A publisher, a stock-image agency or a competing platform has written to you about the data behind your model. Or you are the one weighing a claim, because your licensed photographs appear in a competitor’s product and you cannot tell whether they were copied, scraped or hallucinated. Either way the board asks the same question: what does Chinese law say about training data, and what would a Chinese court do with this?

The honest answer is that the law is less settled than the headlines about China’s AI rules suggest, and that the outcome in most disputes turns on doctrines written long before generative models existed: originality, substantial similarity, the closed list of copyright limitations, and unfair competition. A second answer follows from the first. Because the doctrine is unsettled, the case is usually decided on the file. A claimant who cannot show a clean chain of title, a specific output and a plausible route from the work to the model loses, however strong the moral argument. A developer who cannot show where each dataset came from, under what licence, and what it did to keep tainted material out, has no defence beyond argument.

This guide is written for both sides of that table. It covers the statutory position, the decisions that matter, and then the operational work that decides cases: how a claimant assembles a file, and how a developer documents lawful sources before anyone asks.

Who this is for

Three groups use this material. Foreign AI companies training on Chinese-language corpora, or licensing content from Chinese publishers, image agencies and platforms. Rights holders and their advisers whose material may already sit inside a model, including overseas owners whose works were scraped from Chinese platforms. In-house counsel and investors running diligence on a Chinese AI company’s data provenance, where the question is not whether the company has a policy but whether it can produce records.

Stage: Execute. The guide assumes a claim is threatened or filed, or that a corpus is being audited before a release. The preventive material is in the sections on provenance records and the checklist at the end.

The statutory starting point: no open exception for training

Chinese copyright protects original works of literature, art, science and engineering fixed in a tangible form. The 2020 amendment, in force since June 2021, modernised the subject matter and expanded the bundle of rights, but it kept the architecture that matters here: limitations on copyright are a closed list, not a standard.

The listed exceptions cover quotation for commentary, private study and research by an individual, classroom teaching, news reporting, and translation or adaptation in narrow circumstances. Each carries conditions, and each is read narrowly. There is no text-and-data-mining exception comparable to the one in EU law, and no equivalent of the US fair use test. A Chinese court has no general power to create a new exception because a technology is useful. The implementing regulations apply a three-step test to the listed exceptions, which means that even a use that appears to fall inside an exception must not conflict with the normal exploitation of the work or unreasonably prejudice the author’s legitimate interests. Commercial training at scale struggles with that second limb.

That leaves three routes to lawfulness. First, the material is not a protected work: raw facts, many datasets, and official documents fall outside copyright. Second, you have a licence that covers machine learning use. Third, the copying is so insubstantial that it does not amount to reproduction of protected expression. For a large commercial training run, most companies rely on the first and second, because the third is unpredictable and, in a corpus of millions of documents, unlikely to be true across the whole set.

The first route is broader than foreign counsel usually assume, and narrower than engineering teams usually assume. Facts, numbers and data points are not protected, but a database’s selection and arrangement may be. A court judgment or a statute is excluded from copyright, but the headnotes, commentary and translations published around it may not be. Product listings are largely factual; the photographs, descriptions and user reviews attached to them are not. A corpus assembled from marketplace listings will contain a great deal of unprotected fact and a thin but real layer of protected expression, and the claim will attach to the second layer.

Two administrative regimes sit alongside copyright and are often confused with it. The Generative AI Service Measures, effective August 2023, require providers to use lawful data sources and not infringe intellectual property. The AI-Generated Content Labelling Measures, effective September 2025, require marking of synthetic content. Complying with either does not create a copyright defence. They are separate duties enforced by regulators, and a finding under one does not decide a civil claim, although the regulator’s view of your data practices will be read with interest by any claimant’s lawyer.

What Chinese courts are actually deciding

Reported decisions cluster into three families, and the pattern matters more than any single ruling.

Output-side ownership. A Beijing internet court held in 2023 that an image produced through sustained human direction of a generative tool could be protected as a work, because the person’s choices in prompting and selection amounted to intellectual creation. Later cases have refused protection where the human contribution was thin. The lesson is that your users may own outputs, or nobody may, depending on how the tool is used and what your terms say.

Provider liability for output. A Guangzhou court held a generative service provider liable where its model produced images substantially similar to a well-known animated character, emphasising that the provider had not adopted reasonable preventive measures. This line of reasoning drives product design: filters, blocked prompts, complaint channels and fast removal are treated as evidence of care. It also tells a defendant what to build before a claim, and a claimant what to look for after one.

