License vs Scrape: How AI Companies Should Source Chinese Case Law

Every legal AI team that decides to add China coverage runs the same internal debate. One engineer says: the judgments are public, the courts published them, why pay—we'll just scrape them. Someone in legal says: are we sure we're allowed to do that? And a product lead, eyeing the roadmap, asks the only question that matters: which path actually gets us a reliable China feature, faster?

This is the classic build-versus-buy decision, applied to a corpus with some unusual properties. It deserves an honest answer rather than a sales pitch, so this piece lays out what scraping Chinese case law really costs—in coverage, legal exposure, and maintenance—against what a license actually buys you, and where each one genuinely wins.

The seductive case for scraping

The appeal is real, and worth stating plainly so we're comparing honestly. Chinese court judgments were published on China Judgements Online, the Supreme People's Court's public repository. Public means free. A competent team can stand up a crawler in a sprint, pull documents, and feel like they've sidestepped a procurement cycle and a line item. For a quick proof-of-concept—"can our model reason over Chinese precedent at all?"—that instinct is not wrong. If the goal is a demo over a few thousand documents, scrape away.

The trouble starts the moment "demo" becomes "product." Three things that don't show up in the first sprint show up in the next two quarters: the data isn't all there anymore, the legal questions don't go away, and the pipeline never stops needing you.

Cost one: the coverage isn't there to scrape

The biggest change since the era when scraping China case law was straightforward is that the open tap has been turned down. We've documented the full curve elsewhere, but the shape is what matters here. Public availability peaked in 2020 at 19.2 million judgments published in a single year, then fell to roughly 78%, 54%, and 27% of that peak across 2021, 2022, and 2023 respectively. In 2024 the system reorganized: a comprehensive database moved behind the courts' own internal network, and a smaller curated reference library remained public.

For a scraper, that history is decisive. A crawler can only collect what is exposed today, and what is exposed today is an uneven, shrinking slice—heavily weighted to certain years, thin in exactly the recent window many products care about most. You don't find out what's missing until a user asks about a 2022 dispute and your system has nothing on point. Scraping gives you the documents that happen to still be reachable, not the corpus. The gap between those two is invisible at build time and very visible in production.

Cost two: the legal questions don't resolve themselves

This is the part engineering teams tend to underweight and legal teams lose sleep over. We are not your counsel and this is not legal advice—but the structural point is straightforward: scraping and licensing carry different risk profiles, and only one of them comes with answers in writing.

A scraping approach leaves several questions open at once. The terms of use on the sources you crawl are one. The cross-border considerations that attach to PRC-origin data—judgments contain personal information, and China's data and personal-information regime treats outbound transfer as a regulated act—are another. And if a customer, a regulator, or a counterparty later challenges how the data underneath your product was obtained, a scrape gives you no contractual basis to stand on and no one to point to. You absorb all of it.

A license doesn't make law go away, but it changes your position. Sourcing basis and permitted uses—training, retrieval, indexing, redistribution inside your product—are written down. Responsibility for the provenance of the corpus sits with the provider, not buried in your crawler's git history. For an AI company that will itself be sold to enterprises with their own procurement and security reviews, "we license our China data under written terms" is a far easier sentence to put in a due-diligence questionnaire than "we scraped it."

Cost three: the pipeline is a tax you pay forever

Suppose you accept the coverage limits and your counsel is comfortable with your method. You still own the hardest part, and you own it indefinitely. The crawler is maybe 10% of the work. The other 90% is what turns raw scraped pages into something a model can actually use:

• Parsing. Judgment formatting is inconsistent across courts and years. Extracting clean text, separating the reasoning from the operative ruling, and stripping boilerplate is ongoing work, not a one-time script.
• Deduplication. The same judgment appears in multiple forms and places. Without robust dedupe you inflate your counts and skew any statistics or training distribution built on top.
• Field normalization. To filter, cite, and retrieve, you need stable structured fields—case number, court, date, cause of action, parties, outcome—on every record. Deriving those reliably from free text across decades of documents is a real NLP project in its own right.
• Breakage. Every time a source changes its layout or access rules, something in your pipeline breaks, silently. You inherit a permanent maintenance burden that competes with your actual roadmap.

None of this is glamorous, and all of it is load-bearing. The honest accounting of "free" scraped data is the fully loaded cost of the engineers who keep that pipeline alive—quarter after quarter—instead of building the product your customers are paying for.

The comparison, laid out

Put side by side, the trade is clearer than the "it's free" framing suggests.

So when does scraping actually win?

It would be dishonest to claim never. Scraping is the right call for a genuinely narrow, bounded job: a single court, a single cause of action, a fixed window, where approximate coverage is fine and the result is an internal research artifact that will not ship inside a commercial product. If that's your use case, a license is overkill and a weekend crawler is the pragmatic answer.

The calculus flips the moment any of the following is true: you need full-corpus grounding so retrieval doesn't miss on-point precedent; you need a representative distribution for fine-tuning or for a benchmark you'll trust; you need stable fields for citations a user can verify; or you need to be able to tell your own enterprise customers, on the record, where your data came from and that you're licensed to use it the way you do. At that point the scraper's hidden costs—coverage holes, open legal questions, perpetual maintenance—exceed the price of buying the corpus already solved.

The bottom line for legal AI teams

"The judgments are public, so the data is free" is true and misleading in the same breath. The documents may be public; an AI-ready, full-coverage, structured, rights-cleared corpus is not something you find lying in the open—it's something that gets built, and the only question is whether your team builds it or buys it. For a throwaway experiment, build. For anything that ships, the math almost always favors licensing, because what you're really buying isn't the bytes—it's the coverage guarantee, the structure, the legal footing, and every engineering quarter you'd otherwise spend on data plumbing.

That's the category SinoVerdict occupies. We license a structured, machine-readable corpus of more than 130 million Chinese court judgments, coverage through 2023 with incremental updates on request, delivered as a bulk dataset, a REST API, and an MCP server, with stable fields on every document and AI-use rights in writing. For the deal mechanics and a due-diligence checklist, see our guide to licensing Chinese court judgment data; for how the documents are structured for retrieval and extraction, see our walkthrough of the case law API and document fields.