Research update — 22 September 2026: Added the Supreme People’s Court’s confirmation that China’s unified national case-and-office platform is now online in trial operation, with end-to-end electronic files advancing the infrastructure needed for judicial AI. This does not establish routine majority use of the forecast’s four AI functions; the forecast remains at 84%.
The second mind in a Chinese courtroom will live inside the court network. It will open the electronic file, arrange the evidence, identify the disputed facts, find the controlling law, compare similar cases and compose the first draft of the judgment. The human judge will question that work, change it and carry responsibility for the result. The machine will still have read the case first.
This future has escaped the laboratory. Shenzhen’s artificial-intelligence system already accompanies a case from filing to judgment. Suzhou’s Future Judge Assistant reads entire files and produces legal documents. Beijing’s court model generates proposed legal positions and similar-case reports. In March 2026, Supreme People’s Court president Zhang Jun said a combined intelligent adjudication system—智审系统, or zhìshěn xìtǒng—was being tested in 23 courts across 11 provincial-level regions. Shenzhen’s AI-assisted adjudication system Suzhou’s Future Judge Assistant Beijing’s digital adjudication platform Report on the 2026 national trial
China’s official destination is equally explicit. The Supreme People’s Court ordered a substantially complete judicial-AI application system by 2025 and 全流程高水平智能辅助支持—high-level intelligent assistance across the entire process—by 2030. The Supreme People’s Court policy on judicial AI
The decisive transition will arrive between those dates. ParallaxSee forecasts that by the end of 2028, an approved general-purpose judicial AI will be routinely used in Chinese courts responsible for most of the country’s caseload, performing file analysis, issue identification, authoritative legal retrieval and judgment drafting before human review. House confidence: 84%.
China will call it assistance. In daily judicial work, it will function as an active partner.
01 — China has already built the first generation of the AI case partner.
A courtroom chatbot would be a small achievement. China is building something much more consequential: an artificial-intelligence layer inside the machinery through which cases are filed, heard, decided, reviewed and enforced.
The project began before large language models. In 2017, the Supreme People’s Court instructed courts to develop electronic files, speech transcription, automatic document generation, similar-case retrieval, knowledge graphs and intelligent decision support. The ambition was to turn the accumulated record of Chinese adjudication into an operational system rather than a passive archive. The 2017 Smart Court construction opinion
Generative models supplied the missing interface. Earlier software could classify a document or retrieve a related judgment. A legal large model can now read the whole dossier, converse with the judge about it and assemble several specialised tools into one continuous account of the case. The model does not merely fetch a statute. It explains why that statute may govern a disputed fact and places the explanation inside a proposed judgment.
Shenzhen shows how far this has advanced. Its system divides adjudication into 85 procedural nodes and provides four connected modules: intelligent filing, file review, hearing assistance and document production. The system presents examination points and disputed issues, asks the judge to confirm decisions at defined stages, and then uses those decisions to generate the judgment. Every intervention can be corrected and traced.
During its initial trial, the Shenzhen system assisted with 291,000 filings and generated 11,600 draft documents. In one complex family dispute, the court reported that model-assisted processing cut the adjudication period by more than half. These are operational figures from working courts, not a benchmark constructed for a technology conference.
The active partner has therefore arrived locally. The 2028 question concerns scale: when the exceptional court becomes the national template.
02 — The machine will begin by turning the case file into a map.
A Chinese electronic dossier can contain complaints, defences, contracts, invoices, chat records, bank transfers, photographs, expert reports, hearing transcripts and earlier procedural decisions. The first technical task is not eloquence. It is reconstruction.
The judicial model will combine optical character recognition, speech recognition, document classification and entity extraction. It will identify the parties, dates, sums, claims and cited rules. A timeline will connect an agreement to a payment, a delivery, an objection and the later lawsuit. Contradictory statements will be placed beside one another. Missing signatures or exhibits will become visible.
The model will then build an evidence graph. Each disputed proposition—whether goods arrived, whether notice was given, whether damage was foreseeable—will be connected to the documents and testimony that support or challenge it. The judge will be able to ask a direct question and receive an answer grounded in the page and paragraph from which it came.
