The State of Chinese Legal AI Data, 2026: A Supply-Side Market Map

If you are building a legal AI product and you have decided it needs to cover Chinese law, the first hard question is not "how good is the data" — it is "where do I even buy it." The supply side of Chinese case law is crowded, confusing, and mostly built for someone who is not you. Government portals, decades-old domestic subscription databases, Western incumbents with a thin English shelf, grey-market scrapers, and a newer category of AI-native licensors all technically offer "Chinese legal data," and they are almost impossible to compare on a single page because each was designed for a different buyer with a different job to do.

This is a map of that supply side as it stands in 2026. Not a ranking — a map. The point is to show you which category fits a legal AI vendor's job, and why most of the obvious names don't, despite being excellent at the job they were actually built for. (For the demand-side question — how much data exists and in what shape — see The State of Chinese Case Law Data, 2026; this piece is about who supplies it.)

The buyer's job changes everything

Before the categories, one framing that makes the whole map legible: a data source is only "good" relative to a job. A lawyer's job is to read cases. A legal AI vendor's job is to feed a system that reads cases. Those are different jobs with almost opposite requirements:

Hold this table in mind and the map below explains itself: most Chinese legal data providers are excellent lawyer-sources and poor vendor-sources, not because their content is weak, but because they were built for the left column.

Category 1 — Official channels

The authoritative origin of Chinese judgments is the courts themselves and the government databases that publish them. This is where the data is born, and in principle it is the most trustworthy source. In practice, for a legal AI vendor, three things make official channels a starting point rather than a solution. First, they were never designed for bulk machine consumption — the interfaces assume a human looking up one case at a time. Second, open accessibility has narrowed over recent years; the volume that was once broadly downloadable is no longer a stable, complete feed you can build a product on. Third, the raw output is unstructured prose with no extracted fields, no deduplication across instances, and no normalized taxonomy. Official channels define the ground truth, but turning them into a usable corpus is the entire problem — one we break down in The 170M Judgments Problem.

Category 2 — Domestic aggregators

The deepest, best-curated Chinese legal content sits with the established domestic platforms — the PKULaw / 北大法宝 lineage, Wolters Kluwer's China offering, and their peers. These are mature, respected products with rich editorial layers: classification, citators, annotations, statute linkage. If your job is to be a Chinese-reading lawyer doing research, they are outstanding.

The mismatch for a legal AI vendor is not quality — it is the access model and the license. These platforms sell per-seat subscriptions to a human-facing search interface. The unit of delivery is a document for a person to read, not a structured field set for a system to ingest. Their terms are generally built around viewing, not bulk extraction for training or grounding an external AI product, and the content is Chinese-native with no English layer designed for cross-lingual retrieval. You can put a lawyer in front of them; you cannot cleanly drop their corpus into your pipeline. They occupy the top-left of the framing table, and a vendor needs the top-right.

Category 3 — Western incumbents

LexisNexis, Westlaw, vLex and the other global research platforms all carry some China material, usually as an English-language secondary-source layer: selected translated cases, country guides, practice notes. For an international firm's occasional China question this is a reasonable convenience. For a vendor trying to power a China feature, the limitation is severity of scale: what these platforms hold in English is a curated, human-reference slice — thousands of documents, not the full corpus — and it was assembled for reading, not for machine retrieval at completeness. We've detailed why this shape doesn't serve AI in Why Western Platforms' China Coverage Isn't Built for AI. The takeaway for the map: incumbents are a distribution and trust asset, not a primary data source for Chinese case law at corpus scale.

Category 4 — Raw scrapers and grey data

There is always a cheaper path: scrape what is public, or buy a dump from someone who did. The appeal is obvious — bulk volume at near-zero licensing cost. The problems are equally obvious once the data has to stand behind a product:

• Legality and redistribution. Commercial use and redistribution of scraped court data rest on unclear footing, and that risk transfers to you and your customers. (Informational background, not legal advice.)
• Coverage you can't characterize. A scrape gives you whatever was reachable on the days it ran — uneven, undocumented, impossible to state completeness on to an enterprise buyer.
• No structure, no grounding. You inherit raw prose: no extracted fields, no deduplication, no normalized cause-of-action taxonomy, no stable citation back to source. The expensive engineering is still entirely ahead of you.
• Freshness rots. A one-time dump is stale the day after; maintaining a scraping operation against changing access is its own ongoing cost and risk.

For a weekend prototype, fine. For a commercial legal AI product whose answers must be defensible, the hidden cost of cleaning, structuring, maintaining, and legally defending scraped data typically dwarfs the cost of licensing a corpus built for the purpose. We compare the two paths directly in License vs Scrape.

Category 5 — AI-native structured licensors

The newest category exists precisely because of the gap the first four leave open: a source built from the start to be ingested by an AI system rather than browsed by a person. The defining properties are the top-right column of the framing table, delivered as a product:

• Structured and machine-readable — fields like case number, court level, document type, cause of action, parties (de-identified), statutes, and outcome extracted from the prose, not left buried in it.
• Complete and deduplicated at corpus scale, with the same matter reconciled across instances and a normalized cause-of-action taxonomy.
• Citation-grounded — every record resolves back to the original judgment by stable identifier, so an AI answer can be verified rather than hallucinated.
• Licensed to ingest — the terms contemplate storing and serving the corpus inside your product, not just viewing it.
• Delivered as infrastructure — bulk dataset, REST API, and MCP server, with a refresh cadence so the corpus stays current.

This is the category SinoVerdict occupies. The point of the map is not that it is "better" than a domestic aggregator in the abstract — it is that it is built for a different job. The aggregator wins the lawyer's job; the structured licensor wins the vendor's. A full walkthrough of assembling exactly this into a product is in Building China Coverage Into Your Legal AI.

The map on one page

Scored for the legal AI vendor's job specifically — the same source would score differently for a practicing lawyer:

How to read the map for your own build

Three practical reads come out of it. First, don't confuse a great lawyer-source with a vendor-source — "PKULaw is the gold standard" is true and irrelevant if your job is to feed a pipeline. Second, the incumbents are a channel, not a quarry — valuable for distribution and trust, not for sourcing the corpus. Third, the real choice for most vendors is scrape-and-build versus license-structured, and that choice is an engineering-cost and legal-risk decision, not a data-quality one: the content origin is the same public record either way; what differs is whether someone has already done the structuring, deduplication, grounding, and licensing work, or whether you will.

The bottom line

The Chinese legal data market looks crowded, but for a legal AI vendor it thins out fast once you score each category against the actual job: ingest a complete, structured, grounded, license-clean corpus into a product. Official channels are the origin but not the answer; domestic aggregators and Western incumbents are built for human readers and human seats; scrapers trade legal and engineering risk for a low sticker price. The category built for the vendor's job is the AI-native structured licensor — and that is where SinoVerdict sits: a structured, machine-readable corpus of more than 170 million Chinese court judgments with extracted fields, a normalized taxonomy, an English layer, and citation grounding back to every original document, delivered as a bulk dataset, a REST API, and an MCP server.