Introduction
When a defective electric heater causes a fire, the law has a recognisable route to compensation. When a defective medical device injures a patient, the same statutory framework may apply. But when a conversational artificial intelligence system supplies dangerously inaccurate medical guidance, or a social media recommendation system repeatedly exposes a vulnerable young person to material associated with serious psychological injury, the legal analysis becomes markedly less secure. The harm may be physical or psychiatric; the technology may be delivered entirely online; and the product’s behaviour may be generated by interacting systems rather than traced to a single defective component. The question is no longer simply whether a manufacturer made a dangerous object. It is whether the legal concept of a product, developed for a predominantly physical economy, can regulate harmful digital functionality without imposing liability for every technological error or distressing communication.
Law Teacher writing support
Writing a law essay?
We choose a writer who is an expert in your area of law and qualified to at least the academic level you select. Most of our writers hold a master’s degree or PhD, have teaching experience, or both.
Send us your question and, if you wish, any marking criteria. We’ll write an essay to your instructions.
On 8 October 2026, the Law Commission for England and Wales published Product Liability, Consultation Paper No 279, proposing substantial changes to Part I of the Consumer Protection Act 1987 (CPA). Provisional proposals would treat software and AI systems as products regardless of their delivery method; bring design components of online platforms, including recommendation algorithms, within the regime; broaden the range of potentially liable businesses; clarify the assessment of defect; and address psychiatric injury, data loss, product modification and unusually severe evidential inequality. The Commission has invited responses until 14 January 2027. These are consultation proposals, not amendments already in force, and final recommendations are anticipated later in 2027.[1][2]
The central argument of this essay is that there is a principled case for extending product liability to harmful software functionality, including particular AI systems and platform-design features. Liability should follow the creation and commercial exploitation of an unsafe system, rather than depend on whether its instructions reached the user through a physical object, an application or a remote server. That conclusion is nevertheless qualified. Classification as a product is only the first step: defect, legally recognised damage and causation must still be proved, and the scope of liability must respect the distinction between system design and expressive content. The most difficult question is whether the proposed evidential reforms will make those requirements realistically enforceable while maintaining an intelligible distinction between product liability and an implicit general guarantee of safe outcomes.
The answer therefore requires more than an assertion that AI is risky. It requires analysis of the structure and policy of the CPA; the meaning of the Supreme Court’s existing defect jurisprudence; the technical and conceptual boundaries between software, information and services; the proposed treatment of psychiatric harm; and the relationship between compensation, online regulation and freedom of expression. It also requires attention to an important asymmetry: a developer may know the model version, training interventions, optimisation objective and internal safety evaluations, while a claimant may know only the output encountered and the injury which followed. A law of strict liability which leaves that asymmetry untouched may provide formal protection without an effective remedy.
The existing statutory architecture: strict liability, not absolute liability
The CPA was enacted to implement the 1985 European Product Liability Directive. Section 2(1) makes specified persons liable where damage is caused wholly or partly by a defect in a product. The primary defendants are the producer, a business representing itself as producer by its name or mark, and an importer falling within the legislation. A supplier may become liable under the conditions in section 2(3), particularly if it does not identify an appropriate upstream business when reasonably asked. Where several persons are liable for the same damage, their liability is joint and several. These rules seek to place product risks upon economically responsible entities while sparing injured people the need to prove the particular act of negligence that produced an unsafe item.[3]
That is not the same as a general warranty against injury. Section 3(1) defines a defect by reference to the level of safety to which people generally are entitled, not the safety an individual claimant happened to anticipate. All circumstances matter under section 3(2), including marketing, presentation, reasonable use, instructions, warnings and the time of supply. The legislation expressly prevents an inference of defect from the mere fact that a subsequently supplied product is safer. The question involves a normative standard of safety; a product can fail while satisfying that standard, or be defective although no individual employee acted carelessly.[4]
The claimant ordinarily bears the burden of demonstrating a defect, compensable damage and a causal connection between them. Once those elements are established, the producer may invoke statutory defences. Section 4 includes, among other matters, absence of relevant supply, a defect arising after the relevant time, compliance with a legally imposed requirement and the development-risks defence. The last concerns whether scientific and technical knowledge at the relevant time enabled the defect to be discovered. Liability is therefore without proof of fault but is neither universal nor an instrument for compensating every unfortunate outcome.[5]
The damage rules also restrict the regime. Section 5 covers death, personal injury and specified property damage, not the product’s own diminished value or the ordinary commercial disappointment that it performs badly. Property must generally be ordinarily intended for private use, occupation or consumption, used mainly in that way by the claimant, and the statutory £275 threshold applies. Section 45 defines personal injury to include impairment of a person’s mental condition. Consequently, it would be wrong to suggest that psychiatric injury is wholly outside the existing Act. The uncertainty concerns its boundaries, causation and the extent to which claimants other than the directly injured person may recover.[6]
These distinctions are particularly important for generative AI. A false answer may cause irritation, wasted work or financial loss but not a form of damage for which the CPA provides compensation. A harmful statement might be addressed through contract, professional negligence, defamation or another cause of action. It does not follow that merely describing an output as a ‘hallucination’ establishes a defective product, any more than the breakdown of an ordinary household appliance conclusively establishes an unsafe one. The reform question is whether the established safety regime should cover additional categories of technology, not whether every error made by a machine should become actionable.
What the authorities already teach about defect
Hastings: the public is entitled to safety, but not freedom from all risk
The most authoritative recent domestic analysis is Hastings v Finsbury Orthopaedics Ltd [2022] UKSC 19. The claimant had received a metal-on-metal hip prosthesis which was later revised. His case alleged that inherent risks of wear and adverse tissue reactions made the implant defective. The Supreme Court unanimously upheld the dismissal of his claim. It affirmed an objective standard of entitled safety, assessed in the circumstances at the time of supply. Information emerging later could be considered insofar as it informed the actual level of risk associated with the product at that earlier date; it could not simply be used to substitute later safety expectations for those then applicable. Proof of defect and causal connection remained with the claimant.[7]
The factual reasoning was as important as the formulation of the test. A revision rate and regulatory alerts appeared troubling, yet expert analysis revealed limitations arising from small samples, different patient populations, surgeon-specific outcomes and confounding variables. The Court distinguished a superficially alarming statistical pattern from reliable proof that this particular prosthesis supplied less safety than the public was entitled to expect. The decision does not excuse dangerous products; it recognises that even sophisticated medicine cannot always deliver a risk-free outcome. It also demonstrates how statistical uncertainty can defeat a claim under supposedly claimant-friendly strict liability.[8]
That lesson transfers uncomfortably but usefully to AI. A claimant may establish that a model sometimes generates unsafe advice, or that a recommendation feature is correlated with adverse mental-health outcomes. Neither fact necessarily proves that the version supplied to this claimant had a legally deficient level of safety, or that the particular harm resulted from that deficiency. The comparison of exposed and unexposed users, like a comparison between hip implants, may be distorted by prior vulnerability, selection, usage patterns and changing behaviour. Equally, the law must not require developers to eliminate every possible erroneous output before their systems can be supplied. The inquiry must consider the product’s intended function, foreseeable use, seriousness of risk and available precautions, without confusing proof of safety with assurances of perfection.