Input-side claims over corpora. Suits by authors, photographers and agencies against model developers are fewer and slower than in the United States, and several are framed in contract or unfair competition because pure copyright is hard to prove. Evidence, not doctrine, is the bottleneck. The claimant must connect a specific work to a specific model, and Chinese procedure offers no discovery to help.

Dispute typeWho usually suesMain legal basisWhat decides the case
Output reproduces a character or artworkCharacter owner, film or game companyCopyright infringement, unfair competitionSubstantial similarity plus the provider’s preventive measures
Ownership of AI outputUser, platform, employerCopyright authorshipDegree of human creative input and contract terms
Training corpusAuthors, publishers, agenciesCopyrightWhether works were copied and whether a licence or exception applies
Scraped platform contentPlatform operatorAnti-Unfair Competition Law, contractRobots protocol, terms of use, substitutive effect
Dataset misappropriationDataset ownerTrade secret, contractSecrecy measures and access trail

The table is also a map of the evidence each claim requires. A character case needs comparison evidence and a record of the provider’s safeguards. An ownership case needs the prompt history and the selection record. A corpus case needs the chain from work to dataset to checkpoint. A scraping case needs the platform’s terms, its robots file and proof of the effect on its service.

Copyright requires copying of protected expression. Scraping prices, listings, reviews or specifications may not infringe copyright at all, which is why Chinese platform litigation runs mainly on the Anti-Unfair Competition Law. The general clause requires business operators to observe the principles of fairness, honesty and commercial ethics, and courts use it to condemn conduct that free-rides on another operator’s investment. A separate provision addresses conduct on the internet, and the 2024 rules on network unfair competition give the regulator a basis to act without litigation.

The factors that recur are practical rather than doctrinal: whether the data was public; whether the scraper bypassed login walls, rate limits or anti-bot measures; whether the volume and frequency exceeded normal human use; whether the resulting product substitutes for the source; whether the platform’s terms prohibited the conduct; and the effect on the platform’s incentive to keep investing. A robots.txt file is not a statute, but Chinese courts have treated it as industry practice and as evidence of the operator’s stated position, so ignoring it is difficult to explain later.

For a claimant, that list is a pleading checklist. Show that the data was not simply public but curated at cost; that the terms prohibited extraction; that the defendant’s conduct was systematic; and that the output competes with the platform’s own service. For a developer, the same list is a design specification. Record the robots file you read, the date you read it, and the version of the terms in force; respect rate limits; keep out of login-protected areas; and be able to show that your product does not substitute for the source.

Separate regimes add teeth. Circumventing technical protection measures can engage criminal law on unauthorised access to computer systems, and any personal information in the corpus brings the Personal Information Protection Law into play, with its own consent, notice and minimisation duties. A scrape that is defensible on copyright grounds can still be indefensible overall. For a foreign developer, the practical consequence is that a single crawl can generate three exposures in three different forums: a civil copyright claim, an unfair competition claim by the platform, and a regulatory investigation into personal information handling.

Databases, compilations and AI outputs

China has no sui generis database right of the European kind. Protection comes from three sources. Compilation copyright can cover a database where the selection or arrangement is original, but it protects the selection and arrangement, not the underlying data, and an arrangement generated automatically is unlikely to qualify. The Anti-Unfair Competition Law fills the gap where substantial investment has been misappropriated. Contract and API terms do the rest, although click-through terms are tested in court rather than assumed to bind.

On outputs, the analysis mirrors the ownership cases. Purely machine-generated content is generally not a work, so no one owns it, but that does not make it free to use: it may reproduce third-party expression, disclose personal information, or breach the terms of the model you used. Your customer contract should state who owns outputs, what warranties you give, and how you respond to third-party notices. The warranty is where foreign developers most often take exposure they did not intend, because a standard global terms of service promises that the service does not infringe, and Chinese law may not support that promise for a model trained on an unlicensed corpus.

How a claimant builds the file

A Chinese copyright claim over training data succeeds or fails on six questions, and each has a document behind it. The sequence below is the order in which a well-prepared claimant works, not the order of a pleading.

One: a work you own, with a chain of title. Choose a work where ownership is clean and provable. Registration with the copyright administration is voluntary, but a registration certificate is treated as prima facie evidence of authorship and is the cheapest document to obtain. Where registration is absent, assemble the creation record: drafts, raw files, EXIF data, publication dates, and the contract by which rights passed to you. For a publisher, that means the publishing agreement with the author and, where relevant, separate grants for translation and adaptation. For a photographer, the RAW files and the licensing history. A claim built on a work whose title chain has a gap will be met with a standing objection that consumes the first hearing.

