The core assets of an AI company are not its desks and chairs but its algorithms, models, data and code. Yet many AI start-ups discover only at the diligence stage of a financing that the algorithm is unprotected by patent, that the provenance of the training data cannot be established, that the core code is not subject to any trade secret measures, and that the open-source licence under which the model was obtained imposes restrictions on commercial use.
Patent protection for AI algorithms
The revised examination standards applicable from 2024
In January 2024 the revised Guidelines for Patent Examination introduced more detailed standards for the examination of AI-related patents. In December 2024 the China National Intellectual Property Administration issued the Guidance on Patent Applications for AI-Related Inventions (Trial), which provides operational guidance on the drafting of AI patent applications.
Points of examination:
| Dimension | Requirement |
|---|---|
| Sufficient disclosure | Because AI algorithms exhibit “black box” characteristics, the specification must contain enough information to satisfy the sufficiency requirement |
| Public order | No patent shall be granted for an invention that contravenes the law, social morality or public interest |
| Data compliance | No patent shall be granted where the collection of data, the management of labels or the setting of rules contravenes the law |
| Evidence where necessary | Evidence of the inventive contribution may be required where necessary |
Patent portfolio strategy
- Layered filing — the functional layer (algorithm and model), the foundational layer (data architecture) and the interface layer (human-computer interaction) should be designed and filed separately.
- PCT plus national phase — priority should be given to jurisdictions with substantial market volume and a high risk of litigation.
- Freedom-to-operate analysis — an FTO analysis should be introduced at project inception rather than at launch.
- Competitor monitoring — a patent watch mechanism should be established and reviewed at fixed intervals.
Copyright in training data and the fair use question
The core controversy
Article 24 of the Copyright Law sets out thirteen exhaustive categories of fair use. Whether the use of works for model training falls within those existing categories remains contested in both academic commentary and judicial practice.
Compliance points
| Source of data | Compliance risk | Response |
|---|---|---|
| Crawling of public web pages | Copyright infringement | Review robots.txt and copyright notices |
| User-generated content | No authorisation for training | Express provision in the user agreement |
| Purchased datasets | Opaque provenance | Require the supplier to provide provenance evidence |
| Open-source datasets | Licence restrictions | Review each licence individually |
Legal basis: Article 7 of the Interim Measures for the Administration of Generative AI Services requires that training data be “obtained from lawful sources” and that, where intellectual property is involved, the lawful intellectual property rights of others not be infringed.
Ownership of copyright in AI-generated content
Judicial precedents
Beijing Internet Court, the “AI text-to-image” case (first instance 2023, final appeal 2024): the user generated images through repeated adjustment of prompts and parameters, which was held to reflect “personalised expression and intellectual input” and to satisfy the originality requirement. Infringement was established and RMB 500 was awarded. The case confirms that a user of AI-generated content may enjoy copyright.
Shanghai, the first AI large-model copyright infringement case (2025): the defendant used an AI model to generate character images associated with the novel Battle Through the Heavens and was ordered to pay RMB 50,000 (RMB 30,000 in economic loss and RMB 20,000 in reasonable expenses).
Principles of ownership
- The AI model is not a legal person and cannot be an “author”.
- A user who contributes original labour — prompts, parameter settings, selection of outputs — may be recognised as the creator.
- Platform and user may allocate copyright by contract.
Software copyright registration
Registration process
| Step | Time |
|---|---|
| Account registration and identity verification | 1 day |
| Online filing and upload of materials | 1–2 days |
| Examination | 60 working days |
| Issue of certificate | Upon approval |
Value of registration: software copyright registration is a precondition for recognition as a high-tech enterprise and a standard asset in financing diligence.
Trade secret protection
Trade secret inventory for an AI company
The following assets should be brought within the scope of trade secret protection:
- Proprietary training methods
- Industry datasets
- Model parameters and weights
- Prompt libraries
- Technical management specifications
- Core code repositories
Protective measures
| Measure | Implementation |
|---|---|
| Inventory of secret points | Identify and document each algorithm, dataset and code base, avoiding vague “blanket confidentiality” |
| Access control | Tiered permissions over code repositories, with strict control over access to core assets |
| Audit trail | Data loss prevention, access logging and anomaly monitoring |
| Incident response | Preservation of evidence immediately upon discovery of infringement |
| Digital controls | Tiered permissions, data masking and encrypted storage |
Open-source model licence compliance
Comparison of mainstream model licences
| Model | Licence | Commercial use | Key restrictions |
|---|---|---|---|
| Llama 3 | Meta Community Licence | Permitted | Separate application required above 700 million monthly active users; training of competing models prohibited; derivative models must carry the “Llama” name |
| Qwen series | Apache 2.0 | Permitted | Licence notice and NOTICE file must be retained; closed-source use permitted |
| MIT | MIT Licence | Permitted | Copyright notice must be retained |
| DeepSeek | DeepSeek Licence | Permitted | Specific terms must be observed |
The “open-washing” trap
Some models are described as open-source but attach conditions that restrict:
- The right to distribute as a service
- Brand association
- The naming of derivative models
- Large-scale commercial use
Recommendation: before adopting an open-source model, review each restriction on commercial use in the licence, clause by clause.
Representative disputes
| Case | Court | Point |
|---|---|---|
| Four illustrators v Trik AI | Beijing Internet Court | Whether the use of artistic works for AI training constitutes fair use |
| AI text-to-image case | Beijing Internet Court | A user who contributes original labour may enjoy copyright |
| Generative AI platform infringement | Guangzhou Internet Court | The world’s first judgment against a generative AI platform for infringement |
| AI voice cloning | Beijing Internet Court | China’s first finding of infringement of personality rights by AI-generated voice |
| AI illustrator criminal case | Beijing Tongzhou Court | Reproduction of artistic works by AI for a profit of RMB 270,000 held to constitute a criminal offence |
Frequently asked questions
Q: Can an AI algorithm be patented? A: Yes. The revised Guidelines for Patent Examination issued in 2024 set out detailed standards for AI patents. The essential requirement is that the specification disclose the technical details of the algorithm sufficiently; a statement of concept alone will not suffice.
Q: If we build a commercial product on an open-source model, must we open-source our code? A: It depends on the licence. Apache 2.0, under which Qwen is released, does not require derivative code to be open-sourced, whereas GPL-family licences do. Llama 3 imposes a 700 million monthly active user threshold and restrictions relating to competing products. The licence should be reviewed before use.
Q: If an employee leaves and takes training data with them, is that a trade secret infringement? A: Where the training data satisfies the three elements of a trade secret — not known to the public, commercially valuable, and subject to confidentiality measures — a claim in trade secret infringement may be brought. An inventory of secret points and an access control system should be established in advance.
This article is prepared on the basis of the Patent Law, the Copyright Law, the Anti-Unfair Competition Law, the Interim Measures for the Administration of Generative AI Services and the revised Guidelines for Patent Examination issued in 2024. For assistance with an intellectual property strategy for an AI company, please contact Geason Law Offices.