Written by Mr Charles Feng, Director of the International Business Department and a Senior Partner of Tahota (Beijing) Law Firm, IP Expert and collaborator of the China IP SME Helpdesk.
Abstract: On September 7, 2026, the Supreme People's Court issued the Opinions of the Supreme People's Court on Legally Adjudicating Disputes Involving Artificial Intelligence (hereinafter referred to as the "Opinions"), marking China's first systematic judicial adjudication document specifically addressing AI-related disputes. Comprising five parts and twenty-four articles, the Opinions are grounded in existing legal frameworks such as the Civil Code, the Copyright Law, and the Personal Information Protection Law. They respond to a range of complex issues emerging in judicial practice during the era of generative AI, including infringement of personality rights, AI hallucination torts, boundaries of training data, copyright of AI-generated content, liability for autonomous driving, and evidentiary rules. The Opinions elevate scattered effective case law precedents from internet courts and intellectual property courts nationwide into unified adjudication guidelines applicable across the country. This paper, combining current statutory laws and typical domestic effective judgments, conducts an in-depth commentary from the perspectives of legislative cohesion, substantive tort rules, intellectual property adjudication logic, procedural evidentiary mechanisms, institutional limitations, and future paths for improvement. It clarifies the legal positioning of AI as a technological tool, analyzes the implementation path of the "equal emphasis on development and security" principle in individual judicial cases, reflects on the interpretive space still inherent in the rules, and provides theoretical references for judicial application and industrial compliance.
Keywords: Generative Artificial Intelligence; Deepfake; AI Hallucination; Training Data; Copyright; Fault Liability
Introduction
The iterative advancement of large model technology has comprehensively integrated artificial intelligence into all aspects of social production and daily life. New types of disputes, such as AI face-swapping, AI voice-cloning for commercial promotion, large-scale web scraping by large models for training data, dissemination of false information via AI hallucinations, AI-generated text and graphic works, and accidents involving intelligent assisted driving, have flooded into people's courts. Prior to the issuance of the Opinions, China lacked specialized judicial interpretations to uniformly govern AI-related disputes. Adjudication relied primarily on judges applying traditional civil laws through expansive interpretation, leading to discrepancies in adjudicative standards across different regions: courts held varying understandings of the "reasonable scope" regarding the use of publicly available personal information for large model training; in cases of reputational infringement caused by AI hallucinations, the boundaries of platform obligations fluctuated; and opinions diverged significantly on the copyrightability of AI-generated content and whether training data constituted fair use.
The Personality Rights Chapter of the Civil Code has already established principled provisions against deepfakes. The Personal Information Protection Law regulates the handling of personal information. The Interim Measures for the Management of Generative Artificial Intelligence Services imposes obligations on service providers from an administrative regulatory perspective. However, administrative regulations cannot directly serve as the basis for adjudicating civil torts, and the judicial level lacks unified substantive and procedural adjudication guidelines. Judicial practice has long been in a state of "applying old laws to new technologies." While individual cases have generated numerous typical precedents, a top-down supply of unified rules has been absent.
Against this backdrop, the Supreme People's Court issued the twenty-four-article Opinions. Adhering to the fundamental legal principle that artificial intelligence does not possess civil subject status, the Opinions establish three major principles: people-oriented development, support for innovation-driven development, and the reinforcement of safety guardrails. They integrate personality rights, personal information, consumer rights, intellectual property, product liability, and civil procedure evidence rules, marking a transition from case-by-case exploration to a systematic system of adjudicative rules. The Opinions do not create new civil rights but rather interpret and refine existing laws within the context of AI scenarios, functioning as a judicial interpretation document for the digital technology sector. This paper will analyze the rules' alignment with current laws, examine their highlights and practical difficulties in application, identify institutional gaps, and offer reflections on judicial application and industrial compliance, referencing domestic effective judgments.
I. Institutional Positioning of the Opinions: Translating Technology Scenarios within the Existing Legal Framework
(I) No Creation of Legal Subjects, Adhering to the Instrumentalist Theory
Article 2 of the Opinions explicitly stipulates that artificial intelligence itself does not possess civil subject status, and relevant legal liabilities shall be borne by civil subjects such as developers, service providers, and users. This constitutes the logical cornerstone of the entire document. Within China's civil legal system, civil subjects are limited to natural persons, legal persons, and unincorporated organizations. AI algorithms and large models are merely technological tools used by humans and cannot independently enjoy rights or assume obligations. This legal principle has been consistently affirmed by domestic effective judgments. In the nation's first case involving torts caused by AI hallucinations, the Hangzhou Internet Court ruled that automatically generated output from AI promising compensation does not create legally binding obligations, and that generative AI is a service rather than a product, meaning algorithmic output cannot be equated with a legal entity's expression of intent.