Two: a fixed copy of the output. You need to show the model reproducing protected expression, not merely producing something in a similar style. Style, ideas, facts and genre conventions are not protected. The workable method is to run a series of prompts, record each prompt, the timestamp and the output, and preserve the record by notarisation or by a trusted timestamp before the model is updated. Several runs are necessary because output varies. The comparison then has to isolate the protectable elements: the specific arrangement of words, the particular composition of an image, the distinctive features of a character. An expert comparison report is often the difference between a claim that survives and one that is dismissed as an argument about genre.

Three: a route from the work to the model. This is the weakest link and the point at which most claims stall. Chinese civil procedure has no discovery in the American sense, and a defendant is not obliged to hand over its corpus. What is available is narrower: an application for evidence preservation, in which the court attends and preserves specific material; a request that the court investigate a particular matter; an order to produce a defined category of evidence, with the possibility that refusal is held against the party; and expert appraisal of the model. Each requires you to describe what you are looking for with precision. A request for “all training data” will be refused as a fishing expedition; a request for the dataset manifest, the source list and the version control record for a named checkpoint has a chance.

Some claimants find the link without the court. Model cards, technical papers and conference presentations sometimes name the datasets. Dataset vendors publish their contents. A former employee may provide the manifest. A licence you already hold may carry audit rights, which is the most reliable route of all and the reason licence drafting matters to claimants as well as developers.

Four: damage, quantified. Chinese damages in copyright cases remain modest by international standards, and the court has discretion. The claimant can plead actual loss, the infringer’s gain, a reasonable licence fee by reference to comparable transactions, or statutory damages, which carry a ceiling of RMB 5 million. Punitive damages are available for intentional infringement in serious circumstances. In practice the licence-fee benchmark is the most persuasive where the work has a licensing market, and interim relief is worth more than the award. A claimant who needs the output stopped should apply for behaviour preservation at the outset and support the application with evidence of continuing harm, because a damages award that arrives after the product has saturated the market compensates very little.

Five: parallel routes. Copyright is not the only pressure point. An administrative complaint to the copyright authority can produce an order to cease and a fine, at low cost and on a shorter timeline than litigation. An unfair competition complaint to the market regulator serves the same purpose where the conduct is scraping rather than reproduction. A platform takedown notice is cheap and fast where the infringing output is hosted by a third party. Where personal information is involved, a complaint to the cyberspace authority adds a regulator to the file. These routes do not produce damages, but they change the defendant’s cost of continuing.

Six: forum and time. Civil copyright claims may be brought at the defendant’s domicile or where the infringement occurred, and the internet courts in Beijing, Hangzhou and Guangzhou handle a large share of online claims. The limitation period is three years from the date the claimant knew or ought to have known of the infringement and the infringer. Continuing infringement is treated separately, which matters where a model has been in service for years.

How a developer documents lawful sources

The developer’s task is to be able to answer, months later and under pressure, four questions about any dataset: where it came from, what right you had to use it, what you did with it, and what you would do if a court told you to stop. A company that can answer those four questions has a defence worth arguing. A company that cannot has a settlement.

The provenance register. One record per dataset, maintained as a controlled document rather than a spreadsheet owned by an engineer. The fields that matter in a dispute are: a dataset identifier and version; the source and the counterparty; the acquisition method (licensed, purchased, open licence, public crawl); the date and volume; the licence reference and its file; the restrictions (field of use, commercial use, sublicensing, attribution, geographic limits); the personal information assessment; the reviewer and approval date; a hash of the delivered dataset; and the retention location. The register should be capable of producing, for any model checkpoint, the list of datasets used and the version of each.

Licence terms that actually cover training. Many content licences predate generative AI and say nothing about machine learning. A licence that grants “reproduction for internal business purposes” may not cover the creation of a model that is later distributed. The clauses to negotiate are: an express right to use the content to train, fine-tune and evaluate models; a right to use the resulting model weights and outputs commercially; the right to sublicense to affiliates and to contractors; a warranty of title from the licensor; an indemnity for third-party claims; audit and reporting rights, which cut both ways and should be mutual; record-keeping obligations; and a termination and wind-down mechanism that tells you what to do with the model if the licence ends. That last clause is the one everyone forgets, and it determines whether a licence termination destroys the product or merely the dataset.

