A foreign AI company usually asks the intellectual property question at the wrong moment. The question arrives when a Chinese partner asks for the model, when a customer’s procurement team sends a security questionnaire, or when an investor’s counsel opens a data room. By then the two decisions that matter have already been made, and neither was made deliberately: who owns what the engineers built, and what was disclosed before anything was filed.

Chinese law does not reward good intentions in either area. Ownership follows the documents, and the documents are usually silent. Disclosure is irreversible, and the system has no general grace period to undo it. A company that has built a strong model and a weak paper trail holds an asset it cannot sell, license or enforce without first repairing the record, and the repair is always more expensive than the original paperwork would have been.

This guide is for the point before any of that happens: the stage at which the company is deciding which rights to pursue and what to fix before it shows the model to anyone. It sets out what each Chinese certificate actually proves, how to choose the instrument for each asset, how ownership is determined for employees and contractors, and what evidence has to exist before a dispute rather than after it.

What a Chinese certificate proves, and what it does not

Four certificates matter in an AI company, and each proves something narrower than the holder usually assumes.

A software copyright certificate from the China Copyright Protection Center proves that a particular code base was deposited on a particular date by a particular applicant. Registration is voluntary, copyright arises when the code is written, and there is no substantive examination. The certificate is treated as preliminary evidence of ownership, which is enough to shift the argument in litigation and enough to satisfy an app store, a grant assessor or a procurement process. It does not prove that the code is original, that the applicant wrote all of it, or that the applicant owns the parts written by contractors.

The deposit mechanics explain the limit. Registration requires source code excerpts, usually the first and last thirty pages at fifty lines per page, together with documentation and ownership information. A code base of two hundred thousand lines is represented by three thousand, and the choice of which three thousand is a drafting decision. Companies that deposit the build scripts and the interface files hold a certificate that describes the least valuable part of the product. Companies that deposit the inference engine may be disclosing, in a government record, the part they intended to keep secret. The workable approach is to register the modules that are stable, presentable and non-core, and to keep the core in trade secret form.

A trademark registration from CNIPA proves that the applicant holds the mark for the listed goods and services. China grants the mark to the first applicant rather than the first user, so the certificate is also a statement about who moved first. The classes matter as much as the mark: an AI company usually needs the class covering software, the class covering platform and technical services, and, where the product reaches consumers, the class covering the retail and advertising use of the brand. A registration in one class does not stop a competitor from using the same sign in another.

A patent grant from CNIPA proves that the claimed technical solution was examined and allowed. It does not prove that the patent is valid, that the claims cover the competitor’s product, or that the patentee has the right to sue. Validity is tested in an invalidation proceeding, scope is tested in an infringement analysis, and standing depends on the chain of title being intact.

A trade secret has no certificate at all. What the owner holds is a body of evidence: a register of secret points, access records, signed acknowledgements and a documented control environment. That evidence is what a court examines when the claim is brought, which is why trade secret protection is the one category in which the paperwork is the right.

Choosing the instrument: four questions

The choice between patent, copyright, trade secret, design and trademark is not a matter of preference. It follows from how the asset can be copied.

QuestionIf the answer is yesIf the answer is no
Will the competitor’s product reveal how it works?Patent, because secrecy will not survive the marketTrade secret, because disclosure buys nothing
Can the asset be reverse engineered from a shipped device or a public API?Patent, and treat the shipped artefact as publicTrade secret, with technical controls
Is the value in the expression rather than the function?Copyright, registered for the evidencePatent or trade secret, depending on the mechanism
Is the asset a sign under which the product is sold?Trademark, filed early and in the right classesNo filing needed
Is the asset the appearance of a device or an interface?Design patent, which protects appearance rather than functionPatent or copyright
Will the asset be licensed to a counterparty who will inspect it?Patent, because the licence is easier to value and to auditTrade secret, with a licence that defines access

Two answers are worth spelling out because they are counter-intuitive for software companies. The first is that a model which is served behind an API is not automatically secret. Systematic querying can reconstruct behaviour, and a competitor who observes input and output pairs can train a substitute. Secrecy in that setting depends on rate limits, monitoring and contractual terms, not on the absence of a download. The second is that a patent on the training method does not protect the weights. It protects the method, and a competitor who trains differently but reaches the same performance has not infringed. Companies that want to stop that competitor need the trade secret, and companies that want to license the technology need the patent.

The combination that fits most AI products is a small patent family on the parts that can be seen, copyright registration on the code that can be shown, trade secret on the weights and data, and a trademark on the name. The portfolio map is set out in more detail in the strategy guide; this article is about the sequence and the paperwork that makes each element hold.