(II) The Value Balance of the Three Fundamental Principles: Rights Protection, Risk Prevention and Control, and Technological Innovation
"Supporting innovative development and reasonably accommodating the space for technological trial and error" does not equate to technological immunity. The inherent limitations of technology can serve as factors in assessing fault but cannot independently constitute a defense. This directly addresses AI hallucination disputes: while probabilistic output errors are an objective reality of large models, service providers cannot use "AI hallucination" as a blanket exemption from liability. They must still fulfill obligations related to algorithm control, risk warnings, and content review. This is consistent with the adjudicative logic in the Hangzhou Internet Court's AI hallucination case, where the court did not demand that the platform guarantee 100% accuracy of every generated output but did require the platform to provide risk warnings and employ industry-standard technical measures to reduce the probability of erroneous outputs.
(III) The Relationship between the Opinions and Existing Legal Norms
• Personality Rights: Based on Article 1019 (prohibition of deepfakes) and Article 1023 (protection of voice rights) of the Civil Code, the Opinions extend to protecting the personality interests of deceased natural persons and the application of injunctions against personality rights infringement.
• Intellectual Property: Relying on the principles of originality and fair use under copyright law, rules for open-source license contracts, and standards for patent authorization and confirmation, the Opinions adapt to the industrial realities of large model training, AI-generated content, and open-source large models.
• Procedural Rules: Under the framework of the Civil Procedure Law, the Opinions improve mechanisms for ascertaining technical facts, reviewing AI-generated evidence, and preservation measures, while regulating acts of using AI to forge evidence that obstructs litigation.
II. Commentary on Adjudication Rules for Tort Disputes: Reconstructing Civil Liability Determination Beyond Technological Appearance
Article 3 of the Opinions stipulates that where AI is used to infringe civil rights and interests, and there are no special provisions for no-fault or presumed-fault liability, the principle of fault liability shall apply. Determining fault involves examining the subject's degree of control over the AI, the benefits derived, and the duty of risk prevention and control. This standard enables dynamic stratification of fault:
• Generative AI Service Providers: Those providing APIs or web services to the public bear obligations of algorithm control, risk warning, and content review.
• Users or Deployers: Those who directly operate the AI system or use its outputs bear duties of care corresponding to their level of control and the specific application scenario. Developers: Those who design algorithms and train models bear a duty of care regarding。 the inherent safety and compliance of the technology itself.
The principle of "who controls the risk and who benefits bears the corresponding duty of care" effectively resolves the past judicial tendency of "one-size-fits-all platform liability."
Article 1019 of the Civil Code already prohibits using information technology to forge others' portraits, and Article 1023 explicitly provides that the voice of a natural person is protected by reference to portrait rights. However, technological developments have spawned numerous new scenarios: using AI-synthesized voices for commercial promotion, using AI to "resurrect" the deceased, and illicit activities like AI-powered "undressing" apps. Prior to the Opinions, judicial guidance for these scenarios was lacking.
Referring to a typical effective case: In the case of Li v. Xuzhou [Anonymized] Cultural Media Company, the defendant scraped the plaintiff's public speech audio without permission and used AI to synthesize a similar voice for book promotion. The court found this infringed the natural person's voice rights and awarded substantial damages. This adjudicative spirit is reflected in the Opinions. In another case involving AI face-swapping in a short drama, the defendant swapped an actor's portrait onto a character in a short drama for commercial playback without authorization, and the court found it infringed the right of publicity.
(III) Rules for Handling Public Personal Information in Model Training Scenarios: Balancing Industrial Development and Individual Rights
This provision is one of the most closely watched articles in the entire Opinions from an industry perspective. Article 27 of the Personal Information Protection Law stipulates that handling publicly available personal information requires obtaining individual consent if it has a significant impact on individual rights. The Opinions adapt this provision to the context of large model training. In practice, the key points of contention are what constitutes a "reasonable scope" and what constitutes a "significant impact on personal rights." Drawing on judicial practice, the assessment should consider three aspects: purpose, scale, and consequences. This includes whether the purpose is general large model R&D or targeted profiling; whether the training aggregates sensitive personal information on a large scale; and whether the output can identify specific individuals or lead to personality discrimination or reputational harm.
