China AI Liability: Who Pays When an Algorithm Causes Harm

Introduction

China regulates artificial intelligence in layers — content rules, algorithm rules, data rules — but it has not yet enacted a dedicated AI liability law (a draft has long been under discussion, and none is in force). The practical consequence is that a person harmed by an algorithm does not sue under "AI law." They sue under the Civil Code, under the Personal Information Protection Law (PIPL), or under the network and product rules — and the choice of route decides who bears the burden of proof.

Route One: General Fault Liability

The default is Article 1165 of the Civil Code: whoever through fault infringes another's rights bears tort liability. Under this route the claimant proves fault — a heavy lift against an algorithm whose training data, model and decision logs sit inside the defendant. In practice fault liability is where a claim starts and usually struggles, which is why the more specific routes below matter.

Route Two: Product Liability

If the AI is embedded in a product — a vehicle, a medical device, an appliance — the Articles 1202 to 1207 of the Civil Code on product liability apply: the producer is strictly liable for a defect that causes harm, the seller has its own obligations, and punitive damages are possible where the producer or seller knowingly fails to act after a defect causes death or serious injury. Product liability flips the practical burden: the claimant shows the product, the defect and the harm; the producer argues exemptions. Where a "defect" sits inside an algorithm — as opposed to hardware — is exactly the point the courts are still working out.

Route Three: Network Infringement and the Notice-and-Takedown Analogy

For harms spread through online platforms — defamation, impersonation, infringement of portrait or reputation generated or amplified by AI — the Articles 1194 to 1197 of the Civil Code govern network users and network service providers, including the familiar notice-and-takedown mechanics and liability for failure to act after notice. A platform that ignores a valid notice can be liable for the expanded damage; a platform that also knows and fails to take necessary measures shares liability.

Route Four: PIPL's Reversed Burden

Where the harm involves personal information, Article 69 of the PIPL shifts the burden: a personal-information processor bears liability for damage caused by its processing unless it can prove it was not at fault. This presumption of fault is the single most claimant-friendly rule in the Chinese AI stack, and it is frequently the route that actually succeeds — see China Cybersecurity and PIPL: Enforcement and Defense.

How It Fits Together in Practice

• A self-driving car harms a pedestrian: product liability against the producer, with the defect question centred on the algorithm.
• An AI chatbot defames a person: network infringement against the user and, where notice was ignored, the platform.
• An algorithm misuses personal data and causes loss: PIPL Article 69, with the processor having to prove it was not at fault.
• General algorithmic harm with none of the above: back to general fault liability under Article 1165 — the hardest path.

How Tianni Law Firm Can Help

Tianni Law Firm (江苏天倪律师事务所), based in Nanjing, Jiangsu, advises on the Chinese side of AI-related disputes, from selecting the right liability route and preserving evidence of what an algorithm actually did, to defending or bringing claims under the Civil Code, PIPL and the product and network rules — alongside the regulatory exposure described in China AI Regulation: Compliance for Foreign Companies. This article is general information, not legal advice. The Civil Code, PIPL and the generative-AI administrative measures continue to develop, and a dedicated AI law may change the map — so the applicable rules should be confirmed against the current official texts before acting.