Editor’s Note — DCC.
Zhu Xiaofeng’s piece in 《政法论坛》(Tribune of Political Science and Law, a top-tier Chinese law journal) takes on the question every GenAI-deploying multinational will eventually face: when a personal- information harm arises from a GenAI service, but the specific causal link to any one actor (model designer, model provider, model user, data provider) cannot be established, who pays? Zhu’s answer is doctrinally bold and operationally consequential: apply Civil Code Article 1254 — the rule originally designed for harm from objects falling from buildings of indeterminable origin — by analogy. The framework is currently a doctrinal proposal in a top journal; it is the kind of proposal that, in Chinese legal practice, often crystallizes into judicial-interpretation doctrine and then into rulemaking within 18-36 months. Multinationals deploying GenAI in China should design liability allocations and indemnities against this contemplated framework now.
The structural problem
Generative AI personal-information torts involve a typical actor cast:
- Model designer — built the foundational model
- Model provider — operates the GenAI service
- Model user — the entity (often a downstream enterprise) deploying the model
- Data provider — supplied the training data or in-context data
PIPL Article 69 paragraph 1 establishes presumed fault for PI processors — so once a specific causation link is identified, the actor must prove non-fault to escape liability. Multi-actor liability allocation considers comparative fault and contribution.
The problem: in GenAI cases, the specific causation link is structurally hard to establish. Three reasons:
- Algorithmic black-box — the model’s internal decision pathway is opaque even to the operator.
- Training-data cleansing — data has been transformed, aggregated, and stripped of provenance.
- Interactive learning — the model’s behavior changes through ongoing user interaction, fragmenting causation across actors and time.
Result: the victim cannot prove which specific actor’s conduct caused the PI harm. Under the standard “burden on plaintiff to prove causation” rule, the victim recovers nothing. The PIPL’s stated purpose — comprehensive PI protection — is structurally defeated.
Why the existing frameworks don’t fix this
Zhu walks through the existing doctrinal candidates and explains why each fails.
Civil Code Article 1165 (general tort) — fails
The construction-element approach: plaintiff must prove (a) injury, (b) tortious conduct, (c) causation. The plaintiff cannot prove causation in the GenAI black-box context, so the claim fails. Article 1165 cannot accommodate the unclear-causation scenario.
Civil Code Article 998 (interest balancing) — fails
For non-material personality-rights infringement, Article 998 allows the judge interest-balancing discretion in evaluating culpability. But Article 998’s discretion operates within the construction-element framework — it doesn’t relax the causation requirement. The plaintiff still has to provide prima facie evidence of causation between the conduct and the harm. The Article 998 framework cannot substitute for the missing causation evidence.
Burden-shifting alone — fails
Some Chinese scholarship has proposed shifting the burden of proving causation to the defendants. Zhu accepts this is a necessary half-step but argues it’s insufficient: even with burden-shifting, where multiple potential defendants each independently demonstrate non-causation for their specific conduct, the victim still recovers nothing. The structural problem is not just who bears the burden; it is that no specific defendant can be identified as the cause.
Inferring liability from “principal direct tortfeasor” — fails
Where downstream third-party actors are involved, the model-designer / provider / data-provider’s contribution is often absorbed into the analysis of the principal direct tortfeasor’s act. The exception is Remsburg v. Docusearch (US 2003), but the US precedent has been narrowly cabined. Chinese courts have not adopted a comparable framing.
Zhu’s framework: three doctrinal justifications
Zhu argues for an analogical extension of Civil Code Article 1254 (unclear causation in building-falling-object cases). Three doctrinal foundations:
1. Communication-safety theory (交往安全)
The principle: whoever creates or maintains a risk to others has an obligation to take all appropriate and reasonable measures to control the risk and prevent its materialization. Applied to GenAI:
- GenAI model designers and providers opened and maintain the risk (the GenAI technology cannot operate without large-scale PI processing, and the model’s interactive-learning behavior generates ongoing PI risk).
- GenAI model users maintain the risk (their use of the technology is what activates the harm potential).
- Data providers contribute to the risk (their data supply enables the PI-processing risk surface).
Each is therefore subject to a communication-safety obligation; failure to discharge it grounds liability even where specific causation is unclear.
2. Gain-and-risk allocation theory
The principle: whoever benefits most from a risk should bear most of the downside. Applied:
- Model designers, providers, users, and data providers benefit substantially from GenAI deployment.
- PI subjects bear the harm cost but receive minimal benefit.
Allocating all downside to the PI subject — on the technicality that specific causation isn’t proven — produces an inversion of the gain/risk-allocation justice principle. Better: distribute the harm cost across the actors that benefit.
3. Causation-proof difficulty + harm-prevention
PI causation in GenAI is structurally difficult to prove, not just fact-specifically difficult. The doctrine should accommodate this. Additionally, requiring potential actors to bear shared liability creates incentives for them to prevent harm in the first place — supporting the harm-prevention function of tort law beyond the compensation function.
The combined justification: communication-safety + gain-risk-allocation + structural-causation-difficulty + harm-prevention provide overlapping doctrinal grounds for the unclear-causation rule.