(IV) Protection of Consumer Rights: Algorithmic Price Discrimination and AI-Impersonated Celebrity Commercial Fraud
Algorithmic price discrimination ("big data killing") has long been discussed in academia, but proving subjective malicious intent and differentiated trading conditions in individual cases poses significant evidentiary challenges. The Opinions confirm that algorithmic differential treatment causing harm should incur tort liability, incorporating algorithmic discrimination into the civil tort evaluation system. In the context of AI-impersonated celebrity endorsements, where many businesses use AI-synthesized celebrity voices or virtual images to mislead consumers into believing a celebrity is endorsing a product, this constitutes consumer fraud. Consumers can claim punitive damages under the Consumer Rights Protection Law, a rule clarified in the Opinions.
Article 7 of the Opinions states: "The fact that an AI hallucination is an inherent technical defect shall not automatically exempt a service provider from civil liability; promises or guarantees automatically output by AI do not necessarily create legally binding obligations." This article contains a dual obligation: first, platforms bear a strict duty to review illegal or harmful information; second, concerning factual errors (hallucinations), platforms are not required to guarantee 100% accuracy for all outputs but must fulfill warning and technical optimization obligations. The determination of fault is a key discretionary point for judges, allowing flexibility based on the prevailing industry technological standards.
Article 10 of the Opinions strictly distinguishes between "AI products" and "online AI services": only AI with a physical carrier (e.g., robots, autonomous vehicles, smart hardware) falls under the scope of product liability and its no-fault liability regime. Pure online generative AI services do not fall within the scope of product liability. This clarification resolves a long-standing industry confusion and prevents the misapplication of product liability rules to online algorithmic services.
III. New Intellectual Property Rules
Intellectual property is one of the most concentrated areas of AI disputes. Articles 12 to 16 of the Opinions cover five major issues: liability for AI-generated infringing content, liability for using open-source software, patent authorization and confirmation, AI technology contracts, and fair use of training data. This addresses the full chain of legal risks, from input-side training to output-side generation.
Article 12 clarifies that the fact that content is AI-generated shall not automatically exempt a party from infringement liability. If a user, knowing or having reason to know of a pre-existing work, uses AI to generate content substantially similar to that work without a fair use defense, they shall bear liability. This definitively rejects the misconception that "AI-generated" content equates to non-infringing, non-copying output. The focus of infringement analysis is no longer on whether the tool used is human or AI, but on whether the output is substantially similar to the pre-existing work and whether the actor had subjective knowledge of the pre-existing rights. Developers may use evidence such as training logs and model operation records to prove non-use of specific works as a defense. This imposes compliance requirements on large model enterprises to retain such records.
Article 13 stipulates that in disputes involving the use of open-source software in large models, courts should determine liability for breach of contract and tort based on the license agreement text, industry practices, and the acts of modification and distribution. Parties who violate open-source license conditions and disseminate derivative code shall bear civil liability.
IV. Fair Use for Training Data and AI Technology Contracts
Article 16 addresses disputes concerning the use of copyrighted works for large model training. It requires courts to balance the rights of copyright holders with industrial development, examining factors such as the purpose and manner of use, the nature of the work, and its market impact to determine whether the use constitutes fair use. Regarding AI technology contracts, the Opinions suggest that courts should differentiate between the inherent limitations of technology and a party's failure to meet explicitly agreed-upon technical specifications. It is not appropriate to simplistically deem technological limitations as a performance breach. This underscores the importance for businesses to define quantitative delivery metrics and address model limitations explicitly when drafting AI development contracts.
V. Procedural Rules and Trial Mechanisms
(I) Mechanisms for Ascertaining Technical Facts: Collaborative Role of Technical Investigators, Expert Assistants, and Judicial Appraisers
The Opinions stipulate the status and role of technical investigators, expert assistants, and judicial appraisers in participating in litigation and cooperatively ascertaining technical facts.
(II) Judicial Review of AI-Generated Evidence
Instances have occurred abroad where lawyers used ChatGPT to generate pleadings citing fictitious case law, leading to sanctions by courts. Similarly, domestic judicial practice has seen parties submit AI-generated false legal provisions and evidence. AI can forge documents, chat logs, and expert opinions at low cost, challenging the civil evidence system. The Opinions clarify that AI-generated materials do not automatically possess evidential value and cannot be directly admitted; they must undergo substantive review. Sanctions for using AI to forge evidence are also provided for, which is highly significant in practice.
