What Western Legal AI Vendors Get Wrong About Chinese Court Data

A Western legal AI team that has already shipped US and EU coverage tends to approach China as "the same problem, one more jurisdiction." That instinct is the root of almost every stalled China-coverage project we see. The data engineering playbook that worked on common-law case law — crawl the public dockets, parse the citations, index the English — rests on a stack of assumptions that quietly do not hold in China. This piece names the five that do the most damage, why each one breaks, and what Chinese court data actually demands instead. It's a companion to our structural argument in Why Western Platforms' China Coverage Isn't Built for AI — that piece is about the incumbents' data; this one is about the mistakes a vendor makes building their own.

Mistake 1: Treating it as a scraping problem

In the US, "get the case law" is largely a crawling exercise: public dockets, bulk download programs, and open APIs mean the data is retrievable, and the engineering is about cleanup. Teams carry that mental model to China and scope a scraper as the core task. It's the wrong center of gravity.

Chinese judgments are published through a centralized channel with no public bulk endpoint, access that has tightened over time, and full text in Chinese. A scraper, even a good one, returns an opaque, partial pile: you cannot say what fraction of judgments you have, which causes of action or regions are under-represented, or how many documents are near-duplicates of each other. The genuinely hard work — coverage characterization, field extraction, de-duplication — sits after retrieval and is exactly what scraping doesn't touch. We walk through that structuring work in The 170M Judgments Problem; the headline is that crawling is the easy 10%.

Mistake 2: Importing common-law precedent structure

This one is subtle because it shapes the retrieval design, not just the ingestion. Common-law legal AI is built around a citation graph: cases cite cases, precedent binds, and "find the controlling authority" is a well-defined retrieval target. Teams reuse that framing in China and build for a citation network that isn't there.

China is a civil-law jurisdiction. Prior judgments are not formally binding the way common-law precedent is. The Supreme People's Court issues a small set of Guiding Cases that lower courts should refer to, but the everyday corpus has no stare-decisis citation graph to traverse. What the data does support — extremely well, at volume — is similar-case and pattern analysis: how courts in this region have decided this cause of action, what damages ranges appear, how outcomes move with fact patterns. That's a statistical, retrieval-heavy task over a large structured corpus, not a precedent-walk. Design for the former and the product works; design for the latter and you spend months looking for a graph that doesn't exist.

Mistake 3: Trusting general machine translation

The plan reads reasonably on a slide: scrape the Chinese, run it through a strong translation model, index the English, done. It demos well and fails in production, because legal language is where generic machine translation is weakest.

• Terms of art collapse. Distinct causes of action or statutory references translate to the same everyday English phrase, so two legally different cases become indistinguishable to your retriever.
• Identifiers drift. Case numbers (an-hao), court names, and article references get paraphrased or mangled — destroying the one thing a legal answer must preserve, a resolvable pointer to the source.
• You translate away ground truth. If the English becomes the artifact you reason over and the link to the Chinese original is gone, every output is an unverifiable paraphrase.

The fix isn't "avoid English" — English access is what makes the corpus usable to a non-Chinese-reading team. It's that the English has to be engineered as an access layer over authoritative Chinese sources, with a bilingual taxonomy and citation grounding, not a lossy replacement. We unpack that in Is There an English-Language API for Chinese Court Judgments?

Mistake 4: Assuming published = complete and representative

Western teams used to near-complete public dockets assume the same of China: that "the published judgments" stand in for "all judgments." They don't. Publication is selective. Settled and withdrawn matters don't appear; categories touching privacy, minors, or state secrets are withheld; publication rates vary by region and period; and a meaningful number of judgments were published and later removed.

If you train on or compute statistics over that pile as if it were the population, your model and your numbers inherit a selection bias you can't see. The discipline that separates a usable corpus from a misleading one is coverage characterization: being able to state what years, regions, court levels, and causes of action are well-covered and what is thin or missing. A vendor that can't tell you its coverage shape can't help you bound your own claims — this is the difference between a dataset and a liability, and it's a recurring theme in our 2026 market map.

Mistake 5: Underestimating the structuring tax

Even teams that clear the first four often budget China like a parsing task and discover it's a multi-quarter data-engineering program. Two things blow up the estimate at 170M+ scale:

• Field extraction from prose. Case number, court, level, date, cause of action, and outcome live inside the judgment text, in formats that drift across courts and years. A rule that's correct for 99% of documents is still wrong on millions.
• De-duplication. The same dispute appears as first-instance, appeal, and retrial; republished copies and near-duplicates abound. Without aggressive de-duplication your "counts" are inflated and your retriever surfaces the same matter five times.

This is high-effort, low-differentiation work: two vendors who both do it correctly end up competing on product and workflow, not on who cleaned the text better. That economics — the structuring layer is expensive and doesn't differentiate — is precisely why most teams are better off licensing a structured corpus than rebuilding one, the license-vs-scrape calculus.

What Chinese court data actually requires

Put the five mistakes in reverse and you get the spec for doing it right. None of it is exotic; all of it is the work scraping skips.

This is what SinoVerdict is built to deliver: a licensed, structured corpus of 170M+ Chinese court judgments with normalized fields, de-duplication, a cause-of-action taxonomy, English access, and citation grounding back to every original judgment — available as a bulk dataset, a REST API, and an MCP server, with daily updates. The full assembly is laid out in Building China Coverage Into Your Legal AI.

The bottom line

The Western legal AI playbook is good — it just encodes assumptions about open dockets, binding precedent, English text, complete coverage, and cheap parsing that are all artifacts of common-law jurisdictions. China breaks every one. The teams that ship China coverage are the ones that drop the scraping reflex, design for similar-case retrieval instead of a precedent graph, treat English as a grounded access layer, demand coverage characterization, and respect the structuring tax — or hand those layers to a partner who has already paid it. Get the assumptions right and 170M+ judgments become a real product surface; carry the wrong ones across the border and the project stalls in the same place every time.