What counts as disclosure, and what it costs

Chinese novelty is destroyed by public disclosure anywhere in the world before the filing date, and there is no general grace period to repair it. The categories are broader than a founder’s intuition. A published paper destroys novelty. So does a conference presentation, a thesis defence, a product sold without restriction, a webinar, a public repository and a model released as open weights. An investor deck shown under a confidentiality agreement is not public; the same slide shown at a public pitch competition is.

The AI-specific complication is that one event can be a disclosure of one asset and not of another, because a disclosure only destroys novelty in what it makes available. A demonstration that shows the interface and the output, without revealing the training objective or the data pipeline, may leave the method claim intact while destroying any claim to the interface design and any trade secret in the demonstration itself. A paper that describes the architecture and the loss function discloses the method, and no later application can recover it. The analysis therefore runs asset by asset, asking what the audience learned and which right that knowledge defeats.

Four events cause most of the damage, and each can be controlled. The first is the public talk. A technical meet-up in Shenzhen is a friendly audience and a public disclosure, and the recording circulates. The second is the investor process. A term sheet negotiation with a strategic investor who is also a competitor is a disclosure risk that a confidentiality agreement reduces but does not eliminate, because the agreement binds the recipient rather than the audience the recipient later addresses. The third is hiring. An interview that walks a candidate through the architecture is controlled if the candidate has signed a confidentiality agreement; the same conversation at a conference is not. The fourth is release. Publishing weights is publication, and no licence can recall what has been downloaded.

The control that works is a disclosure register with an approval gate. Every external presentation, paper, repository push, customer pilot and interview involving technical detail is logged with the date, the audience, the asset concerned and the approval. Before the gate opens, counsel confirms whether a filing is needed and whether the material contains a secret point. The register takes an hour a month to maintain and answers, years later, the question that otherwise cannot be answered: what did the company disclose, to whom, and when. In a dispute about validity, that record is the difference between a patent that stands and one that is invalidated by the company’s own marketing.

Employment: how service inventions are actually allocated

Inventions made by an employee in the course of work, or mainly using the employer’s material and technical resources, are service inventions that belong to the employer. The rule is broad, and it is broader than most foreign founders expect, because “in the course of work” is interpreted by reference to the employee’s duties and the employer’s business rather than to the hours in which the idea arrived.

Three scenarios account for most of the disputes. The engineer hired to build the recommendation system who improves it at home over a weekend is within the rule, because the work relates to the duties. The engineer hired to run infrastructure who invents a training accelerator on the side may be outside it, because the invention is not related to the duties and no employer resources were used, which is why the company needs a separate assignment for anything it wants to hold. The founder who built the prototype before the company existed is outside the rule entirely, because there was no employer, and the assignment from the founder to the entity has to be a document rather than an assumption.

The employer’s rights come with obligations. The employer must pay the inventor a reward when the patent is granted and remuneration when the invention is exploited, and the amounts may be fixed in internal rules adopted through the procedure the law requires and communicated to staff. The right to be named as inventor stays with the individual and cannot be contracted away. Omitting the engineer who actually conceived the solution is a defect that can be raised against the patent later, and it is a common one in companies where the patent was drafted by whoever was available.

There is a tail that catches companies on both sides of a hire. An invention made within one year of an employee leaving, which relates to the work the employee did or the tasks the employer assigned, is treated as a service invention of the former employer. For a hiring company, that means the engineer who joins from a competitor may bring a claim with them, and the model they build in the first year may be exposed. For a company losing staff, it means the one-year window is worth documenting rather than assuming.

Non-compete obligations are a separate instrument and a weaker one than companies assume. Chinese law allows them for senior staff, technical staff and others with confidentiality duties, but the maximum term is two years and the employer must pay monthly compensation during the period. Courts scrutinise the scope, the compensation and the necessity, and an agreement without compensation is difficult to enforce. A non-compete is not a substitute for access control: the engineer who has already copied the weights is not stopped by a clause that binds them after they leave.

The documents that do the work are unglamorous. An employment contract that defines the duties by reference to the actual role and the employer’s business. An invention assignment clause that covers work product, inventions, improvements and data, in present-tense terms. A confidentiality clause that identifies the categories of confidential information rather than reciting a formula. A remuneration policy that has been adopted and shown to staff. An onboarding record confirming that the employee received and signed all of it. Each of these is cheap at hiring and expensive to reconstruct.

Contractors, joint development and acquired code

Contractors are where the chain of title most often breaks, because the default rule is not what foreign companies expect. Copyright in commissioned work belongs to the contractor unless the contract says otherwise, and a development agreement that says the contractor “will deliver the software” does not transfer the copyright in it. The assignment has to be express, and it should be drafted in present-tense terms so that the rights pass on creation rather than on a future promise to assign.