Suzhou has already assembled the foundations. Its 未来法官助手—Future Judge Assistant—uses multimodal processing to analyse 1.6 million electronic case files inside the court’s private network. The second version can process legal texts longer than 50,000 Chinese characters in one operation. During a nine-month reporting period, 653 judges and assistants used it in 94,500 cases. Every working day it completed an average of 479 intelligent file reviews, generated 70 legal documents and answered 80 case questions. The court reported that file-reading and drafting time fell by approximately half, with user satisfaction above 96%.
By 2028, this structured first reading will be the ordinary opening of a routine Chinese case. The judge will receive both the raw dossier and the machine’s navigable model of it. Judicial attention will begin at the disputes and anomalies the system has already surfaced.
03 — China’s legal model will answer from a controlled library.
A public general-purpose chatbot is trained to continue language plausibly. A court system needs a verifiable answer drawn from the law in force, the authoritative interpretation and the approved case. China is therefore building a specialised model around an official legal memory.
The most important component is the 法信法律基座大模型—the Faxin Legal Foundation Model. Its developers report using 320 million professionally labelled legal documents containing 3.67 trillion Chinese characters, together with the Faxin Outline, a legal knowledge structure assembled over more than a decade and containing 180,000 classifications. The underlying general model is supplemented through legal pre-training, supervised fine-tuning and repeated evaluation. The Supreme People’s Court introduction to Faxin
The large model supplies language and synthesis. Retrieval supplies authority. When a judge asks whether a contractual clause is enforceable, the system will search the current statutes, judicial interpretations, the Supreme People’s Court’s case database and its internal answer service. It will place the retrieved passages in the model’s context and require the proposed reasoning to point back to them.
This architecture is commonly called retrieval-augmented generation. In a court, its deeper purpose is provenance. A convincing sentence carries little value unless the judge can inspect its legal source. Every claimed rule must lead somewhere outside the model—to a promulgated text, an approved case or a page in the record.
Faxin is designed as national infrastructure rather than an isolated product. The Supreme People’s Court says it will be incorporated into the unified national court network and used to create digital case assistants, legal research experts and specialised knowledge services. That shared foundation will allow a successful provincial workflow to spread without every local court commissioning a different model.
The Chinese judicial AI of 2028 will feel conversational at the surface. Underneath, it will behave like a controlled research system whose claims remain attached to a state-maintained legal library.
04 — The model will assemble a visible chain of legal reasoning.
Legal judgment is not a contest in fluent prose. The machine must connect a rule to facts that the court is entitled to accept, address the opposing explanation and expose the path to its conclusion. Chinese researchers are trying to make that structure computational without hiding it inside an uninspectable answer.
Wang Zhu, writing through the Chinese Academy of Social Sciences, proposes a 准三段论—a quasi-syllogism—for judicial AI. The system separates the governing legal norm, the disputed issue and the legally relevant facts. It first disentangles these elements from judgments and files, then couples the appropriate rule to each issue, and finally generates an assisted inference that a judge can examine. The Chinese Academy of Social Sciences explanation of the quasi-syllogism
In simple language, the machine will show its work in four columns: what must legally be proved, what each side says happened, what the evidence supports and what result follows if those findings are accepted. A damages calculation can expose every input. A limitation-period analysis can show the triggering date. A similar-case recommendation can display both the resemblance and the decisive difference.
This structured approach gives the judge more than a single proposed answer. The system can generate alternative paths: one result if a disputed payment is treated as performance, another if it is treated as a deposit; one result if a message constitutes notice, another if it does not. The judge can alter a factual finding and observe how the legal consequence changes.
China’s large volume of standardised judgments makes this method attractive. The model can learn recurring forms of legal argument while the knowledge graph constrains each case to recognised elements. The result is a partnership between probabilistic language and explicit legal structure.
The model’s most valuable product will not be a prediction of who wins. It will be an editable map showing how each possible judgment can be reached.
05 — The national rollout has begun two years before the forecast.
Local experimentation is giving way to selection and consolidation. Shenzhen supplied a full-process model. Suzhou demonstrated long-file analysis and repeated daily use. Beijing added automatic legal views, similar-case reports and model-based supervision of procedure and legal application. The next step is to turn those separate strengths into a common product.
