Podcasts
AI & Data

The Fault-Based Turn: How China's Supreme Court Assigns AI Liability

Syed Shahzad RazaSeptember 14, 20266 min read
The Fault-Based Turn: How China's Supreme Court Assigns AI Liability

On September 7, 2026, the Supreme People's Court of China released a set of opinions addressing AI-related disputes, establishing a framework where plaintiffs generally bear the burden of proving fault. Under this standard, courts evaluate liability based on the specific application scenario and the extent of the developer's preventive measures. The approach places the evidentiary burden on the party alleging harm, who must show that the developer fell short in its preventive measures.

The Evidentiary Burden in China's AI Disputes

The September 7, 2026, Supreme People's Court opinions change what generative AI providers in China must be ready to show. By requiring plaintiffs to prove fault, the court introduces a negligence-based inquiry into the heart of intellectual property and data disputes. This means that simply demonstrating that an AI model produced a specific output that resembles a protected work may no longer be sufficient to establish liability. Instead, the litigation must pivot to the developer's internal processes, specifically their training data sources and their preventive measures.

For developers, this creates a "compliance by design" imperative. To defend against infringement claims, providers must be prepared to produce evidence demonstrating that they took reasonable steps to prevent unauthorized data usage or infringing outputs. This includes proving the legitimacy of training-data sources and showing the implementation of technical guardrails. The legal question is no longer just "what did the AI produce?" but "what did the developer do to ensure the AI acted within legal bounds?"

This shift has profound implications for the cost of litigation. If the burden of proving fault lies with the plaintiff, the complexity of discovery increases significantly. Plaintiffs will likely seek to access a developer's internal audits, training logs, and data-filtering protocols to substantiate their claims of negligence. For developers, the cost of litigation may shift from defending against the "result" to defending their "process."

A fault-based approach turns the defense from a technical argument about output into an argument about developer conduct.

Key Insight

The Technical Boundary: Output vs. Alteration

While Chinese courts focus on the conduct of the developer, US courts appear focused on the technical mechanics of the machine's output. In the case of Doe v. GitHub, filed on September 16, 2026, the Ninth Circuit Court of Appeals addressed the limits of the Digital Millennium Copyright Act (DMCA). The court's decision in this case highlights a narrow interpretation of what constitutes a violation of Section 1202(b) in the context of generative AI.

The plaintiffs in the GitHub matter argued that AI tools like Copilot or Codex violated the DMCA by removing or altering Copyright Management Information (CMI). However, Judge Miller, writing for the panel, concluded that the plaintiffs failed to state a valid claim under 1202(b). The core of the ruling rests on a technical distinction: the "output" generated by the AI does not necessarily constitute a "removal or alteration" of CMI from a copy of an existing work. Instead, the court viewed the AI as creating entirely new works. This distinction is vital. If the AI's output is legally classified as a "new work" rather than a "derivative copy," the legal path to claiming DMCA violations via CMI removal becomes significantly more difficult.

This creates a potential divergence in global legal strategies. In the US, a plaintiff's challenge may be stymied by the technical nature of how generative models function, regardless of whether the training data was used. In China, the challenge moves past the nature of the output to the behavior of the corporation. This suggests that even if a model creates "new" works, the developer's failure to prevent certain types of data usage could still trigger liability under a fault-based regime.

Feature US Approach (Doe v. GitHub) China Approach (SPC Opinions)
Primary Legal Focus Technical mechanics of output (DMCA 1202(b)) Conduct and preventive measures of developer
Key Legal Question Does the output remove or alter CMI? Did the developer take adequate preventive measures?
Burden of Proof On plaintiff (to prove removal/alteration) On plaintiff (to prove developer's fault)
Developer Requirement Low, if output is categorized as "new work" High, must provide evidence of training data and safeguards

The Convergence of Liability and Safe Harbors

A critical, yet currently unanswered, question is whether the "safe harbor" rules typically applied to service providers will be extended to generative AI developers. The September 7, 2026, opinions from the Supreme People's Court suggest that courts may consider applying these rules by analogy to AI providers. If safe harbors are applied, developers might be shielded from liability for infringing content produced by users, provided they act quickly to remove such content upon notification.

However, the "fault-based" nature of the Chinese approach complicates this potential shield. In traditional safe-harbor frameworks, the focus is on whether the provider had actual knowledge of the infringement and acted to mitigate it. In a fault-based model, the inquiry is more proactive: did the developer's architecture itself facilitate the infringement due to a lack of preventive measures? If the latter is true, the "safe harbor" might only protect developers from the actions of their users, but not from the inherent risks of their training processes. This creates a dual-layered liability structure where a developer could be protected against user-driven copyright infringement but still held liable for "faulty" training methodologies.

Scenarios for Global Compliance

As these legal theories mature, two primary scenarios emerge for international AI developers:

  • Scenario A: The Procedural Defense Model. Developers focus their resources on "auditability." This involves creating deep, immutable logs of all training data and automated testing of model outputs to prove that they have taken all reasonable "preventive measures." In this scenario, legal risk is mitigated through technical documentation and rigorous compliance protocols.
  • Scenario B: The Output-Centric Defense Model. In jurisdictions like the US, where the focus remains on whether the output is a "new work," developers focus on the technical architecture of the transformer and the statistical independence of the output from the training data. The goal here is to prove that the model is "transformative" rather than "derivative."

Implications for the AI Ecosystem

The divergence between the US and Chinese approaches creates a complex terrain for several key stakeholders. The distinction between "output mechanics" and "developer conduct" changes how risk is assessed and managed.

For executives and product owners, the Chinese model necessitates a shift in internal governance. It is no longer enough to have a "fair use" or "transformative use" legal argument for the model's output. The legal department must work closely with data engineers to ensure that the "preventative measures" are not just present, but are provable and auditable. The ability to survive a "fault" inquiry will depend on the quality of the technical documentation surrounding data curation and filtering.

For investors, the risk profile of AI startups is moving from the "output" to the "pipeline." An investor might find that a startup with a highly impressive, creative AI model still carries massive liability if that model was trained without a transparent or "fault-free" process. The valuation of AI companies will increasingly depend on the robustness of their data supply chain and their ability to provide "evidence of non-infringement" in the form of preventive technical measures.

Watch List for Policymakers and Regulators

As the legal terrain settles, several indicators will determine which model gains dominance in the global market. Policymakers and regulators should monitor the following developments:

  • The Application of Safe Harbors: Watch for whether the Supreme People's Court or subsequent regulations explicitly extend safe-harbor protections to generative AI providers. If they do, the threshold for "fault" may effectively become the standard for liability.
  • The Definition of "Preventive Measures": Monitor how courts define what constitutes an "adequate" preventive measure. Will it be limited to technical filters, or will it include legal indemnification for users?
  • Cross-Border Precedents: Watch for whether US courts move away from the technical distinction of "new works" and toward a standard of "duty of care" for model developers. If the US adopts a fault-based model similar to China's, it would signify a global shift toward a negligence-based regime for AI.

The intersection of technical output and developer conduct is where the next generation of AI litigation will be fought. Whether the focus remains on what the machine produces or how the company built the machine, the question of liability is moving from what a model produced toward how its developer behaved. Developers must now prepare to defend their processes as rigorously as they defend their products.

Sources: cdn.ca9.uscourts.gov, fangdalaw.com.

Written bySyed Shahzad Raza

Shahzad is an AI & Data leader, author, and strategist with 17+ years of experience helping organizations adopt AI responsibly and at scale. He advises C-suite executives on AI strategy, governance, and organizational readiness, and helps clients transform AI investments into measurable business value through responsible adoption and enterprise enablement.