A civil case in Hangzhou closes in about twenty minutes. The parties never enter a building; an intelligent platform sorts the filings, surfaces the key questions, and drafts most of the judgment. A human judge reviews the text, signs it, and sends it to both sides.
The internet court blueprint
China opened the world's first internet court in Hangzhou in August 2017, followed by Beijing and Guangzhou in 2018. The model is fully digital: online filing, blockchain-anchored evidence, remote hearings, and electronic service replace the physical courthouse for a large class of online-dispute cases.
Similar experiments exist elsewhere—Singapore, South Korea, and Australia all pilot AI in tribunals—but China's version is unusually centralized and scaled, built as a national template rather than scattered local trials.
The numbers show how far the format has spread:
- In 2023, the three internet courts handled 70,635 cases.
- Online filing reached 98.05%, remote hearings 99.71%, and electronic service 88.18% of applicable matters.
For a judiciary processing a flood of e-commerce and platform disputes, the appeal is simple: keep pace with the volume without multiplying headcount.
The "digital court" in daily use
The AI layer is not only for internet-specific cases. At the Kunshan People's Court in Jiangsu, an AI assistant automatically sorts litigation files, summarizes case elements, and highlights the central legal questions—such as whether a property manager met its duty of care. After the hearing, the judge generated more than 70% of the judgment draft with a single click.
Across the system, judges use AI tools for legal research, document drafting, and consistency checks. The consistent stance from the courts is that the machine supports the human; it does not decide.
Province-scale trial platforms
The most ambitious example is "Phoenix" (凤凰), an intelligent trial platform built by a Zhejiang University team:
- It has been rolled out across 55 pilot courts in Zhejiang and has handled over 15,000 cases, with reported judicial efficiency gains near 40%.
- It now reaches more than 10,000 courts nationwide.
- The team open-sourced the LegalMind legal large model (大模型) and a voice-synthesis component, with monthly usage reported above 1.5 million calls.
- LegalMind was pretrained on 3,000+ Chinese laws and 50,000+ regulations, and was built with input from serving judges who mapped the reasoning logic of civil cases.
This is the pattern worth noting: a legal large model (大模型), openly released, embedded into everyday case handling rather than sold as a standalone product.
A profession adapting
The demand side is moving too. China had 830,000 lawyers and 45,000 law firms as of September 2025, according to the Ministry of Justice—a large, fragmented market where junior routine work dominates.
A 2025 survey of legal professionals found that 96% already use generative AI tools in daily work, though only 28% reported firm-wide regular deployment, and 47% said their firm planned an AI strategy within two years. Market research estimates China's AI legal-services market at roughly US$186 million in 2025, led by natural-language and generative-AI tooling.
The gap between near-universal personal use and only 28% firm-wide deployment is the familiar pattern seen in other markets: individuals adopt faster than institutions can govern. The open question is whether that lag protects quality or simply delays inevitable reorganization.
Why the centralized model scales
The reason China's version looks different from scattered Western pilots is centralization. Internet courts, the Kunshan assistant, and the Phoenix (凤凰) platform were built as templates meant to propagate, not as one-off experiments — so a legal large model (大模型) trained once can be deployed across thousands of courts rather than re-implemented per jurisdiction. That top-down distribution is what produces both the efficiency gains and the concern: a single model's blind spot, once embedded nationwide, repeats at scale rather than staying local. For other judiciaries weighing AI assistance, the Chinese case is a natural experiment in standardization versus local control. The efficiency argument is hard to dismiss when caseloads outrun headcount, yet the accountability question grows with centralization — the more a draft is machine-produced and uniformly applied, the more important it becomes that any error is catchable, appealable, and traceable to a human. Speed without that loop is the risk the system still has to prove it has contained.
Honest limitations
- One-click drafts invite over-reliance. When 70% of a judgment is machine-generated, the risk is "template justice"—plausible text with thin reasoning that a pressurized judge signs without full review.
- Hallucination is a live hazard. Fabricated citations or misread precedents are hard to catch without manual verification, and the cost of a miss is borne by a litigant.
- Deployment is uneven. Internet-savvy and wealthier provinces lead; rural and lower-resource courts lag, widening a digital justice gap.
- Transparency is thin. There are few public benchmarks on how accurate AI recommendations are, or how appeals resolve when a draft is contested.
- Court data is sensitive. The records used to train and run these models carry privacy and due-process weight, yet retention, access, and audit-trail practices vary across jurisdictions and are rarely published.
What readers can do now
- If you litigate in China, expect an AI-assisted process; ask your lawyer specifically how filings and drafts are checked for errors before they are filed.
- If you build legal tech, the proven wedge is verifiable, standardized work—drafting, retrieval, scheduling—not outcome prediction, which courts and clients should treat as off-limits.
- If you study the system, judge it by auditability and appeal outcomes, not by speed metrics alone.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the smart-court push proves AI earns trust in justice only when every machine draft remains fully reversible by a human judge.