(III) Coordination with Criminal Law and Supporting Trial Mechanisms
Articles 21 to 24 build supporting trial mechanisms: promoting mediation and diversified dispute resolution, improving case retrieval systems, issuing typical cases to unify adjudicative standards, enhancing coordination with regulatory authorities through judicial recommendations, and improving the talent pool for technical investigators. Given that AI technology evolves far faster than legislative amendments, the Supreme People's Court's subsequent release of typical cases will be the most critical pathway for refining the rules.
While the Opinions have established a framework for adjudicating AI-related disputes, limitations inherent in judicial interpretations and rapid technological change mean that several areas retain interpretive ambiguity requiring further refinement through judicial practice and typical cases.
VI. Outstanding Issues and Institutional Gaps
First: The stratification standard for service providers' duty of care in AI hallucination scenarios needs further refinement. While Article 7 prohibits blanket immunity based on hallucinations, the obligations for platforms should differ across application scenarios. The level of risk prevention and control required of a general-purpose conversational large model should differ from that for a high-risk professional scenario (e.g., law, medicine, finance). Service providers in high-risk scenarios should bear higher warning and verification duties. This stratification is not explicitly addressed in the Opinions and awaits development through case-by-case adjudication.
Second: The ownership of copyright in AI-generated content is not positively addressed in the Opinions. The document focuses primarily on liability for infringing content generated by AI but does not reiterate the standard for ownership of AI-generated outputs. Based on existing effective judgments, the core criterion remains whether a natural person has invested original intellectual effort. Simple prompt-based one-click generation does not create a copyrighted work; works that embody a user's personalized choices and arrangements through deep prompting, iterative refinement, and parameter adjustments do qualify as works. Although not written into the articles, this remains the applicable rule for adjudicating relevant cases.
Third: The standard of proof for applying personality rights injunctions requires practical refinement. Article 8 allows rights holders to apply for an injunction against personality rights infringement if they can prove "irreparable harm will result if not promptly stopped." While AI deepfake information spreads rapidly, the practical challenge for courts is determining the evidentiary threshold for such injunctions. Preventing abuse of injunctions must be balanced against protecting rights holders and upholding freedom of technological expression.
Ⅶ. Practical Implications for Legal Practice and Compliance
AI technology continues to evolve, and legal systems are inherently lagging. The Opinions of the Supreme People's Court on Legally Adjudicating Disputes Involving Artificial Intelligence does not seek to create a completely new set of AI laws outside the existing civil law system. Instead, it adheres to the underlying logic of instrumentalism, projecting existing legal principles from the Civil Code, Copyright Law, and Personal Information Protection Law onto AI scenarios, distilling previously scattered effective case precedents into unified national adjudication guidelines.
(I) Core Rules and Practical Compliance
The issuance of the Opinions unequivocally establishes the judicial authorities' core adjudicative orientation: tolerating technological trial and error but not indulging it; supporting innovation without granting immunity. Compared to previous emphasis on administrative filing and algorithm registration for formal compliance, these judicial rules substantively address civil tort liability, fault standards, burden of proof allocation, and liability boundaries.
Article 5 clarifies the rules for using public personal information in large model training, dismantling the industry misconception that "publicly available data can be freely used for commercial training." The rules are: the information must be legally public; the individual has not expressly objected; and there is no significant impact on the individual's rights. Enterprises must establish a full-process data compliance review mechanism. This includes:
• Classified Data Management: Categorize training data into public general information, public sensitive information, copyrighted works, and private information. Strictly prohibit scraping private information, unauthorized portrait/voice data, or classified information for model training.
• Usage Restrictions: For ordinary public data (e.g., resumes, public social media posts), use only for general model technical iteration, not for user profiling, targeted tracking, or commercial discrimination.
• Preventing High-Risk Features: Strictly control the development of features enabling "doxing" or aggregating fragmented information to reconstruct private details (phone numbers, addresses, location trails).
• Regular Audits: Conduct regular compliance audits for data sets, retain records of data sources, desensitization logs, and audit reports to meet judicial scrutiny and evidentiary requirements.
Article 7 denies "blanket immunity for AI hallucinations" and establishes fault-based liability. Courts will assess whether the service provider fulfilled obligations for algorithm control, risk review, warnings, and technical optimization.