Four clauses belong in every development, annotation and data-labelling agreement. An express assignment of all intellectual property in the deliverables and in anything created in the course of the engagement. A waiver of moral rights to the extent the law permits, because the right of attribution and the right of integrity can otherwise survive the assignment. A warranty that the deliverables contain no third-party code or data except as disclosed, with an indemnity behind it. A confidentiality and access clause that covers the contractor’s own staff and subcontractors, because the annotation vendor’s subcontractor is inside the perimeter whether or not the contract says so.

University collaborations need a different route. Inventions made by academic staff using university resources belong to the university, and the professor cannot assign what the institution owns. The licence or assignment has to come from the technology transfer office, and it usually comes with terms about publication, equity, milestone payments and the university’s retained right to use the technology for research. A spin-out that has been operating on a professor’s assurance discovers the problem at its first financing, when counsel asks for the assignment and there is none.

Joint development is where control is lost most quietly. The default position gives each party a share, and each joint owner may exploit the result and grant non-exclusive licences. That means a Chinese partner can license the jointly developed model to a competitor and remain within the law. Exclusivity, field of use, prosecution control, cost sharing, enforcement and improvements all have to be written into the agreement, because none of them is implied. A memorandum of understanding that says the parties will “jointly own the results and cooperate on commercialisation” allocates nothing and postpones the argument to the point where the asset is valuable.

Acquired code brings its own history. A repository with two hundred contributors has two hundred potential claims unless the contributions were made under terms that assigned them, and the open-source components inside it carry licences that the acquisition may not be able to satisfy. The diligence question is not whether the code works but whether the seller can show how each part came to be owned.

The evidence file for enforcement

Enforcement in China is not slow because the courts are slow. It is slow when the plaintiff arrives without the documents that make a quick remedy available. Four routes exist, and each has a documentary precondition.

RouteForumWhat it requiresRealistic timing
Civil action with behaviour preservationPeople’s court, intellectual property tribunalA registered right or a documented trade secret, evidence of infringement, and a bondWeeks to a first ruling on preservation; a year or more to judgment
Evidence preservationPeople’s courtSpecific identification of where the evidence sitsDays to weeks, if the application is precise
Administrative complaintMarket regulation authorities, or the copyright administrationA registration certificate and evidence of the infringing conductWeeks to months
Criminal reportPublic security and the procuratorateLoss or gain at the thresholds applied in practice, and evidence of the takingMonths, and discretionary

The evidence file that supports all four is assembled before the dispute. It contains the registration certificates and the patent grant, the chain-of-title documents showing that the entity bringing the action owns the right, the trade secret register and the access logs, the version history showing what existed when, the signed acknowledgements from staff and contractors, and the records of any disclosure that was authorised. A company that has these can move in days. A company that has to reconstruct them spends the first months of the dispute on its own paperwork while the infringer keeps selling.

Two evidential points are specific to this field. Online evidence, including a competitor’s model output, a repository or a product page, is usually captured by a notary or through a court-recognised platform, because a screenshot alone is easy to challenge. And in trade secret litigation, once the plaintiff presents reasonable evidence of infringement, the burden can shift to the defendant to explain where its information came from, which is why the register and the logs matter: they are what convert a suspicion into a case that survives the first hearing.

Documents, filings and timeline

The work fits into roughly ninety days if it is run as a project rather than left to whoever has time.

StepAuthorityTypical timingWhat foreign companies miss
Assignment and confidentiality pack signed by founders, staff and contractorsInternalBefore any work begins, and as a condition of hiringThe offer letter is used as the contract, and it contains no assignment
Invention harvest and freedom-to-operate searchInternal, against the CNIPA databaseTwo to four weeksThe search covers English abstracts only, so Chinese-language prior art is missed
Filing decision, recorded in writingInternal, with patent counselBefore the next external disclosureThe decision is made by default, and the default is to file nothing
Patent applicationCNIPAThe filing date fixes priorityFiled after the conference paper, and the disclosure cannot be undone
Software copyright registrationChina Copyright Protection CenterSeveral weeks after filingThe deposit captures the interface and leaves the inference engine uncovered
Trademark applicationCNIPACommonly nine to twelve monthsFiled after the Chinese launch, by which time a local applicant has filed
Open-source and model licence scanInternalBefore the first external buildThe scan covers direct dependencies and misses the container
Evidence file assemblyInternalBefore any dispute, and refreshed annuallyNobody owns it, so it is assembled under time pressure

Where foreign companies get this wrong

The first mistake is hiring on an offer letter. The letter states the salary, the title and the start date, and it says nothing about inventions, confidentiality or data. The engineer is competent, the model is built, and the company later discovers that the assignment it needs was never signed. Reconstructing it requires the employee’s cooperation, which is cheapest when they are still employed and most expensive when they are not.