Wilkes, Gee and the danger of smuggling negligence into strict liability
In Wilkes v DePuy International Ltd [2016] EWHC 3096 (QB), Hickinbottom J rejected the proposition that a hip-stem fracture necessarily proved a statutory defect. The product was assessed in its proper context, including its utility, compliance with relevant standards, warnings, the rarity of failure and reasonable design compromises. Importantly, the judge explained that a claimant need not always identify why an unsafe product failed: the unsafe condition, rather than its precise engineering cause, may be capable of proof. The judgment illustrates both the generous conceptual structure of the CPA and the evidential difficulty of proving that the requisite safety standard was breached.[9]
Gee v DePuy International Ltd [2018] EWHC 1208 (QB) similarly required an assessment of what safety persons were entitled to expect of the particular medical product, considering its risks and benefits rather than treating a potentially serious adverse outcome as conclusive. Both decisions figured in the Supreme Court’s explanation of the governing law in Hastings. The Commission observes that such litigation often resembles negligence in substance: arguments focus on design choices, warnings, testing and available alternatives, though dressed in the language of entitlement. Claimants may plead negligence alongside the CPA precisely because the supposedly distinct approaches converge in technically complex disputes.[10]
This presents a conceptual problem for reform. If a claimant must prove that a developer could feasibly have adopted safer training methods, safeguards or evaluation procedures, the claimant may effectively be asked to reconstruct what a reasonably competent developer would have done. That reintroduces evidence of care and avoidability into a cause of action intended to relieve claimants of having to show carelessness. Yet excluding design reasonableness altogether is equally implausible where the alternative would be a demand for impossible safety. There is no perfectly clean division. The task is to prevent a flexible defect inquiry from collapsing into a fault requirement that the statutory regime was created to avoid.
Why the definition of product matters
The current statutory definition in section 1(2) is framed around goods and electricity. The historical paradigm is a manufactured item which can be identified, supplied and examined. Software is not necessarily excluded from a claim merely because it contributes to an accident: where a defective embedded controller makes a physical device unsafe, the device’s producer may already face liability. But the application of the CPA to independently supplied intangible software, a cloud-hosted AI system or a continuously changing recommendation service is uncertain. This is a real legislative boundary, not just an artificial inconvenience in pleading.[11]
The UK Jurisdiction Taskforce’s July 2026 Legal Statement on Liability for AI Harms under the Private Law of England and Wales reinforces the distinction. As an expert statement of present law rather than binding judicial authority, it concludes that contractual obligations and negligence often apply to AI-related loss, but that present statutory strict liability ordinarily attaches to AI integrated into physical products. Its analysis is particularly useful because it precedes the Law Commission’s consultation and does not assume that a new liability statute is already operative. It also acknowledges the fact-sensitive nature of negligence, the limitations of claims against general-purpose model developers, and the different risks created by specialised applications.[12]
The Commission’s proposed solution is primarily functional rather than physical. Consultation Questions 1 and 2 would bring software and AI systems within the CPA however supplied, including downloads, remote delivery and AI systems supplied in an intangible form. It expressly envisages software-as-a-service within the definition of product, although it separately raises questions about which accompanying services should remain outside liability. The premise is that dangerous code performs the same physical or behavioural function regardless of whether it arrived with a purchased object or was executed on a distant server.[13]
Consider two insulin-monitoring systems: one installed on a physical monitor, the other accessed through a phone application connected to an online service. If both fail to deliver an essential warning because of identical software logic, the patient’s entitlement to safety should not rationally depend on the distribution channel. Equally, an algorithm may perform the same autonomous selection and ordering operation whether the underlying model runs locally or on the provider’s infrastructure. Making liability turn on embodiment would invite architectural avoidance and protect a deployment model rather than a safety interest.
Nevertheless, the proposed extension creates hard questions of its own. What exactly is the product when a foundation model, a fine-tuned model, retrieval software, a user interface and external applications combine to produce an answer? Is the relevant product the model weights, the system configured for a specific task, the individual run of a programme, or the deployed environment as a whole? If a claimant alleges that a system was unsafe, identifying the product determines the producer, the date of supply, the applicable defence and the evidence required. A definition which simply lists ‘AI’ without addressing these boundaries may transfer uncertainty from classification to causation and defendant identification.
What is AI, and where does the product end?
Consultation Paper No 279 does not settle a rigid definition of AI. Instead, it asks whether core characteristics should be identified, referring to machine-based inference from inputs, outputs such as predictions or recommendations, varying degrees of human specification and possible continued learning. Importantly, the Commission asks separately whether AI models should constitute products in their own right distinct from AI systems. This is an exploratory issue, not a concluded provisional proposal that every model weight or training dataset is independently a defective product.[14]
At least four legally distinct things can exist within one commercial application. A base model supplies learned capabilities. A developer or deployer constructs an interface and system instructions. The system may retrieve further information or call tools. Finally, a human user supplies prompts, exercises judgment and acts on the results. Harm might arise from a basic limitation of the model, an unsafe downstream adaptation, inaccurate source material, a flawed tool integration or reckless human reliance. The mere fact that AI is present does not settle which actor controlled the harmful feature. Nor should incorporation of another company’s model automatically absolve the business which intentionally deploys it for a sensitive purpose.
The distinction between an AI model and its configured application is consequential. A general-purpose model may be used to generate poetry, translate text, control a physical device or provide health information. Its producer cannot foresee every context with equal specificity. An application marketed as a clinical decision-support tool, however, selects a particular purpose and target user population. The safety to which users are entitled may properly be higher, because erroneous outputs are more likely to be relied upon in a context where error is serious. Product liability can take these differences into account without announcing a universal clinical standard for all conversational software.
