Standard software contract templates do not cover the risks peculiar to AI projects. An AI company needs to add specific clauses — on training restrictions, liability for hallucination, and the discontinuation of models — across five categories of core contract.
The AI development agreement: eight core clauses
Clause list
| No. | Clause | Core point |
|---|---|---|
| 1 | Ownership of intellectual property | Allocation of rights in the model architecture, training parameters and derivative works |
| 2 | Data ownership | Allocation of rights in customer data, training data and model output |
| 3 | Service levels | Definition of model accuracy, response time and availability |
| 4 | Data privacy and security | Requirements for processing, storage, transmission and destruction |
| 5 | Infringement liability | Allocation of liability for infringing output, data leakage and algorithmic discrimination |
| 6 | Discontinuation and retirement | Provision for model updates, discontinuation and replacement |
| 7 | Regulatory undertakings | Compliance with the generative AI measures and other regulations |
| 8 | Acceptance | Deliverables, acceptance procedure and handling of defects |
Key risk points
Liability for hallucination: an AI model may produce incorrect information (“hallucination”) and cause loss to the user. The development agreement must address:
- who bears loss caused by hallucination
- whether there is a cap on compensation
- whether any exclusion of liability applies
Disclosure of API dependencies: in a 2025 case before the Shenzhen Intermediate People’s Court, the core functionality of the software called a non-open-source external API without disclosing the dependency, and the source code was not delivered. The court ordered rescission of the agreement, repayment of RMB 300,000 and compensation for loss.
The model licence: five key restrictions
| Restriction | Description | Typical case |
|---|---|---|
| Commercial authorisation | Whether commercial use is permitted | Permitted under Apache 2.0; restricted under some licences |
| Monthly active user threshold | A separate application is required above the threshold | Llama 3 requires an application to Meta above 700 million monthly active users |
| Competing-product restriction | Training competing models is prohibited | Expressly prohibited under Llama 3 |
| Naming of derivatives | Naming rules apply to derivative models | Llama derivatives must carry the “Llama” name |
| Background IP | Scope of the licence over the supplier’s background IP | Term and assignability must be specified |
The data processing agreement: eleven mandatory clauses
Legal basis
The 2025 model text of the Data Processing Services Contract issued by the National Data Administration and the State Administration for Market Regulation, together with Article 28 of the GDPR.
Mandatory clause list
| No. | Clause | Core requirement |
|---|---|---|
| 1 | Purpose and term | The purpose, scope and term of processing must be specified |
| 2 | Subject matter | The type, category and volume of data must be listed |
| 3 | Security measures | Encryption, access control and audit logging |
| 4 | Personnel security | Confidentiality obligations of personnel with access |
| 5 | Sub-processing | No sub-processing without consent |
| 6 | Return and deletion | Written confirmation of deletion within 30 days of termination |
| 7 | Breach notification | Immediate notification of any leakage |
| 8 | Cross-border transfer | Handled in accordance with the cross-border data rules |
| 9 | Liability | Specific performance, remediation and compensation |
| 10 | Audit rights | The controller may audit the processor |
| 11 | Assistance with data subject rights | Assistance in responding to requests from data subjects |
API terms of use: nine core clauses
| Clause | Point |
|---|---|
| Scope of use | Call scope, quotas and call frequency |
| Authentication | API key, OAuth 2.0 and other methods |
| Service levels | Availability, response time and service credits |
| Content safety | Content moderation and filtering policy |
| Data use | Whether the API provider may use customer data to train models |
| Intellectual property | Ownership of API output |
| Limitation of liability | Liability for loss arising from API failure or incorrect output |
| Termination and change | Notice period for discontinuation or change, and transitional arrangements |
| Regulatory undertakings | Compliance with applicable law |
Important note: calling an offshore large-model API, such as OpenAI, to serve the domestic public requires not only API terms but also algorithm filing, large-model registration and a data export security assessment. In a Chongqing case, direct calls to an offshore API resulted in a fine of RMB 30,000 to 50,000.
The AI SaaS contract: ten special clauses
Traditional SaaS clauses do not cover the risks peculiar to AI. The following clauses must be negotiated separately:
| Clause | Risk |
|---|---|
| Restriction on training | A conventional “improvement of the services” clause may be read as permitting the use of customer data to train AI |
| Ownership of output | Intellectual property in AI-generated content |
| Liability for hallucination | Compensation for loss caused by incorrect output |
| Explainability | The supplier should provide an explanation of model behaviour |
| Bias and discrimination | Monitoring and mitigation of bias by the supplier |
| Discontinuation and replacement | Notice period, data migration and replacement arrangements |
| Compliance certification | SOC 2, ISO 27001 and similar |
| Audit rights | Audit of processing activities |
| Data sovereignty | Location of storage and restrictions on transfer |
| Cap on compensation | Specific provision for AI-related risk |
Key risk
The “hidden training licence” trap: conventional SaaS wording such as “improve”, “build upon” or “enhance” may be construed as permitting the supplier to use customer data to train AI models. The customer must restrict this expressly in the contract.
Contract disputes
Case one: AI digital human development agreement (Shenzhen Intermediate People’s Court, 2025)
Facts: the core functionality of the software called a non-open-source external API; the dependency was not disclosed; the source code was not delivered. Decision: rescission of the agreement, repayment of RMB 300,000 and compensation for loss. The shareholder of the one-person company was held jointly liable.
Case two: AI server computing power sale (Beijing Huairou Court, 2026)
Facts: the buyer paid a deposit of RMB 3 million and then failed to pay the balance; the seller removed and resold the servers. Decision: the buyer was ordered to pay RMB 16.4 million.
Lessons
- API calls must be expressly provided for in the contract — source, open-source status and commercial authorisation.
- The obligation to deliver source code must be stated expressly.
- Liquidated damages and the scope of compensation must be specified.
- The shareholder of a one-person company must attend to the independence of company property.
Frequently asked questions
Q: Can a standard SaaS contract be used for an AI product? A: No. A standard SaaS contract lacks the AI-specific clauses on training restrictions, hallucination liability, discontinuation and bias monitoring. Used unchanged, it may permit customer data to be used to train the supplier’s models.
Q: When must a DPA be signed? A: Before processing begins. Where personal information is processed, a DPA is a statutory requirement rather than an optional clause.
Q: If we call a third-party API and our customer data is used for training, what can we do? A: The API terms should state expressly that the provider may not use customer data for model training. Anthropic and others have undertaken not to train on data of customers of paid services. Where training has already occurred, a claim for breach of contract may be brought.
This article is prepared on the basis of the 2025 model DPA text issued by the National Data Administration and the State Administration for Market Regulation, Article 28 of the GDPR and the judicial decisions referred to above. For assistance with the review or drafting of contracts for an AI company, please contact Geason Law Offices.