The Hangzhou Internet Court has recently concluded a network tort dispute arising from a generative AI “hallucination”, described by the media as the first case of its kind in China. The facts were roughly as follows: a user queried a general-purpose generative AI application about a certain matter, and the information given by the AI differed from official information; after the user pointed this out, the AI made further commitments and suggested the user protect their rights through litigation. The user relied on this and sued the service provider.
Three key findings
First, AI has no legal personality. The court made clear that AI itself cannot be a civil subject, and that the legal consequences of its generated content are borne by qualified subjects such as the service provider. This finding clarifies the legal effect of statements such as “the AI promised compensation” — AI output does not constitute an independently attributable declaration of intent.
Second, fault-based liability applies, not strict liability. The court held that generative AI is a “service” rather than a “product”. Product liability applies strict liability, whereas generative AI services are dynamic: different users, at different times, with different questions receive different outputs, lacking the fixed use and quality-inspection standards of a product. Ordinary tort law’s fault-based principle therefore applies.
Third, causation is not established. Applying the adequate-causation standard, the court found that the inaccurate information generated by the AI did not substantially intervene in or influence the plaintiff’s decision-making, so no causation existed. Weighing fault and causation together, the court found that the defendant’s conduct was not at fault and caused no harm to the plaintiff’s rights, and dismissed the claim. The judgment is now effective.
Why the case matters
The value of this case lies not in “AI being exempt from liability” but in the analytical framework it provides. The court did not simply exempt liability on the ground that “the AI generated it”; instead it broke the question into three layers — legal personality, the attribution principle and causation — and reasoned through each. This framework is likely to be followed in later cases.
For companies, the caution is this: the conclusion rests on the factual basis that the plaintiff’s decision-making was not substantially affected. If AI output is used directly by a user for decision-making, and the service provider has clear gaps in content review, risk warnings or error-correction mechanisms, the conclusion could differ.
Key points for service providers
First, improve content boundaries and disclaimers. For questions in professional fields such as examinations, applications, medical care and law, set clear capability-boundary warnings and avoid making definitive commitments. See our note on AI application ethics and security guidelines.
Second, build error-correction and feedback mechanisms. The path for handling a user’s identification of an error should be clear, avoiding situations where “the AI makes a counter-commitment”. See reporting channels for AI misuse.
Third, distinguish “tool” from “opinion”. If the product is positioned as an assistive tool, the interaction design should avoid presenting output as professional opinion. See our note on compliance points for anthropomorphic interaction services.
Fourth, retain output and interaction records. In a dispute, complete interaction records are the key to showing that due care was exercised. See our note on the Supreme People’s Court work report on AI.
The other side of the message: users
The case also signals to content users that AI output is not a professional conclusion that can be relied upon. Courts have already made clear that users of generative AI bear a necessary duty of verification; see our note on the verification duty of generative AI users.
The significance of this judgment is that it turns “AI hallucination” from a technical problem into an analysable legal problem. Liability does not shift automatically because “the AI said it”, nor is it automatically exempt because “AI generated it” — it depends on the specific assessment of conduct, fault and causation.
This article is based on public reporting of the relevant case before the Hangzhou Internet Court and its key findings. It is for general reference only and does not constitute legal advice.