There is also a methodological problem. Machine-learning systems may exhibit stochastic outputs, changes following deployment, and non-transparent internal processes. An output is not a conventional interchangeable component which can always be examined and replicated. A court may therefore need to specify an alleged defect by reference to the system’s propensity to cause a defined category of harm, its safeguards, or its behaviour under foreseeable conditions, rather than one unrepeatable exchange. The risk is that a sufficiently broad description becomes circular: the system is said to be defective because it caused harm, and the harm proves the defect. The lesson of Hastings is that a court must still evaluate whether that risk fell below entitled safety, not simply whether it materialised.
The boundary between functional code, information and services
The Commission approaches software as both instructions and the functionality which their execution produces, but expressly invites views on whether that definition is workable. It provisionally supports including digital manufacturing files and other information which controls the automated production of physical goods. By contrast, pure information remains outside the ordinary product concept. A digital file containing a set of factual assertions is not necessarily equivalent to code which instructs a machine to perform a potentially dangerous task. Whether particular machine-readable inputs should be treated as a product is deliberately left open in Consultation Question 10.[15]
This distinction is attractive in principle. A defective set of instructions which causes an automated cutting machine to move unsafely acts functionally like a physical component. A factual article which inaccurately describes a chemical’s safety operates through interpretation and communication. Different values are implicated, including editorial independence, intellectual freedom and the need to prevent product liability from becoming a general regulator of factual truth. In practice, however, systems such as AI assistants can convert information into operational instructions. A generated command may be executed automatically, and an advisory output may be deliberately presented to a user as the system’s principal function. A purely formal division between ‘information’ and ‘function’ could make liability depend on superficial interface design.
The consultation accordingly distinguishes software-as-a-service from an associated service provided through software. It provisionally proposes excluding services connected with a product, even where they are needed for the product’s operation, and asks whether bespoke software supplied as a service requires particular treatment. The proposed exclusion of ‘related services’ differs from elements of the newer EU framework. A professional giving individual advice through AI is not transformed into a product manufacturer simply by choosing a digital medium; equally, deploying a general-purpose tool as a specialised application should not automatically deprive the application of its character as software.[16]
A defensible boundary would focus on whether the complained-of danger arises from reproducible software design, configuration, functionality or the way a digital system is supplied and maintained, rather than from the merits of an individual human professional judgment. This would leave legal, clinical and financial advice principally subject to their established professional frameworks while permitting product claims where an unsafe piece of software independently causes injury. It would not eliminate overlap: the same event might support a negligence claim against a clinician and a statutory claim against a defective monitoring application. The law already tolerates overlapping causes of action when their elements differ.
Social media: can a recommendation algorithm be a dangerous product?
The proposal which most clearly challenges conventional product categories concerns social media and other online platforms. The Commission does not provisionally propose that every post, video, opinion, search result or user-generated message is itself a product whose harmful content exposes the platform to strict liability. It identifies design components and features which organise how the platform works: ranking and recommendation algorithms, infinite scroll, filters, visibility mechanisms, notification systems, age safeguards and parental controls. Consultation Question 6 would treat these components as products; Question 7 explores how the relevant online-platform category should be defined.[17]
The rationale is not merely that platforms have software behind them. A recommendation system makes specific design choices about which material to promote, to whom, at what moment and how frequently. The provider may optimise for engagement, retention, advertising performance or other measurable outputs. Infinite scroll and autoplay may alter the conditions in which users encounter content, while default settings and safeguards influence whether vulnerable users can avoid it. These are engineered functions. They can be examined as design decisions even though their immediate inputs include expression supplied by third parties.
The Commission’s distinction between content and architecture deserves support, but it cannot be assumed to be self-executing. A recommendation mechanism produces its risk precisely by selecting and amplifying content. A claimant alleging that a recommender repeatedly directed a vulnerable child towards self-harm material necessarily refers to the content selected; the provider may argue that the actual source of harm was third-party expression or pre-existing illness rather than a defect in the selection mechanism. A court must be able to use content as evidence of how the function operated without converting the truth or offensiveness of individual posts into the defect itself. Otherwise the design/content distinction would either become an obstacle to every claim or a route to disguised content liability.
The Commission acknowledges the difficulty of examining platform components in isolation. A platform may offer strong age verification but weak parental controls, or effective reporting but exceptionally aggressive recommendation features. Users encounter the integrated system, not a series of separate modules. The consultation therefore proposes a holistic defect assessment even where a particular component is identified as the product. The legal question concerns the safety of that feature in its actual surroundings, including the combined operation of other protections. That is sensible, although it complicates identification of the legally responsible producer where third-party tools, contracted moderation teams and embedded recommendation models are involved.[18]
Care is required when translating public concern about social media into a civil cause of action. Associations between a pattern of platform use and psychological distress are not proof that a particular provider’s algorithm caused a particular psychiatric disorder. Some people who are already unwell may actively seek troubling material, while others may be harmed by exposure that they would not otherwise have encountered. The direction and strength of causal influence may vary between products, ages, modes of use and personal circumstances. A legally coherent response must therefore allow serious claims to be investigated without assuming that every instance of high engagement, prolonged use or subsequent illness establishes an unsafe design.
Freedom of expression and the risk of regulating content by another name
The proposed treatment of algorithms sits alongside Article 10 of the European Convention on Human Rights and the domestic protection of expression. The Commission argues that a claim properly directed at platform design, rather than third-party content, generally avoids the central free-expression difficulty, and that any engagement of Article 10 must be considered proportionately. It further emphasises that liability should address serious psychiatric injury attributable to unsafe functioning rather than mere offence, distress or disagreement with lawful material.[19]
That framing identifies a legitimate objective, but the expression issue is not exhausted by the legal label ‘design’. The ranking of speech affects whether it is encountered. A platform facing a large claim may change recommendation criteria, suppress categories of lawful but controversial material, or withdraw certain features rather than risk costly proof of their safety. The potential for such indirect effects does not defeat the case for civil liability; publishers and broadcasters already encounter legitimate legal constraints. It does mean that a court should ask whether the complained-of design danger can be described in terms of a demonstrable safety failure without making liability depend on the viewpoint of lawful content.
Several distinctions could help. Conduct which deliberately targets a vulnerable person with increasingly harmful recommendations presents a different safety problem from a single encounter with distressing political commentary. Mechanisms which bypass stated user preferences, override safety settings or systematically undermine age protections are distinguishable from ordinary decisions about relevance. Repeated exposure generated by the provider’s optimisation model differs from user-initiated access to controversial material. These are evaluative factors, not presently enacted statutory conditions or proof that the mechanism caused a recognised illness. Their value is in locating the potential defect in functional conduct that courts can scrutinise with evidence.