Crawl discipline. For public web data, the record should show that the crawl respected the signals available at the time. Snapshot the robots file with a timestamp and a hash, and keep the version of the terms of service in force on the crawl date. Log the rate limits applied, the user agent, and any exclusion list. Do not circumvent login walls, paywalls or anti-bot measures; if a business unit proposes to, require a written legal opinion before the crawl rather than after. Maintain a removal log: when a rights holder asks for exclusion, record the request, the date, the action taken, and whether the material was removed from future training runs only or also from the deployed model.

Technical lineage. A register is a statement of intent; the lineage is the proof. Keep dataset versions immutable, hash each shard, and record which shards fed which training run. Maintain a manifest for each model version listing the datasets, the preprocessing steps and the checkpoints. Retain the manifest even where the raw data is deleted for cost or privacy reasons, because the manifest is what you will produce in court. Where the law or a licence requires deletion, the practical question is whether you can retrain without the subset, so the pipeline should be built so that a dataset can be removed and the model rebuilt.

Governance that produces evidence. A legal review before a dataset enters the pipeline, with a named approver and a dated record. Procurement diligence on dataset vendors, including questions about how they obtained the data and whether they will stand behind it. An annual review of the register against the deployed models. Training for engineering staff on what may and may not be ingested, because the crawl that creates the problem is usually run by someone who did not think of it as a legal act. Where the company files a generative AI service with the regulator, the filing materials require a description of data sources, and the register should be the document from which that description is written. The filing is not a clearance, but a register that matches the filing is strong evidence of consistency, and a mismatch is a gift to the other side.

Output controls. Because provider liability has been decided on the adequacy of preventive measures, the record should include the filters in place, the blocked prompt list, the similarity checks run on outputs, the complaint channel and the response time achieved. Log moderation decisions. A provider that can show it removed infringing output within days of notice is in a different position from one that cannot.

The defences the file must support. Four defences are worth building towards, because each is provable from records rather than argument: a licence that covers machine learning use; subject matter that is not protected; copying too insubstantial to amount to reproduction; and independent development, evidenced by a clean-room record showing that the work was produced without access to the claimant’s material. On the unfair competition side, the corresponding defences are that the data was public and that the use is not substitutive of the platform’s service. A developer who cannot point to the record behind each defence is left arguing about doctrine, which is the position the claimant wants.

Documents, filings and timeline

The table sets out the steps that generate the documentary record, the body involved and where foreign parties lose time. It applies to both sides, with the roles reversed.

StepAuthorityTypical timingWhat foreign companies miss
Preserve online evidence of outputNotary public, or a trusted timestamp serviceDays, if instructed earlyModels are updated; an output captured after the update may not be reproducible
Apply for evidence preservation against the developerPeople’s courtWeeks from a complete application; security may be requiredThe application must name the material; “the training data” is not a description a court will accept
Obtain copyright registration or reconstruct titleCopyright administration; private recordsWeeks for registration; longer for a chain with gapsA foreign publisher’s rights may be held by an affiliate, and the claimant must be the right entity
File an administrative complaintLocal copyright or market regulation authorityMonthsRegulators act on local targets; an offshore developer may be outside their reach
Send a platform takedown noticeHosting platformDaysPlatforms require a statement of rights and a precise URL, not a general complaint
Review filing materials for consistencyProvincial cyberspace administrationAt filing and on changeThe data description in the filing will be compared with the corpus in any later dispute
Commence civil proceedingsInternet court or the court of the defendant’s domicileFiling to first hearing usually several monthsService on an offshore defendant follows a treaty route and adds months
Commission a technical comparison or appraisalExpert institutionMonthsThe expert needs a defined question; an open-ended remit produces a report that decides nothing

Where foreign companies get this wrong

Fair use is imported with the business plan. A US-trained team assumes that training is transformative and therefore permitted. Chinese law has no such standard, and a defence built on transformation will be met with the closed list. The same assumption produces internal documents that read, in translation, like an admission that the use was unlicensed.

“Publicly available” is treated as a licence. Public availability affects the unfair competition analysis, not the copyright one. A photograph on a public page is still a protected work, and the fact that it could be viewed does not authorise reproduction in a corpus.

Provenance is reconstructed after the claim. Diligence requests, filings and litigation all ask the same question, and the answer has to come from records made at the time. A corpus rebuilt from memory and invoice files is a liability, because the gaps are visible and the other side will point at them.

The filing is treated as a clearance. A generative AI filing confirms that the service meets the regulator’s requirements for data sources. It does not adjudicate copyright, and it does not stop a rights holder from suing. Companies that rely on the filing as a defence are relying on the wrong instrument.

