On 7 September 2026, China’s Supreme People’s Court (SPC) issued its Opinions on the Lawful Adjudication of Disputes Involving Artificial Intelligence (Fa Fa [2026] No. 10, the “Opinions”). The 24 provisions offer systematic guidance on liability, intellectual property, data use and evidence in AI-related disputes. SPC announcement and full text (Chinese).
For businesses, the practical significance lies in how legal scrutiny will focus on identifiable parties, actual conduct and verifiable evidence. Companies need to assess whether an output infringes third-party rights and be able to explain where relevant data came from, what people contributed, and what measures they took when risks emerged.
Liability depends on fault and the ability to control risk
Paragraph 3 establishes fault-based liability as the default where legislation does not expressly provide for strict liability or a presumption of fault. Relevant factors include the application context, technical transparency, potential risks, preventive measures, and the parties’ ability to foresee and control harm.
The fact that AI generated the content does not, by itself, excuse the parties involved. Equally, developing or supplying a model does not automatically make a business liable for every output. In its official explanation, the SPC stresses the need to avoid excessive burdens on innovation and to align responsibility with control and duties of care. SPC press briefing Q&A (Chinese).
Consider a company using a third-party model to create advertising images. An employee uploads another party’s work and repeatedly requests similar images. A dispute would require separate consideration of the user’s instructions, the company’s publication of the output and the service provider’s role. Contractual allocations of responsibility can help manage the relationship, but cannot replace a legal assessment of the conduct involved.
Businesses should therefore assign clear responsibilities for reviewing input materials, checking generated content and handling complaints.
Training remains contested; a defence needs supporting evidence
For AI-generated content that infringes copyright, paragraph 12 identifies factors including the source of training data, each party’s involvement, preventive measures and financial benefit. A developer raising a non-infringement defence must substantiate it with evidence concerning training data sources, training records, model operation and relevant technical grounds.
This strengthens evidentiary requirements for the party holding technical information. It does not settle the legal characterization of training on copyright works without permission. The SPC expressly states that this issue, and the copyrightability of AI-generated content, remain contested and are not resolved by the Opinions. SPC press briefing Q&A (Chinese).
Public accessibility should not be treated as unrestricted permission to train. Conversely, identifying that a work was used does not establish the full extent of liability without examining the relevant conduct and available defences.
As a practical measure, developers should retain records of data provenance, permission scope, processing and model versions, tailored to the data and uses involved. Such records can support a specific defence and should be managed consistently with trade-secret and personal-information protections. The Opinions do not impose an unconditional duty to disclose all training data to the public.
Human contribution matters to both protection and infringement
A business using AI to produce images, copy or designs faces two separate questions: whether the output infringes someone else’s rights, and whether the business can claim rights in its own result.
On infringement, paragraph 12 addresses users who know or ought to know of an earlier work and use AI to generate a substantially similar work without a valid defence. Prompts, reference materials and revision history may help establish relevant facts. Entering an author’s name or a style description alone does not establish every element of an infringement claim.
On protection, the SPC’s earlier Implementation Plan for Judicial Protection of Intellectual Property by People’s Courts (2026–2030) directs courts to examine the individual’s instructions, selection and modification process when deciding whether generated content reflects original human choices and expression. Implementation Plan, paragraph 9 (Chinese).
Businesses should retain drafts, reasons for selecting particular results and revision records that demonstrate creative contribution, alongside the final file. A high number of operations does not automatically meet the legal threshold. The value of the record is in showing what the person contributed to the final expression.
AI-assisted R&D and open-source use have distinct limits
Human contribution also matters in patent protection. Paragraph 14 sets out the technical-solution requirements for AI-related inventions while excluding, among other cases, those without a substantive human contribution. Inventorship requires a natural person to make a creative contribution to the invention’s substantive features. Meeting the subject-matter requirements does not establish that every condition for patent grant has been satisfied.
China’s current Patent Examination Guidelines also contain disclosure requirements for inventions involving AI model construction, training or application in a particular field. The specification must enable a person skilled in the art to implement the technical solution. CNIPA decision amending the Patent Examination Guidelines (Chinese).
R&D records should capture the formulation of technical problems, selection of solutions, experimental validation and human improvements. Keeping only the AI output may be insufficient to demonstrate inventorship or support adequate disclosure.
For open-source software, paragraph 13 allows courts to exempt qualifying developers and providers from liability where they supply certain code modules free of charge on an open-source basis and publicly explain their functions and security risks. This protects open-source collaboration, but is not a blanket exemption for all open-source models or downstream commercial services. Businesses adopting the technology should review licence conditions and assess their own fine-tuning, deployment and operational conduct.
Evidence preservation belongs in everyday workflows
Paragraphs 17 and 18 address evidence production and review. A party controlling evidence may face an adverse finding if it refuses to produce it without proper justification. When AI-generated content is submitted as evidence of infringement, courts are to consider factors including the effect of prompts, similarity, consistency across repeated tests and filtering mechanisms.
Depending on the use, businesses should preserve key records such as the model and version, complete inputs, original outputs, human changes and publication dates. These are practical recommendations, not a uniform statutory checklist for every dispute. Repeat testing may help explain the conditions under which a result arose; inability to regenerate the same content does not automatically undermine the authenticity of an earlier record.
Court submissions require a separate verification step. Under paragraph 19, participants submitting AI-generated litigation documents, case-search reports or similar materials must check the authenticity and accuracy of cited laws, judicial interpretations and cases before submission and explain their use of AI assistance when filing. Legal research should be checked against original sources. The submitting party remains responsible for its judgment.
What businesses can put in place now
The practical direction is to develop management and evidence practices proportionate to the risks of each AI use. For IP teams, the immediate priorities are to verify the rights basis for input materials, document human creative and inventive contributions, review outputs before external use, and retain evidence sufficient to reconstruct key steps. These measures can reduce infringement exposure and help businesses protect their own innovations when disputes arise.
Prepared by ONECHINE IP based on public materials available as of 8 September 2026. This article provides general analysis; individual matters require assessment of the facts, applicable law and subsequent court decisions. English renderings of Chinese document titles and provisions are for reference; the Chinese originals govern.