That step was visible in March 2026. According to a report from Shenzhen municipal media, Supreme People’s Court president Zhang Jun said the national court had combined the Shenzhen system with work from other regions. The resulting intelligent adjudication system was being tried in 23 intermediate and basic courts across 11 provincial-level regions.
This is a deliberately varied test bed. Intermediate courts handle appeals and important first-instance cases. Basic courts carry much of the everyday civil, criminal and administrative workload. Testing in both reveals whether the model can move between a small loan dispute, a family file, a commercial contract and appellate review without losing its evidentiary discipline.
The central authorities can then standardise the parts that matter: approved data sources, security rules, model evaluations, mandatory citations, review checkpoints, user permissions and the record of every human correction. Provincial systems can retain specialised workflows while sharing the same base model and audit language.
The Supreme People’s Court’s 2026 work report describes a national future in which courts work through 一张网、一个平台—one network and one platform. It pairs that project with the prudent development of AI-assisted adjudication. The 2026 Supreme People’s Court work report
On 21 September, that common platform moved from programme to operation. The Supreme People’s Court said the unified case-and-office system for courts nationwide was fully online in trial operation and had already produced gains in consistent legal application, case management and adjudication efficiency. It also directed courts to generate electronic case files alongside every matter and deepen their use. The official record of the national digital-court meeting documents a deployed shared substrate, not a proposal. It does not say that the forecast’s four AI capabilities are already in routine use across most cases, but it removes part of the integration burden between local AI pilots and a national workflow.
A national trial in 2026, provincial integration in 2027 and routine majority use in 2028 form a plausible administrative sequence. Beijing’s published 2030 objective then becomes the period of saturation and refinement rather than the first moment of adoption.
06 — Thirty-six million completed cases create an immense market for minutes.
China’s courts completed 36.2 million adjudication and enforcement matters in 2025 while receiving 37.49 million, increases of 8.9% and 10.8% respectively. The Supreme People’s Court still describes the imbalance between cases and personnel—案多人少—as an acute problem. The Supreme People’s Court’s 2025 judicial statistics
That pressure explains why the first successful judicial AI will concentrate on preparation. Reading a thick file, extracting its timeline, checking whether every claim has been addressed and producing the formal structure of a judgment consume hours without transferring responsibility away from the judge. A system that removes half of that preparation can create capacity equivalent to thousands of additional assistants.
The economic unit is tiny and powerful: minutes saved per case. Ten minutes across 36 million completed matters equals six million working hours. An hour saved equals 36 million. Even after excluding cases poorly suited to automation, the institutional return is enormous.
Uniformity provides a second return. China’s Supreme People’s Court maintains approved cases and judicial interpretations partly to reduce inconsistent application of national law. An AI assistant can compare every proposed result with those sources automatically. It can warn that compensation falls outside the ordinary range, that a required issue is missing, or that a cited provision has been amended.
Supervision will become continuous. Beijing’s platform already uses models to find procedural irregularities and questionable legal application during filing, trial and enforcement. The national system will inspect a case while it is still correctable instead of discovering defects through appeal or later review.
The machine will therefore earn its place through accumulated administrative value: shorter preparation, fewer omissions, earlier correction and more consistent reasoning. China’s extraordinary caseload turns modest improvements into a national transformation.
07 — Chinese institutional design will accelerate the spread.
A judicial model needs more than capable software. It needs access to files, an accepted legal corpus, common technical interfaces and an authority able to decide which system courts may trust. China is assembling those conditions together.
The Supreme People’s Court sets the national judicial-AI policy, operates central legal resources, promotes the unified court network and evaluates local innovations for wider use. Shenzhen and Suzhou can experiment rapidly; Beijing can absorb the successful architecture into a common platform. The institutional distance between pilot and national instruction is short.
This structure also suits the political objective of consistent adjudication. A national assistant can embody new judicial interpretations immediately, promote approved cases and flag departures from centrally expressed standards. Updates can change the research environment of thousands of judges at once.
