A default rule that does not punish error
On 7 September 2026 the Supreme People's Court (SPC) released its first judicial opinions on AI disputes, 法发〔2026〕10号, running to 24 provisions in five parts. Its central choice is the liability default. Under Article 3, where the law does not expressly provide for no-fault or presumed-fault liability, tort liability is decided under the fault principle of Civil Code Article 1165(1). Courts are told to weigh the AI application's scenario, its degree of autonomy, its technical and informational transparency, and its potential risks and scope of impact, according to a translation of the text.
The case for the opposite approach deserves a fair hearing. AI systems are opaque, victims usually cannot see inside them, and a harmed person may have no practical way to prove negligence. Strict liability would push developers to internalize risk, and it is how many legal systems treat dangerous products. That argument has force where an AI is embedded in something that can hurt people physically.
The SPC accepts it in exactly that case. Strict liability under the Product Quality Law applies to AI-enabled products with 'unreasonable danger', and Article 9 limits AI 'products' to those with a physical object as their carrier, which excludes AI services. According to Jones Day's analysis, defectiveness turns on factors including the product's purpose, self-learning capability, update status, degree of user control and compliance with standards. Presumed fault is likewise reserved for places where statute already provides it, such as the Personal Information Protection Law.
This is a sensible allocation. A chatbot that gets a fact wrong and a braking system that fails are different risks, and the opinions treat them differently.
The Hangzhou ruling gave the logic
The opinions follow the Hangzhou Internet Court's January 2026 'hallucination' judgment. A user asked a generative AI app about university admissions, received a wrong campus address, and was then told by the system that it would pay ¥100,000 if it were wrong. The court dismissed the claim. According to Gowling WLG's summary, it held that AI-generated statements do not form the provider's own intent, and that the provider owed a tiered duty of care: strict on illegal content, a clear warning that outputs may be inaccurate, and only 'reasonable technical and governance measures' on accuracy.
That is the right line. If every incorrect output created liability, no provider could ship a general-purpose language model, because hallucination is a statistical property of the technology and not a defect that diligence can eliminate. A duty to warn and to take reasonable measures rewards providers who invest in safeguards instead of those who merely avoid litigation. Treating the service as a service, not a product, is what keeps that design coherent.
Innovation signals in the text
The opinions state three principles: putting people first, fostering innovation-driven development, and safeguarding a safety baseline. Several provisions give the second principle practical content.
- Article 2 takes an 'inclusive and prudent' attitude, and says processing lawfully public personal information for training, where individuals have not expressly refused, is generally not an infringement of personal information rights.
- Article 7 offers generative AI providers a notice-and-action safe harbor: they avoid liability if, after a rights holder's notice with prima facie evidence and real identity, they promptly take measures such as ceasing to generate the infringing content.
- According to the National Law Review, the opinions also grant appropriate liability exemptions to open-source developers and providers.
The notice-and-action design borrows from the intermediary-liability tradition that built the open internet, and it is more proportionate than a general duty to pre-screen all output.
The burden that comes with it
The developer-friendly default has a counterweight. Article 12 provides that where an AI developer raises a non-infringement defense, it shall be ordered to provide the sources of its training data, records of the training process, the model's mode of operation and its scientific theoretical basis. Per Geopolitechs, courts may also draw adverse inferences if a party refuses to produce evidence without justification, though the allocation follows information access: rights holders prove external facts, and model controllers explain internal ones.
There is a fair reason for this. Only the developer knows what went into the model, and a plaintiff cannot prove copying without that information. But a rule that is workable for a large lab with disciplined data lineage can be punishing for a small team using open datasets of uncertain provenance. Disclosure ordered by a court, under procedural protections for trade secrets, is far narrower than the broad public transparency mandates debated elsewhere, and it only bites once a defendant raises a defense. The risk is in implementation: the opinions do not say how business secrets will be protected, and lower courts will have to improvise.
What the opinions leave open
The SPC deliberately avoided the two hardest copyright questions: whether AI-generated content is copyrightable, and when training on protected works needs authorization. Geopolitechs reports the court noting significant disagreement during drafting. Leaving these to case law is defensible, since premature rules could freeze a fast-moving field, but it means developers still lack the certainty that matters most for investment.
Assessment
The SPC has made a choice that pro-innovation observers should welcome: a default of fault-based liability, strict liability confined to physical products where statute already supplies it, a safe harbor for takedown, and a disclosure duty triggered only by litigation. The weak points are the unwritten procedural safeguards for disclosed training data and the unresolved copyright core. Other jurisdictions debating AI liability should note the structure: a judicial default that separates AI services from AI products does more to keep liability proportionate than a single rule applied to every harm.