The second is leaving the founder’s prototype outside the company. The model that won the first customer was built before incorporation, on the founder’s own laptop, and it was never assigned to the Chinese entity. At financing, the diligence report identifies a core asset held by an individual, and the fix requires an assignment, a valuation question and, sometimes, a tax analysis.

The third is signing a joint development memorandum that allocates nothing. The parties agree to cooperate, to share the results and to discuss commercialisation later. Later arrives with a valuable model, two parties who each believe they own it, and no clause that resolves the question.

The fourth is relying on a non-compete instead of access control. The company has a two-year non-compete with the technical lead, pays no compensation during the period, and has never reviewed who can download the weights. When the lead leaves, the clause is difficult to enforce and the download is already done.

The fifth is registering in the wrong name. The copyright certificate or the patent is filed in the name of the overseas parent, because that is where the group’s intellectual property sits, and the Chinese subsidiary then cannot use the registration for the purposes that require a Chinese holder. The correction is an assignment and a recordal, and it should have been a decision rather than a default.

Cost and commercial impact

The instruments differ in cost by an order of magnitude, and the cheapest ones carry the most weight in diligence. A software copyright registration and a trademark application are inexpensive and produce certificates that a court, an app store and an investor accept. A patent family is expensive, because the cost sits in drafting, translation, prosecution and annuities rather than in official fees, and a family maintained across several jurisdictions costs more each year it lives. Trade secret protection costs almost nothing in fees and a great deal in management attention, because the controls have to be operated rather than filed.

The commercial impact shows up in three places. In financing, the intellectual property section of the diligence report is read by people pricing risk, and a clean chain of title moves the price in one direction while a missing assignment moves it in the other. In licensing, a counterparty will pay for a right that can be described, registered and enforced, and will discount a right that depends on a document nobody can find. In a dispute, the difference between an injunction in weeks and a judgment in two years is usually the evidence file rather than the merits.

Practical checklist

The firm advises foreign AI companies on the ownership, registration and evidence work that has to happen before a model is disclosed. The relevant service page is intellectual property, and companies that are still deciding on an entity structure should read the note on incorporation first, because the entity that holds the rights is a decision rather than an afterthought. Two related guides go further: intellectual property strategy for AI companies sets out the portfolio and prosecution routes, and the note on the Shanghai prompt copyright case explains how the courts are approaching authorship in AI-assisted work.

Frequently Asked Questions↑ Back to article

Is software copyright registration required to own code in China?

No. Copyright arises when the code is created, and registration with the China Copyright Protection Center is voluntary. Registration matters because it gives you a certificate that is treated as preliminary evidence of ownership, which is useful in litigation, app-store reviews and grant applications, but it does not cure a weak chain of title.

Can an AI algorithm be patented in China?

Yes, if the claim is framed as a technical solution that solves a technical problem with technical means and produces a technical effect. A pure algorithm, business rule or mathematical method is not patentable on its own. The workable approach is to tie the algorithm to a concrete technical process, device or system and to claim the combination.

How long does a Chinese invention patent take?

Typically two to three years from filing to grant, depending on the field and the number of office actions. Some local intellectual property protection centres offer accelerated examination for eligible technologies, which can shorten the process. Plan patent filing before public disclosure or product launch, because China has no broad grace period.

Can model weights be protected as trade secrets?

Yes, if the weights are not publicly available, have commercial value because they are secret, and you have taken reasonable confidentiality measures. Those measures normally include access control, encryption, logging, confidentiality clauses with employees and contractors, and clear exit procedures. Without documented measures, a trade secret claim is difficult to sustain.

Do open-source licences matter for a product sold in China?

They matter, and Chinese courts have enforced open-source obligations in reported decisions. Permissive licences such as MIT and Apache 2.0 mainly require notice retention, while copyleft licences such as GPL and AGPL can require you to license derivative code on the same terms. Run a licence scan and keep a software bill of materials before shipping.

Who owns IP created by our Chinese employees and contractors?

Inventions made by employees in the course of their work generally belong to the employer as service inventions, but the employer must still pay statutory inventor remuneration. Copyright in work by external contractors may not transfer automatically, so you need a written assignment. For joint development, default rules allow each party to exploit the result unless the contract says otherwise.

What is the fastest way to stop a competitor copying our model?

There is no single fast route. A combination usually works best: a civil injunction application with evidence preservation, an administrative complaint where infringement is clear, and a criminal report if the conduct looks like trade secret theft. Speed depends on having registered rights, documented secrecy measures and evidence ready before you file.

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