• Differentiated Content Risk Control: Establish a system distinguishing general chat scenarios from high-risk professional scenarios, applying graded compliance obligations.
• Real-Time Interception: Implement mechanisms to intercept illegal content (deepfake pornography, defamation, fraud) in real-time, fulfilling platform content review responsibilities.
• User Prompt Management: Improve keyword blocking, account bans for malicious prompting, and traceability of infringing content.
• Record Keeping: In potential hallucination tort litigation, enterprises can rely on complete records of risk control logs, technical iteration history, and risk warnings to demonstrate they exercised reasonable care, thereby mitigating liability.
Article 4 specifically regulates deepfakes, voice-cloning, "AI resurrection" of the deceased, and commercial use of virtual personas. This is a core judicial basis for combating AI-related illicit activities and protecting personality rights.
• Strict Access and Use Standards: Establish stringent protocols for AI products involving virtual avatar generation, voice synthesis, or face swapping. Prohibit unauthorized commercial use.
• Fast-Track Infringement Response: Establish a rapid channel to handle personality rights infringements by platform users. Immediately halt generation, remove content, block dissemination, and ban accounts upon discovery of misuse.
• Proactive Risk Mitigation: In litigation, proactively demonstrate that pre-incident prevention and post-incident mitigation duties were fulfilled to reduce the determination of fault and damages.
Articles 12, 13, and 16 address IP risks across the AI lifecycle: input training, output generation, and open-source usage.
• Training Data Copyright: Abandon the misconception that commercial training automatically constitutes fair use. For copyrighted works, prioritize obtaining commercial authorization. Where authorization is unavailable, strictly apply fair use factors (purpose, nature, amount used, market impact) and retain assessment reports as a defense.
• Output Generation: Ensure that prompts do not instruct AI to imitate pre-existing works or well-known IP styles for commercial purposes. Clarify ownership of AI-generated content in user agreements (e.g., ownership based on user's original intellectual input, while enterprise retains model rights).
• Open-Source Compliance: Adhere strictly to open-source license obligations (e.g., GPL compliance) and retain distribution records.
• Defense Documentation: In IP infringement actions, submit training logs, prompt histories, and model operation data to demonstrate lack of subjective fault or substantial copying, potentially reducing damages.
Article 6 subjects algorithmic price discrimination and AI-impersonated celebrity endorsements to tort liability, supporting claims for punitive damages.
• Algorithm Review: Review algorithms to eliminate discriminatory treatment, false advertising, and misleading marketing practices.
• Swift Remedies: Upon consumer complaints, promptly rectify issues and compensate to avoid litigation and potential punitive damages.
Articles 10 and 11 delineate the liability boundaries between physical AI products and online services.
• Physical Product Safety: Manufacturers of robots, autonomous vehicles, smart hardware must implement strict quality control, regularly review algorithms for defects, vulnerabilities, and decision-making errors, issue timely updates and safety bulletins.
(II) Compliance Recommendations:
1. Commercial Contract Compliance:
All AI technology transaction contracts must include quantitative deliverables (e.g., model accuracy, response speed, error rates, functional scope). Avoid vague acceptance criteria like "industry standards." Include specific clauses on "limitations of AI technology," explicitly stating that hallucinations and probabilistic errors are inherent technical features and do not constitute a breach, defining a reasonable margin of error.
2. Evidence Retention and Litigation Compliance:
• Full-Featured Logging: Retain comprehensive records of model training logs, dataset source records, algorithm iteration logs, user prompts, content review logs, risk handling records, and complaint management. Retention periods should align with statutory limitation periods. Consider encrypted cloud backups.
• Evidence Integrity: Do not tamper with or delete AI operation data. Do not use AI to forge contracts, chat logs, or appraisal materials for litigation, thereby avoiding sanctions for obstructing justice.
3. Establishing a Long-Term Compliance System:
Overall, the Opinions construct an AI governance framework prioritizing safety, accommodating trial and error, clarifying rights and responsibilities, and ensuring proportionality of punishment. The core logic of enterprise compliance is no longer solely avoiding administrative penalties but achieving substantive compliance to preserve evidence, define fault boundaries, and reduce liability risks. Within this framework, enterprises can leverage the space for innovative trial and error granted by the judiciary while operating within legal boundaries, balancing technological innovation and compliant operation.
This article was first published in “Data Law Bulletin 43” by Tahota (Beijing) Law Firm.
- Publication date
- 1 October 2026
- Author
- European Innovation Council and SMEs Executive Agency