There is a broader constitutional consideration. Product liability traditionally evaluates dangerous designs under open-textured standards. Where those designs are speech-distribution systems, standards must be sufficiently predictable to avoid disproportionate deterrence of lawful expression. A proportionality-conscious assessment should require specificity about the alleged mechanism, affected users, seriousness of the injury, degree of developer control and availability of less restrictive design safeguards. The same analysis should reject the proposition that merely displaying lawful material which someone finds upsetting is itself a defect. Such a threshold would better preserve the Commission’s stated distinction than a blanket exclusion for algorithms or a blanket rule that exposure to harmful content proves liability.
The existing Online Safety Act is not a substitute for compensation
The Online Safety Act 2023 regulates defined online services through duties which concern risk assessment, illegal content, protection of children and the design and operation of relevant systems. Sections 9 to 12 address aspects of the duties of providers of user-to-user services, including proportionate measures relating to functionalities, algorithms and risks. Enforcement is principally public and regulatory, through Ofcom. The Law Commission emphasises that this framework does not, in itself, create an individual statutory damages action for breach of those duties. It is directed at systemic risk management rather than determining whether a particular person is entitled to compensation for an injury.[20]
That distinction supports the proposed reform. Regulation can alter platform behaviour for future users but does not necessarily compensate a person who has already suffered a medically recognised psychiatric disorder. A damages action may also expose a different type of accountability: the court must consider the disputed causal chain in an individual case. Conversely, product litigation is a poor substitute for prompt regulatory action. It is retrospective, dependent on funding and evidence, and generally requires actual actionable damage. A platform may pose a material risk of future injury without any particular claimant yet being able to prove the necessary elements of a civil claim.
Regulatory compliance should be relevant but not conclusive. An authorised design may still be legally defective if the statutory safety standard is not met; a breach of a regulatory rule may be powerful evidence of danger but cannot automatically establish that the breach caused a particular person’s injury. To make regulatory non-compliance dispositive without examining the harm alleged would confuse public enforcement with private compensation. The stronger design is complementary: public duties set and monitor preventative expectations; carefully delimited product claims provide remedies for independently proven injury.[21]
What would make an AI system defective?
The Commission does not recommend abandoning the overarching standard in section 3(1). It proposes supplementing that test with seven non-exhaustive principles, codifying or clarifying circumstances in which the safety of a product is insufficient. The principles are unusually important because they demonstrate that the proposed reform is much more than a redefinition of ‘product’. In broad terms, they concern: abnormal harm arising on reasonably expected use; inadequate information to permit acceptance of a risk; insufficient instructions or warnings to avoid it; presentation which obscures or downplays its seriousness; reasonably avoidable risk, evaluated with regard to feasibility and cost; an unjustifiable imbalance between risk and benefit, subject to socially accepted inherent risks; and a materially worse risk profile than contemporaneously available products offering comparable or greater benefits. Where a claimant establishes a principle and the court finds it engaged, defect would follow within the statutory framework.[22]
These principles offer alternative pathways, not seven cumulative requirements. A claim may succeed because an ordinary use produces abnormal harm even if the claimant cannot point to an alternative design. Another may succeed because the product’s warning concealed a serious and avoidable danger. They retain a holistic evaluation, rather than turning into a checklist whose every box must be ticked. In relation to AI, however, their application raises different problems according to the type of system under examination.
Abnormal harm, foreseeable use and meaningful warnings
The first principle is intuitively suited to a medical application which unexpectedly fails to warn of a dangerous condition, but less straightforward for a conversational model producing variable text. What counts as abnormal behaviour for a probabilistic system designed to answer many types of question? A developer may point out that occasional false outputs are an ordinary feature of current generative models. Yet the prevalence of error cannot by itself settle its social acceptability. Widespread technical limitations should not be allowed to lower the safety standard indefinitely where a dangerous use is actively marketed, and users are induced to rely on the answer. The product’s presentation and target audience must therefore interact with the risk analysis.
The second, third and fourth principles recognise a distinction between the existence of a risk and a user’s capacity to understand it. A prominent explanation that an AI assistant can be inaccurate may be relevant when someone uses it for informal research. It may be less meaningful when the interface invites an unwell person to rely on definite treatment advice while qualifying the warning in small print. Warnings cannot necessarily transform an unsafe design into a safe product: their adequacy depends on the use to which the product is predictably put, their prominence, specificity and the practical ability to act on them. The Commission’s proposals expressly distinguish providing information about accepting a risk, instructions for avoiding it, and the accuracy with which its seriousness is conveyed.[23]
For children and other vulnerable users, a warning is particularly unlikely to perform its function if they cannot reasonably understand the hazard or cannot realistically avoid the relevant feature. This should not create a categorical rule that every product must satisfy the needs of the most vulnerable conceivable person. It does support treating an expressly targeted audience, foreseeable vulnerability and the design of safeguards as important circumstances in the entitled-safety inquiry. A system specifically marketed to young teenagers presents a different safety question from specialist software licensed to trained clinicians. The resulting difference is not moral favouritism; it follows from different marketing, capacities, foreseeable uses and consequences of failure.
Safer alternatives, avoidability and the risk of hindsight
The fifth and seventh principles examine whether the relevant harm could have been avoided without sacrificing the product’s benefits, and how the product compared with alternatives available at supply. For an AI application, the evidence might concern well-established safeguards against dangerous advice, the inclusion of appropriate human review, a more reliable model for a particular task, or the availability of safer configuration. Those inquiries could make accountability more concrete. They may also impose an onerous comparative task on a claimant who cannot inspect the defendant’s training data, test results or alternative designs.
The Commission tries to prevent hindsight from being determinative by anchoring avoidability and comparative performance to the time of supply. A safety measure invented years later should not retrospectively prove that an earlier model was defective. There is nevertheless a temporal difficulty unique to services subject to continuing revision. A model can be updated silently, a recommender’s optimisation objective changed, or a protection system withdrawn after deployment. A court must identify the relevant version and period rather than comparing an amorphous ‘AI product’ against a static technological benchmark. Without reliable version records, the time-of-supply qualification risks becoming impossible to apply.
The sixth principle, balancing risks against benefits, is potentially the most controversial. In medicine, some inherent dangers may be justified by therapeutic benefit. But what counts as the benefit of a social media recommender? Commercial engagement and advertising revenue are benefits to the operator; entertainment, communication and access to information are benefits to users. These should not be treated as interchangeable. If a feature predictably exposes a minority of young users to serious harm while increasing platform retention, the private commercial advantage does not necessarily justify the public safety risk. Conversely, a judicial assessment of benefits should recognise that ranking and discovery technologies can support useful connections, creativity and the exercise of rights. The Commission’s exception for risks so well known as to be socially accepted should not become a presumption that familiar technology is safe merely because consumers have little realistic alternative.[24]
Can a claimant prove that an algorithm caused the harm?
