Shenzhen's AI court assistant has handled 600,000 cases

Share
Shenzhen's AI court assistant has handled 600,000 cases

Shenzhen put an AI assistant inside its courtrooms two years ago. It now has numbers to show for it — and foreign judges are booking flights to see the thing.


Shenzhen's Intermediate People's Court says its AI-assisted adjudication system has helped handle more than 600,000 cases since going live on 28 June 2024, describing it as the world's first complete AI-assisted adjudication system and the first to span civil, administrative and criminal proceedings. It is not one model but four modules threaded through the case workflow: filing review, file reading, courtroom support and judgment drafting. The filing module alone, live since January, has assisted more than 310,000 filings and prompted parties to supply missing materials over 9,000 times. Across the 36 working days after the full launch, the court counted 57,052 assisted document readings, 23,701 assisted hearings and 52,554 generated judgment drafts.

The operational numbers are where the case gets made. In July, Shenzhen's courts closed 37,000 civil and commercial cases — up 73.9 percent year on year — with average time to closure down 38.54 days and the backlog down 29.3 percent. Judges still decide: every node in the workflow requires the judge to tick "review," "confirm" or "decide," a design the court says exists specifically to prevent machine adjudication. The system was built in-house over ten months on a dedicated-network allocation of 70 petaflops of sparse compute, trained on a corpus the court describes as trillions of Chinese characters drawn from statutes, the national case database and electronic case files. Shenzhen is not running a private experiment — it is the pilot for the Supreme People's Court's plan to run the country's courts on a single network, under a 2022 SPC opinion on regulating AI in judicial work.

The framing deserves scrutiny, and the court invites it: "world's first" is a claim about completeness, not capability. Read the modules closely and this is an unusually aggressive case-management stack — classify, correct, cross-check, draft — with the judge as signatory rather than operator. That is the honest version of AI in courts today, and it is also the version that raises the interesting question. If the model drafts a judgment and the judge approves 95 percent of what it sees, the human in the loop is real but thin. The court's tick-box design is a guard against that; whether it is enough is a question no press tour answers. Foreign delegations at the China Law International Forum in Shenzhen this week — including judges from Samoa, Greece and Russia — came away wanting to build their own. They should ask how often the tick-box says no.

One number worth reading carefully: Shenzhen's two-tier courts take in more than 700,000 cases a year, which is a caseload figure, not an AI-assisted count. The official assisted total is the 600,000-plus above.

What to watch: whether the Supreme People's Court publishes an independent evaluation of draft acceptance rates, and whether any court outside China adopts this stack.

Would you want an AI drafting the first version of a judgment in your case? Tell us in the comments.

Read more

Byteification turns Qwen3 and Llama 3 into byte-level models

Byteification turns Qwen3 and Llama 3 into byte-level models

The tokenizer is the last piece of the language-model pipeline that was never learned end to end — and a Nature paper out this week shows it can be swapped out for less than one percent of a pretraining budget. Researchers have retrofitted four open subword language models into true byte-level models using less than 1% of a typical pretraining budget, and the result — published 7 October in Nature — closes the performance gap that has kept byte-level models a niche curiosity for nearly a decad

OpenAI's dots go mobile and can start your Codex work

OpenAI's dots go mobile and can start your Codex work

One product update and two benchmark papers this hour, all circling the same question: how much of the work can you actually hand over, and what happens to the part you didn't? OpenAI's dots — the always-on agents that get their own cloud computer — can now be created and configured entirely from the ChatGPT app on iOS and Android, and a dot can start Codex work or pick up an existing Codex thread on your behalf. The October 9 release, billed as the first of a fresh run of updates for dots, mo

Coding agents as a hospital: Cockroach's five-month results

Coding agents as a hospital: Cockroach's five-month results

The agentic-coding debate keeps being framed as speed versus trust. Cockroach Labs just published the most detailed trust-first ledger yet — five months of numbers, including the parts that didn't work. Cockroach Labs ran its codebase like a teaching hospital for five months and says the pipeline merged well over a million lines of code with seven total reverts. In the setup the company calls MOLT Sinai, GitHub issues are patients, merging is discharge, and the humans in charge are Chiefs of M

Today in AI — October 10, 2026

Today in AI — October 10, 2026

A day where the boring machinery took center stage: proof checkers, order books, warehouses and the humble text message — plus fresh arguments about who gets to declare any of it safe. Models & Research * Mathematicians get a reliability primer for the Lean Theorem Prover. A guest post by Thomas Hales on Terence Tao's blog walks through what Lean actually guarantees — and devotes itself to the "Summer of Soundness Bugs," the string of kernel bugs found this summer that let false proofs thr