The active partner will consequently possess two identities. For the individual judge it will be a private case assistant: patient, searchable and familiar with the dossier. For the court hierarchy it will be a national coordination system, making local reasoning more visible and comparable.
That second identity deserves attention. Software that recommends a similar case also recommends which similarities matter. A warning that a draft is unusual can protect equality while nudging a judge toward the institutional centre. Chinese scholar Liu Yanhong describes judicial AI as a form of algorithmic justice in which the judge may retreat, and argues that governance must examine training data, the model and its outputs together. Liu Yanhong’s study of algorithm-centred judicial-AI governance
China will accept this tension because standardisation is one of the project’s intended benefits. The model will become valuable to the judge by remembering the case and valuable to the state by remembering the system.
08 — The final engineering problem is calibrated trust.
Large language models can produce an invented authority with the tone and formatting of a real one. China received a practical warning when a lawyer submitted two highly relevant but fictitious cases generated through repeated model prompts. The court discovered the fabrication when it tried to verify them. The Ministry of Justice report on AI-generated false cases
Legal benchmarks confirm that fluent answers exceed dependable reasoning. LexEval evaluates Chinese models across 23 legal tasks and 14,150 expert-constructed questions, including ethical behaviour. Its purpose reflects the central problem: legal deployment demands accuracy, reliability and fairness that general language performance does not guarantee. The LexEval Chinese legal benchmark LawBench separately models memorisation, understanding and application across twenty Chinese legal tasks, revealing large differences between knowing a rule and applying it correctly. The LawBench paper A newer multi-step benchmark based on the issue-rule-application-conclusion structure reports that even its best tested model remained below 75% recall across the complete reasoning chain. The MSLR multi-step legal-reasoning study
These results define the 2028 architecture. The judicial system will ground answers in controlled retrieval, require citations that open to their sources, split reasoning into inspectable stages and preserve human checkpoints. It will use specialised models inside the court network rather than send sensitive files to a public service. It will measure abstention as a capability: when the evidence or law is insufficient, the correct output is a question for the judge.
The harder danger is automation bias. A judge under heavy workload may inspect a polished draft less aggressively than a rough one. Chinese researchers writing in Zhejiang Social Sciences therefore reduce policy to three principles: active exploration, prudent use and human-centred operation. They identify technical risk and cognitive dependence as separate problems. Chinese research on human–AI collaboration in adjudication
China’s answer will be calibrated reliance. The system will make its strongest claims where the source is explicit and display uncertainty where facts, credibility or values remain contested. Trust will be engineered case by case rather than granted to the model as a whole.
09 — The judge will remain responsible while the model becomes indispensable.
China has drawn a firm legal boundary around the partnership. The Supreme People’s Court’s 2022 policy states that AI results may serve only as references, judicial power remains with the adjudicating organisation and responsibility remains with the decision-maker. In its 2026 report, the court expressed the rule more sharply: 司法责任主体只能是法官—the bearer of judicial responsibility can only be the judge.
That boundary permits extensive automation because it separates activity from authority. The model can perform the first reading, propose the evidentiary structure, retrieve the law, calculate the remedy and draft the reasons. The judge can accept much of this work without transferring the legal act of judgment. A visible record of confirmations and edits will demonstrate where the human exercised authority.
The arrangement also fits the nature of routine adjudication. Many cases contain recurring legal forms but unique documents. The AI excels at applying a stable checklist to a large, messy file. The judge excels at assessing credibility, understanding social context, resolving genuine ambiguity and accepting public responsibility. Each will occupy the part of the process where its contribution is most valuable.
The balance will never be perfectly equal. The system that reads first influences which facts appear central and which precedents frame the choice. By 2028, judicial skill will therefore include interrogating the model: requesting the contrary line of cases, testing an omitted fact, demanding the original source and recognising when a confident structure rests on an uncertain premise.
China’s judicial AI will not sit on the bench. It will sit inside every screen around it. Its work will begin before the hearing and continue through drafting, review and supervision. Judges will remain the authors of the state’s decision, but increasingly they will author it in dialogue with a machine that has already organised the possible answers.
By the end of 2028, AI will have become an active partner in the ordinary work of China’s courts. It will read the case first. The judge will decide what that reading means.