Product liability remains based on actual causal responsibility. A claimant who proves that an algorithm is defective must separately establish that the defect caused actionable damage, subject to the ordinary rules of proof and any enacted evidential modifications. This is the decisive limit on the suggestion that recognising software as a product would immediately unlock compensation for all AI-related injury. A defect may raise risk across a population without having caused the particular claimant’s condition. Equally, a claim may involve demonstrable individual injury without sufficient evidence that the digital system was the cause.[25]
In a straightforward hypothetical case, a dose-calculation application produces an incorrect number because of an identifiable coding error. The patient follows the instruction, suffers a documented physical injury, and the record establishes what the correct dose should have been. Product, defect, injury and causal sequence may be relatively accessible. Different issues arise if the application gave an erroneous suggestion which a clinician reviewed, altered and then used in a treatment decision. The clinician’s judgment may form part of the causal chain. It might be a concurrent cause, a foreseeable response to misleading system design, or, in exceptional circumstances, a sufficiently independent intervention to affect legal causation. Each conclusion would depend on what occurred; using a human intermediary should neither automatically absolve nor automatically implicate the developer.
Now consider a person who develops a recognised psychiatric illness after months of using a recommendation platform. The claimant may have encountered thousands of posts through a combination of active searches, automated suggestions, subscriptions, recommendations from friends and ordinary external circumstances. Establishing that a particular ranking mechanism was unsafe is distinct from establishing that, without its defective functioning, the illness probably would not have occurred or would have been materially different. Counterfactual reconstruction may require data about the sequence and intensity of recommendations, user interactions, prior symptoms, comparative exposure and other possible causes. Much of this evidence may be unavailable to the user and may have changed since the events.
The law’s ordinary starting point is the ‘but for’ inquiry, though legal causation is more complex where several factors operate together. The UK Jurisdiction Taskforce’s July 2026 legal statement emphasises that the ordinary principles can apply to AI, that misuse by third parties presents fact-sensitive questions about intervening acts, and that autonomous behaviour is not inherently a legal escape route for a developer or deployer. Its conclusions are reasoned predictions about existing private law, not an AI-specific statutory causation code. They offer no basis for replacing proof with the slogan that ‘the algorithm did it’.[26]
The distinction between general and specific causation is useful. General causation asks whether a feature is capable of contributing to a defined injury in some circumstances. Specific causation asks whether it did so in this claimant’s case. Controlled experiments, epidemiological analysis, system audits and expert evidence may assist the first question. The second may require an individual chronology, medical evidence and digital records. It would be a mistake to treat a population-level association as automatic proof of an individual’s claim; it would be equally unreasonable to demand a form of certainty that is routinely unavailable in complex injury litigation. The appropriate standard remains the civil standard, applied to the evidence relevant to the actual causal question.
Foreseeability should also be kept distinct. That a developer could foresee some danger may inform whether its system provided the entitled level of safety, and may be relevant to remoteness in an appropriate tort claim. It does not establish that the danger materialised through the alleged defect in this particular case. Equally, a harmful outcome being rare does not necessarily make it unforeseeable, particularly where internal testing has identified the precise failure mechanism. The analysis should resist collapsing foreseeability, defect and causation into a single intuitive judgment about the seriousness of the injury.
Evidence and the structural advantage of the developer
The Commission recognises that technically complex claims already require expensive expert evidence and that producers usually possess the more relevant documents, knowledge and resources. Hastings illustrated how questions of statistical reliability, comparators and confounding could defeat a claimant even in litigation concerning a physical device. In AI litigation the imbalance may be deeper: the product may be inaccessible in the version that caused harm, and the defendant may be the only actor able to reconstruct its recommendations, safety settings and interaction history. A claimant may need access to evidence to formulate an arguable allegation but be unable to obtain sufficient evidence without first commencing costly proceedings.[27]
Potentially relevant material could include version histories, model or recommender configuration, relevant output logs, safety evaluations, recorded incidents, decisions about deployment and evidence of available safeguards. Not every case requires source code or the disclosure of training data. An intelligible description of relevant system behaviour or sufficiently detailed test results may sometimes be enough. Any disclosure regime must also account for trade secrets, security risks, third-party privacy, data retention obligations and disproportionality. The analytical issue is whether the defendant’s legitimate interests can be protected through calibrated procedures without allowing secrecy to defeat meritorious claims as a matter of course.
Existing civil procedure provides routes to disclosure and expert examination, but those mechanisms do not guarantee that a claimant can identify an opaque product defect at proportionate cost. If essential information was never recorded, ordinary disclosure cannot produce it. If a model was replaced or logs deleted, the claimant may be forced to rely on secondary evidence. This suggests that the most useful reform may not be a general duty to explain every AI output, which could be technically unrealistic, but targeted preservation and disclosure rules for information genuinely material to an alleged safety failure. That possibility warrants careful distinction from the specific Law Commission proposals actually under consultation.
Two distinct proposals for evidential difficulty
Discrete factual presumptions: an exploratory question, not a settled proposal
Consultation Question 33 asks whether there should be rebuttable presumptions concerning particular facts relevant to defect or causation where a claimant establishes information asymmetry and it would be fair to require the defendant to rebut the fact. The Commission deliberately presents this as an exploratory question, not a provisional proposal. It expressly resists a general reversal of the burden of proving defect or causation, reasoning that broad presumptions may unduly alter the balance of risk and create substantial procedural complexity. The distinction matters because descriptions of the consultation as simply ‘reversing the burden of proof for AI companies’ are inaccurate.[28]
A narrowly defined presumption might address a fact such as whether a particular safety setting was operating on the relevant date, whether a given version was supplied, or whether an identified safeguard had been disabled. These are matters on which the provider will ordinarily possess better records than the individual user. Requiring the defendant to rebut them may be justified, particularly where the relevant information was created by and remained under the defendant’s control. Yet a collection of small presumptions could generate complex satellite disputes about exactly which facts shift and what evidence rebuts them. The device is most attractive where the burden attaches to a fact which the defendant could readily have recorded or disclosed, not to an unbounded opinion about the system’s overall safety.
