When is AI a product? China’s new approach and the choices facing the UK Law Commission

In September 2026, China’s Supreme People’s Court (SPC) issued new judicial guidance on disputes involving artificial intelligence. Among the issues addressed is a fundamental question for product liability: when should AI itself be treated as a product? China has drawn a relatively clear boundary, keeping product liability connected to AI embodied in physical products rather than standalone AI services. The development is particularly timely for the UK, where the Law Commission is reviewing the UK product liability regime under the Consumer Protection Act 1987 (CPA), including its treatment of digital products and emerging technologies. The EU has already taken a different route by expressly bringing software within its revised product liability regime. These contrasting approaches provide a useful lens through which to consider the choices now facing the UK Law Commission.

Where does product liability begin?

Section 1(2) of the CPA defines a product as ‘any goods or electricity’. AI contained in a physical product presents less difficulty than non-physical products. If an AI-enabled refrigerator is defective and catches fire, the refrigerator is a physical product and product liability can operate in the ordinary way. The harder question is standalone AI supplied as software or through the cloud.

The EU has answered broadly. Article 4(1) of Directive (EU) 2024/2853 expressly includes software within the definition of a product. Recital 13 further makes clear that software can fall within the Directive regardless of its mode of supply or access, including where it is accessed through cloud technologies. Standalone software can therefore fall within the EU’s no-fault product liability regime.

China has drawn the boundary differently through the SPC opinion. The authoritative text is in Chinese and, at the time of writing, the SPC does not appear to have published a full official English translation. The discussion here therefore draws on an English-language analysis published by the European Commission’s IP Helpdesk and refers, where useful, to the unofficial English translation published by China Law Translate.

China’s approach keeps product liability tied to AI with a physical carrier, such as robots or autonomous vehicles. If such a product is defective and the defect causes damage, product liability can apply on a no-fault basis, meaning that the claimant does not have to establish fault on the part of the producer. Standalone generative AI services, by contrast, are outside this product-liability category. This does not mean there can be no liability: the general starting point is fault-based liability, under which the claimant must ordinarily establish fault, damage and causation. The contrast therefore raises a basic policy question: should standalone AI be brought within the special no-fault product-liability regime, or left to other forms of civil liability?

What does ‘defect’ mean when AI can change?

Section 3(1) of the CPA provides that a product is defective where its safety is ‘not such as persons generally are entitled to expect’. But a particularly difficult question arises where an AI product is safe when supplied but subsequently changes through autonomous learning or updates. This sits uneasily with section 4(1)(d), which provides a defence where the defect did not exist at the relevant time. In this regard, Devenney and Howells observe that the traditional focus on the product at the time of supply becomes difficult to apply to evolving AI. 

The EU responds directly to this point. Article 7(2)(c) of the Directive requires defectiveness to be assessed with regard to a product’s ability to continue to learn or acquire new features after it is placed on the market or put into service. Because the Directive covers software, this extends to standalone AI.

China addresses the problem of post-supply development within its narrower category of physical AI products. According to the European Commission IP Helpdesk analysis, determining whether such an AI product is defective requires consideration of characteristics including its autonomous learning capability, subsequent updates, the user’s degree of control, applicable safety standards, and the adequacy of warnings. Overall, China limits product liability to AI in physical products, while adjusting how defect is assessed to reflect that AI can change over time. Claims about standalone AI services remain fault-based.

Control and proof

AI also challenges the idea that the manufacturer's role ends at supply. Developers may continue to update or control a system. Under the EU Directive, Article 7(2)(e) allows continued manufacturer control to be considered when assessing defectiveness, while Article 11(2) restricts the post-supply defect defence where the defect is connected with software, a related service, an update or a necessary safety update within the manufacturer's control.

Proof creates another difficulty. Relevant technical evidence may be held by the developer or manufacturer, while the operation of a machine-learning system may itself be difficult to explain. Articles 9 and 10 of the EU Directive respond through disclosure rules and rebuttable presumptions, including where technical or scientific complexity makes proof excessively difficult. China also addresses evidential imbalance. The IP Helpdesk analysis explains that, in autonomous and assisted-driving disputes, a party controlling relevant driving data may be required to provide complete records.

Where a party controls relevant documentary or electronic evidence but refuses to produce it without justification, the court may draw an adverse inference against that party.

The choices facing the Law Commission

The comparison shows that adapting product liability to AI involves more than simply adding ‘software’ to the CPA. The Law Commission will have to confront several connected questions. Should standalone AI fall within the no-fault product liability regime? How should defect be assessed where an AI product is safe when supplied but later develops harmful behaviour through autonomous learning? When should a developer remain responsible for changes resulting from updates or continuing control? And how can claimants prove defect and causation where the operation of an AI system is technically complex or opaque?

The EU and China provide different answers to some of these questions. The EU has brought software within product liability and adapted its rules on defect, continuing control and proof. China has retained a stronger connection between product liability and physically embodied AI, while adapting the assessment of defect and evidence to AI’s particular characteristics. The challenge for the Law Commission is therefore not simply to decide whether the UK should follow either model, but to determine where no-fault product liability should begin and end, and how a regime designed for conventional products can remain workable when the behaviour of an AI product may continue to change after it has been supplied.

Mohammad Samiul Huq is a non-practicing Barrister and Advocate at the Supreme Court of Bangladesh and a PhD Researcher at City St George’s University of London.