is uneven.** The system is mature in Shenzhen's two court tiers; other Chinese courts are only beginning pilots, and rural or smaller jurisdictions lag. - **Residual risk remains.** Hallucination and bias are managed, not eliminated, and the system depends on clean, well-scanned input files. - **Cost is understated.** The public budget line covers only one slice of a much larger, multi-year effort including compute and training. ## What readers can do now - **If you have a matter in Shenzhen,** expect faster procedural updates and clearer filing feedback — use the 12368-style intelligent service lines rather than waiting for a call-back. - **If you are a legal professional,** learn the system's six scenes and 85 business flows so you can pre-empt what the model will surface and focus your argument on the contested points. - **If you set policy abroad,** borrow the governance shell first — human-in-the-loop plus intranet isolation — before importing any specific model. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-app-shenzhen-ai-court Free to quote with attribution and a link to the original.