The proposed exceptional-difficulty mechanism and the 50 per cent starting point
Consultation Question 34 contains a materially different, more developed provisional proposal. Where a claimant can show an arguable defect, non-fanciful supporting facts, exceptional evidential difficulty from limited available evidence, and a risk not so widely known as to be socially accepted, the defendant would be required to elect between two routes. It could try to rebut a presumption of defect, accepting the burden of doing so. If it failed, defect would be treated as proved for the purpose of assessing full damages. Alternatively, the defendant could accept the alleged defect for the purpose of the mechanism and have damages allocated on a risk-sharing basis. The starting point would be 50 per cent of the otherwise recoverable award, capable of adjustment according to the claimant’s reasonable understanding and acceptance of risk and their contribution to the harm.[29]
Crucially, neither election relieves the claimant of proving actionable damage and its causal connection with the alleged defect, and ordinary statutory defences remain relevant. The mechanism concerns exceptional uncertainty about defect, not a general right to half the claimed amount whenever an AI product is complicated. A claimant whose psychiatric illness cannot be causally attributed to the alleged unsafe design would still fail. Conversely, if a claimant proves causal injury and qualifies for the mechanism, it would no longer always be possible for a producer to win solely because the factual material needed to decide defect cannot be obtained.[30]
This is an intellectually unusual attempt to distribute epistemic risk. The traditional approach allocates non-persuasion to the claimant; if the court cannot decide whether a defect existed, the defendant ordinarily succeeds. The proposed option recognises that uncertainty may have been created or controlled by the commercial actor and that requiring the injured person to bear it entirely may be unfair. The defendant’s choice is intended to preserve commercial flexibility and avoid forcing every exceptional case into a costly inquiry. In effect, the law would use a presumptive full-liability risk to encourage the defendant either to produce sufficient evidence or to accept a partial allocation of loss.
There are substantial objections. First, the 50 per cent figure is not derived from a finding that the product was 50 per cent likely to be defective; probability of defect and the appropriate share of loss are not interchangeable. Secondly, the option of paying a proportion of damages without a conventional defect finding is difficult to reconcile with the CPA’s original premise that liability is based on an objectively unsafe product. Thirdly, the mechanism might discourage smaller developers from defending novel but safe technologies if the costs of disproving a defect exceed the sum in dispute. Fourthly, requiring the court to investigate whether the claimant’s risk was ‘socially accepted’ could reward industry-wide normalisation of dangerous design. These are not reasons automatically to reject the proposal, but they demand principled justification for the scope and measure of risk sharing.
A particular concern arises for psychiatric-injury cases involving algorithms. Proving that an algorithm caused a diagnosed illness may itself be more difficult than proving that a feature was defective. If causal proof remains untouched, the exceptional-difficulty mechanism may mainly assist other sorts of claim, such as those involving malfunctioning devices or unexplained product risks. The proposal therefore risks being presented publicly as a remedy for opaque algorithmic harm while leaving the decisive evidential obstacle in place. One response would be to retain individual causation requirements while allowing carefully bounded factual presumptions, supported by compulsory preservation of relevant interaction records. That combination may be more useful than expanding an abstract damages-sharing mechanism.
Psychiatric injury: direct claims and the limits of ordinary distress
Psychiatric harm is not a marginal issue in the proposed digital regime. Many alleged injuries associated with conversational AI and recommendation platforms concern anxiety, depression, trauma, eating disorders or other mental conditions, rather than the sort of immediate physical accident which characterised the 1987 Act’s paradigm. Section 45 already includes impairment of mental condition in the statutory definition of personal injury. The Commission’s task is therefore not to invent mental injury as a compensable category, but to clarify what threshold and controls should govern it and whether people indirectly affected may claim. It carefully distinguishes an illness which can be established with medical evidence from passing emotional distress.[31]
Consultation Question 37 provisionally proposes that psychiatric illness consequential on physical injury, or on an event which physically imperilled the claimant even without actual physical injury, should be compensable where medically ‘recognised’ or ‘recognisable’. Question 38 addresses psychiatric injury suffered directly without such physical injury or imperilment. In addition to a medically recognised or recognisable condition, it would require that development of that illness fall within the natural range of reactions of a person of ordinary fortitude, assessed in the circumstances. Neither question proposes compensation simply for annoyance, transient anxiety, offensive content or general unhappiness with a digital service.[32]
The requirement of ordinary fortitude creates a difficult interaction with targeted vulnerable users. A person may be unusually susceptible to a harmful recommendation mechanism precisely because the platform has classified them by interests, behaviour or other indications of vulnerability. It would be uncomfortable if the product could be legally criticised for exploiting a foreseeable susceptibility while the injured person’s susceptibility then excluded recovery as an abnormal reaction. The answer should lie in the Commission’s direction that all relevant circumstances matter, not a mechanical application of a hypothetical average adult. The standard nevertheless has a restrictive purpose: the law must differentiate serious, reasonably attributable psychiatric illness from the potentially unlimited consequences of exposure to distressing material.
It is important not to assert that an algorithm has caused a particular disorder simply because the claimant received a diagnosis after using the platform. An expert may establish that the condition is clinically recognisable while being unable to identify its cause. The courts will need to distinguish diagnostic proof, individual causation and attribution to a defect. They may encounter disputes about pre-existing symptoms, cumulative exposure, social and family circumstances, direct searches and the extent to which risk existed independently of the challenged feature. These are familiar problems in personal injury law, but algorithmic personalisation creates new obstacles in reconstructing each person’s exposure.
Indirect victims and the challenge to secondary-victim orthodoxy
The Commission’s proposals for people affected by injuries to others are particularly ambitious. Under the ordinary law of negligence, recovery for psychiatric injury suffered by a person witnessing a loved one’s injury is restricted by controls developed in cases including Alcock. In Paul v Royal Wolverhampton NHS Trust [2024] UKSC 1, the Supreme Court rejected attempts by relatives to recover for psychiatric illness arising from witnessing deaths or medical crises consequent upon earlier clinical negligence in the circumstances before it. The judgment explains the special and limited role of accident-related secondary-victim liability. Those common-law controls should not be assumed to apply unchanged to the statutory CPA, whose text is silent on claims by indirect victims.[33]
The Commission would expressly permit a sufficiently connected indirect claimant to seek compensation for psychiatric illness caused through the physical or psychiatric injury, physical imperilment or death of a directly affected person. It provisionally requires a close tie of love and affection and a recognised or recognisable psychiatric illness within the ordinary-fortitude control. It envisages a defined list of relationships, including spouse, parent, child, sibling and qualifying cohabitant, with others able to prove an equivalent tie. It does not propose importing wholesale the negligence requirements of sudden shock, a discrete accident or proximity in time and space. That choice is specifically designed to accommodate gradually developing harm, including a parent who becomes psychiatrically ill while caring for a child injured by a defective digital system.[34]
There is a coherent policy distinction here. The CPA channels responsibility to businesses producing or supplying goods with an objectively unacceptable safety risk; ordinary negligence may extend to a much wider population of defendants and failures. A wider category of indirect claims could be justified as part of allocating the cost of dangerous commercial products. Nevertheless, the indirect claimant’s harm remains an additional causal step removed from the product, and potentially several people may develop illness after a single direct injury. The proposed close-relationship and medical thresholds are intended to limit that expansion. Their success will depend on the courts maintaining robust proof of the causal chain rather than presuming that familial distress is psychiatric damage.