The output side is ignored while the corpus is audited. The largest reported liability has arisen from what the model produced, not from what it ingested. A company with a clean corpus and no filters is exposed; a company with a questionable corpus and good filters is exposed differently. Both need attention.

A claim is brought on a work the claimant does not own. On the claimant’s side, the most common failure is a title chain that runs to an affiliate, a licensor that retained the right to sue, or a photographer who assigned only publication rights. Standing is checked first and dismissal is cheap for the defendant.

Cost and commercial impact

The cost of a training-data dispute is front-loaded into evidence, which is the opposite of most commercial litigation.

For a developer, the largest item is a corpus audit: sampling the register, checking licences against use, tracing datasets to checkpoints and identifying the subsets that cannot be defended. A technical expert is usually needed to explain what the model contains and how it was trained. Remediation follows: retraining without a tainted subset, rebuilding pipelines so that removal is possible, and renegotiating licences that do not cover machine learning. The cost is real but bounded, and it is a fraction of the cost of an injunction against a product in market.

For a claimant, the cost is the technical survey, notarisation of outputs, expert comparison and the litigation itself. Damages are usually modest, so the economics only work where the claimant has a licensing business to protect, a portfolio of works, or a strategic reason to stop the output. Interim relief and administrative pressure are the return on that investment.

Both sides should price the indirect costs. A disclosed claim affects a financing round and a sale. A customer indemnity invoked by an enterprise buyer is more expensive than the underlying claim, because it arrives with a service-level dispute attached. Insurance is available in some markets for media liability, but policies written before generative AI usually exclude it, and a policy that covers it will ask for the register described above.

Practical checklist

Geason Law Offices advises foreign AI companies and rights holders on training-data exposure in China, from corpus audits and licence drafting to infringement claims and regulatory complaints. The intellectual property and commercial dispute practices handle the two sides of that work, and data compliance covers the personal information issues that arise from the same crawl. For the licensing layer, see training data licensing and PIPL for AI training data; for how the output cases are being decided, see the analyses of the Guangzhou AIGC copyright case and the Shanghai prompt copyright case.

Frequently Asked Questions↑ Back to article

Is it legal in China to train an AI model on copyrighted works?

There is no blanket prohibition and no blanket permission. Training on protected works without a licence is exposed to a copyright claim unless a specific statutory exception applies, and the exceptions are narrow. In practice, legality turns on the source of each dataset, the licence attached to it and whether the use falls inside an enumerated exception, so provenance records decide most outcomes.

Does China have a fair use doctrine that covers AI training?

No. China uses a closed list of specific limitations rather than an open fair use standard, and none of the listed exceptions was written for commercial model training at scale. Courts cannot invent a new exception, so a defence usually rests on the work not being protected, a licence, or copying that is too insubstantial to matter.

Can AI-generated content be protected by copyright in China?

Sometimes. Chinese courts have protected AI-assisted output where a person made substantial creative choices, such as iterating prompts and selecting results, and have refused protection where human input was minimal. The test is whether the output reflects the author's intellectual creation, so the answer depends on your workflow rather than on the tool alone.

What happens if our model outputs something similar to a well-known character?

You face both copyright and unfair competition exposure, and Chinese courts have held generative service providers responsible where output reproduced a protected character and the provider had taken no reasonable precautions. Practical responses include output filters, blocked prompts, rapid takedown and documented moderation. The provider's conduct after notice often matters as much as the similarity itself.

Is scraping Chinese websites and platforms lawful?

Not automatically. Scraping public facts may avoid copyright infringement, but Chinese platform litigation relies heavily on the Anti-Unfair Competition Law, which can condemn scraping that substitutes for the original service or free-rides on the operator's investment. Circumventing technical measures can also create criminal and data-security exposure.

How do we prove what data went into our model if we are sued?

You prove it with records you kept before the dispute: dataset versions, source licences, collection logs, robots protocol snapshots and training manifests. Chinese procedure has no US-style discovery, so courts rely on evidence preservation, court investigation and contractual audit rights. Reconstructing a corpus after a claim is filed is usually impossible.

Should we pause training on Chinese-language data until the rules are clearer?

A blanket pause is rarely necessary, but an unlicensed crawl of Chinese content is a real risk that waiting will not cure. A better approach is to separate licensed, open and scraped corpora, document decisions, and keep the ability to retrain without the tainted subset. Confirm the position with counsel before a launch that depends on scraped data.

Geason Law Offices AI Industry Research Group | Reviewed by He Qianggao (Lawyer License No. 14403201310092337, 10+ years of practice)