The proposed rule: Article 1254 by analogy
Civil Code Article 1254 was originally designed for harm from objects dropping or being thrown from buildings: if the specific actor cannot be identified, all potentially-causal users of the building must compensate the victim, with judges determining proportional allocation among them.
Zhu’s proposed application to GenAI:
(1) Trigger condition. Where (a) PI harm has demonstrably occurred, (b) the harm clearly originates from a specific GenAI service or product, and (c) which specific actor’s conduct caused the harm cannot be established despite reasonable investigation.
(2) Liability scope. Each potentially-causal actor (model designer, provider, user, data provider) compensates the full damage to the victim. The victim is not required to litigate among the actors; the actors must absorb the joint exposure.
(3) Inter-actor allocation. Among the actors, the allocation is proportional (按份关系). Judges determine the specific amounts case-by-case, considering:
- Each actor’s role in opening / maintaining the risk
- Each actor’s safeguards (or lack thereof)
- Each actor’s economic benefit from the risky activity
- Comparative-fault factors
(4) Escape mechanism. Actors that affirmatively prove their conduct is not causally connected to the specific harm can be excluded from the liability pool. This preserves the differential-incentive property — actors that invest in safeguards and can demonstrate non-causation are released from joint liability.
How this connects to the existing framework
Zhu carefully positions the framework as complementary to existing PIPL and Civil Code structures, not as a replacement:
- Where specific causation IS established — PIPL Article 69 paragraph 1’s presumed-fault rule continues to apply.
- Where multi-actor liability applies — Article 1170 (joint dangerous conduct) or Article 1171 (joint conduct) continue to apply as appropriate.
- The Article 1254 analogy applies specifically to — the structural unclear-causation case where the conventional frameworks systematically under-protect the victim.
This is doctrinally tidy: the proposed rule fills a specific gap without disturbing the broader liability architecture.
What this tells overseas compliance teams
-
The unclear-causation gap will get filled. Even if Zhu’s specific Article 1254 analogy proposal isn’t ultimately adopted, some framework will fill the gap — the alternative (continued under-protection of PI victims in GenAI cases) is not politically sustainable as GenAI deployment scales. Multinationals deploying GenAI in China should design liability and indemnity frameworks against the contemplated joint-liability outcome.
-
Joint-liability allocation across the GenAI supply chain becomes the operating norm. If you are a model designer, model provider, GenAI service user, or data provider in any role for a Chinese-market GenAI service, plan for the scenario where you become a defendant in a PI case where no single actor can be specifically blamed. The contemplated framework imposes joint-liability exposure regardless of your specific causal contribution.
-
Demonstrable safeguards and audit-able compliance records are now operationally consequential. The escape mechanism in Zhu’s proposed framework — actors who can prove non-causation are released — creates a differential incentive. Actors that maintain comprehensive PIIA documentation, algorithmic-decision audit logs, training-data provenance documentation, and verifiable safeguard implementations have a defensible exit from the liability pool. Actors that don’t, don’t.
-
Contractual allocation across the GenAI supply chain becomes essential. Designers, providers, users, and data providers should contract — explicitly — on liability allocation, indemnification, and cooperation in the event of a joint-defendant scenario. The contracts should anticipate (a) joint-defendant status, (b) information-sharing obligations to support each party’s non-causation defense, (c) cost-allocation in joint defense and joint settlement.
-
Insurance becomes structurally important. The US insurance-and-industry-fund model Zhu references is one operational answer. Multinationals operating GenAI in China should evaluate (a) cyber/tech-liability policies specifically covering joint-liability exposure from unclear-causation PI scenarios, (b) industry-fund or pool arrangements where available, (c) coverage extensions for the comparable-position actors in your GenAI supply chain.
-
The PIIA + audit infrastructure does double duty. PIIA (PI Impact Assessment) and the PI Audit Measures regime are not just direct compliance obligations — they are the evidentiary infrastructure for proving non-causation in a joint-defendant case. Invest accordingly.
The deeper doctrinal point Zhu’s piece signals: Chinese GenAI tort law is, structurally, going to look very different from the US negligence-and-product-liability approach or the EU AI Act / GDPR approach. The Chinese frame will lean on Civil Code architecture, Continental civil-law doctrine, and joint-liability mechanisms in ways that don’t translate cleanly into Western analogies. Multinationals that build compliance programs against the coming Chinese frame — joint-liability, evidentiary-defense infrastructure, contractual allocation across the supply chain — will operate the regime efficiently. Multinationals that wait for the regime to crystallize will be reverse-engineering liability allocations into already-deployed GenAI services under unfavorable conditions.
— 朱晓峰, 生成式人工智能个人信息侵权因果关系不明时的责任认定 (Liability Determination for Generative AI Personal Information Torts Where Causation Is Unclear), 《政法论坛》(Tribune of Political Science and Law), Issue 6, 2025; reposted via 数字经济与法治 WeChat Official Account, November 25, 2025. Original article (Chinese).
Not legal advice. The above is DCC’s structured summary of Zhu’s analysis, with framing for overseas counsel; the three-pillar doctrinal justification and the Civil Code Article 1254 analogy proposal are Zhu’s.