A family-based presumption also warrants examination in contemporary social reality. A child may rely most strongly on a grandparent, sibling, long-term carer or non-biological caregiver. A fixed list may provide procedural certainty while inadvertently favouring formal relationships over actual emotional bonds. The Commission recognises this by proposing that people outside the list could establish a sufficiently close relationship and by separately asking whether the list should be expanded. Courts should be alert to both risks: overly intrusive inquiry into genuine relationships and unjustified exclusion of carers whose connection does not fit conventional categories.[35]
Who should be sued: the developer, deployer or marketplace?
Extending the definition of product would be of limited value if the law continued to identify only the manufacturer of a tangible item. AI services may have a foundation model developer, a company refining or adapting the model, a business deploying it in a particular setting, operators providing data or infrastructure, and a platform offering access to end users. The Commission therefore provisionally proposes extending ‘producer’ to the person who developed a product. It also considers how liability should work where defective components are integrated into a product or where a product is substantially modified by another actor. That is a more realistic response than treating ‘the AI company’ as a single legally identifiable defendant.[36]
The starting point should be control over the feature which rendered the product unsafe. A provider which develops a narrowly targeted medical application has substantial control over its claims, interface, validation and deployment context, even if the underlying model comes from another business. A general-purpose model developer may control the base system but not every unforeseeable downstream use. A claimant should not automatically have to choose one entity where several materially contributed to the unsafe product. The Commission’s proposed component rules allow responsibility to be allocated according to whether an integration or modification occurred within the original producer’s control, including an authorised third-party component presented as part of the original product offering.[37]
For example, suppose a connected home-monitoring device is sold as including a particular diagnostic application which the buyer must obtain from a named third-party developer. The Commission regards it as defensible to hold both the application developer and the device producer responsible if the application is defective, because the named application forms part of the advertised offering. The position is different where the device can run any of thousands of independently selected applications. The manufacturer does not necessarily adopt every conceivable application merely by permitting installation. This control-based distinction avoids imposing open-ended liability on businesses for products they could neither select nor meaningfully evaluate.[38]
Substantial modification presents a similar issue. The consultation envisages liability for a person who, outside the original producer’s control, materially changes a product’s performance, purpose or type in a way which alters safety expectations or risks, and then supplies the modified product in business. A developer may fine-tune a general model for medical advice, remove risk controls, alter its purpose and offer it commercially. Treating that modified system as a product for which the modifier bears producer responsibility is defensible; otherwise responsibility might be misplaced upon the upstream developer which neither authorised nor controlled the new configuration. Yet a threshold based on ‘substantial modification’ must avoid characterising every ordinary security patch or interface change as a wholly new product.[39]
Online marketplaces are not automatically liable for every defective listing
The consultation also addresses the familiar difficulty of suing an inaccessible seller operating through a digital marketplace. Current statutory responsibility may attach to producers, own-branders, importers and suppliers, but the structure of modern fulfilment and third-country e-commerce can make a suitable defendant difficult to identify or pursue. The Commission provisionally proposes liability in specified circumstances for fulfilment service providers, UK authorised representatives and a new category of ‘Core Participant’. That final category is directed at entities whose conduct materially facilitates supply to the UK, which receive relevant business benefit, and which have influence or control over manufacture or distribution, in circumstances where specified other UK defendants cannot be identified. A platform’s presence in a transaction alone would not suffice.[40]
Two distinct marketplace questions must not be confused. First, an online marketplace might itself deploy a defective algorithm or interface; if that functionality becomes a statutory product, liability concerns its own unsafe design. Secondly, a marketplace might facilitate the sale of a defective item made by someone else; the proposed supply-chain provisions may extend responsibility if the specified conditions are satisfied. The rationale, relevant product and potential defendant are different. The Commission’s attempt to identify decisive influence and business benefit is a sensible restraint, although the meaning of practical control will invite litigation when the platform controls visibility, payments, contractual terms, fulfilment or vendor access to differing degrees.
The existing statutory requirement that a supplier provide identifying information about an upstream actor still has a role, but identification is not the same as effective recovery. A product may be developed by a company outside the jurisdiction, supplied through a separately incorporated subsidiary and backed by an entity with few assets. The Commission proposes procedural improvements where a claimant has used the wrong company name or sued the wrong entity in a corporate group. It does not, however, undertake wider reform of corporate insolvency, enterprise liability or rights against insurers. These are significant limits to the practical reach of any new substantive cause of action.[41]
Free products, advertising-funded platforms and open-source code
In digital markets, price paid by the immediate user is a poor measure of commercial exploitation. A social media service may be free in money terms while generating advertising revenue and benefiting from audience attention or data-driven monetisation. Consultation Question 15 would clarify that a product is supplied for CPA purposes where it is made available in the course of business or with a view to profit, even without payment or other consideration from the particular user. This avoids making redress depend on a business model which hides commercial value from the transaction with the claimant.[42]
Open-source development raises a different problem. Requiring a volunteer maintainer to insure against potentially vast psychiatric or physical injury claims could discourage collaborative development. Yet excluding all open-source software would enable a commercial developer to avoid responsibility by publicly releasing code or building a profitable application on open-source components. The Commission does not propose a blanket exemption. It considers that the existing exclusion where the relevant supply is not in the course of business or with a view to profit provides a starting protection for genuinely non-commercial work, while commercial use or distribution may fall within the regime. This is more precise than saying ‘all free software will be liable’.[43]
There remains an unresolved attribution issue. A small open-source contributor may publish a general-purpose tool without control over its later integration into a high-risk commercial service. The downstream business may modify it, make safety claims and earn revenue. Responsibility should ordinarily be concentrated on the commercial actor which creates or controls the unsafe deployed configuration, subject to particular facts about defective components. The regime should not allow the word ‘open-source’ to immunise a profitable supplier; neither should it expose unpaid contributors to liability merely because a much larger company has chosen to deploy their work irresponsibly. The proposed control and commercial-supply criteria are promising, but their interaction should be expressed with sufficient clarity to avoid uncertainty deterring legitimate development.
AI does not remain the product originally supplied
Physical-product doctrine usually looks to the condition of a product when it leaves the producer’s control. AI products complicate that assumption. Cloud software may be revised frequently; a model’s protective instructions may be changed remotely; a recommender may continuously adapt to user behaviour; and vulnerabilities may be discovered only after a product has been deployed at scale. A developer may retain the power to correct or worsen safety risks long after an account was opened or an application first downloaded. The relationship between time of supply and legal responsibility is therefore central, not peripheral, to the proposed reform.
The Commission provisionally proposes additional circumstances relevant to defect, including a product’s cybersecurity, its interaction with other products, the needs of the intended users, the continuing control of a producer and relevant updates. But it expressly states that its approach would not impose a free-standing statutory duty to supply software updates. Failure to issue an update may contribute to a finding that a product has become defective where the circumstances justify an expectation of continued protection. A manufacturer’s assurances about the period of support, applicable standards, the seriousness of a vulnerability and the feasibility of a fix may be relevant. The mere fact that no update was supplied is not sufficient to establish liability.[44]
That distinction is important. An indefinite duty to maintain all software would be neither economically realistic nor compatible with the ordinary lifecycle of technology. Conversely, a producer which advertises continuing security support and then leaves a dangerous vulnerability uncorrected while retaining control of the system cannot necessarily rely on the product’s original safety. A defensible rule requires evidence of the particular safety entitlement, the producer’s continuing ability to affect it, and the relationship between the omission and harm. The Commission also proposes that, in relevant cases, assessment of safety expectations should take account of the period during which the producer retained control, rather than treating original supply as permanently conclusive.
This approach should be applied carefully to learning systems. A model that produces different results as a consequence of ordinary probabilistic generation is not necessarily ‘modified’ whenever it generates a new answer. Nor should an automated adjustment of a recommender always be equated with a substantial product modification. The law should distinguish routine operation within the original design from a material change in capability, intended purpose, safety controls or risk. Otherwise the concepts of supply and modification could become unstable in almost every AI claim. The consultation’s proposed substantial-modification threshold offers a starting point, but will require technologically literate application.[45]
The development-risks defence: innovation and unknowable danger
Section 4(1)(e) recognises that a producer may be unable to discover a defect because scientific and technical knowledge was insufficient at the relevant time. The rationale is not that an innovator deserves immunity for taking precautions which turn out to have been careless. It is that allocating every genuinely unknowable risk to producers could inhibit socially beneficial products and make insurance or commercial development excessively difficult. The Commission provisionally proposes retaining the defence across all industries, while clarifying that the state of accessible knowledge should be assessed objectively rather than by asking only what a reasonable producer in the particular sector might have discovered.[46]
AI makes the distinction between an undiscoverable defect and a known class of risk especially contested. A developer might genuinely lack a scientific method capable of predicting a rare harmful interaction. Alternatively, it might know from its own tests that a system has a significant propensity to generate dangerous advice but be unable to predict precisely when that error will occur. Unpredictability of an individual output is not necessarily the same thing as scientific impossibility of discovering an underlying product risk. The defence should not be converted into a reward for designing a system whose internal decisions are opaque or for declining to perform available evaluations. Nor should the existence of general public concern about AI mean that every particular defect was technically discoverable.
The Commission also proposes adapting the relevant time where a producer introduces a defect through an authorised modification or leaves a product defective by failing to supply an update in circumstances engaging the CPA. In those cases, the development-risks inquiry would concern knowledge when the product became, or remained, defective through the relevant post-supply conduct, not simply knowledge at the original sale. Separately, the proposed changes would restrict reliance on the defence in section 4(1)(d) where a producer itself causes a post-supply defect within the identified control circumstances. These proposals prevent the original release date from providing permanent cover for risks which come under a producer’s continuing control.[47]
Limitation periods and injuries which appear years later
The limitation regime exposes a further difficulty. Claims under the CPA are ordinarily subject to a three-year period measured by reference to injury or relevant knowledge and a separate ten-year long-stop beginning at the statutory relevant time, normally supply. The long-stop extinguishes the right to bring a claim even where relevant harm is discovered only after the ten years have passed. For continuously deployed AI systems it may also be difficult to identify the relevant supply or modification date. A strict historical long-stop can therefore leave an injured person without statutory redress even where the dangerous functionality was substantially revised much more recently.[48]
The Commission provisionally proposes restarting the ten-year long-stop when a product is substantially modified, using the date the modified product becomes available or is supplied as specified. It also proposes extending the long-stop to 25 years for latent personal injury where the claimant could not initiate proceedings within the initial ten years, with an accompanying consultation question concerning relevant knowledge. These are proposals, not amendments to limitation law already enacted. They attempt to reconcile a finite period of commercial responsibility with the unfairness of a claim expiring before a claimant can know that the injury has occurred.[49]
Both proposals create tensions in relation to AI. If every material software update resets time, the possible duration of liability may be long and uncertain; if no update does so, the law may insulate important later changes. The proposed substantial-modification threshold will therefore bear considerable weight. A second issue concerns slowly developing psychiatric injury: the date of diagnosis may come much later than initial exposure, but it should not be assumed that a diagnosis years afterwards proves an injury was latent for statutory purposes. Courts would still have to examine when the injury existed, whether it was discoverable and which product version or period of use is causally implicated. Long-stop reform cannot remove those factual questions, but it can prevent an arbitrary historical cutoff from deciding them before the merits are heard.
Property damage, destroyed data and the limits of consumer remedies
The new economy creates losses which are not easily accommodated by the CPA’s traditional distinction between personal injury and physical damage to private property. A defective AI application might corrupt a person’s private files without damaging a physical disk. A software update could delete photographs or records, and a faulty algorithm might render important information unusable. The Commission provisionally proposes recognising destruction and corruption of data as compensable damage, while excluding data used exclusively for professional purposes and reducing awards for data used partly professionally. It expressly distinguishes destruction or corruption from data leaks and breaches of data protection rules. Whether mere loss of access should be treated equivalently is put forward for consultation, rather than settled as a proposal.[50]
These distinctions deserve support but remain difficult in practice. If files are temporarily inaccessible because a service is offline, there may be no destruction or corruption; yet a person’s inability to reach an essential record can cause serious consequences. Conversely, allowing every period of service unavailability to produce a product-liability claim risks converting the CPA into a general guarantee of continuity or contractual performance. A functional but bounded definition should focus on the loss of integrity, usable possession or recoverability of data, while preserving separate contractual and consumer remedies for ordinary service failure.
The Commission also proposes loosening the current requirement that damaged property be used only for private purposes, so that mixed private and professional use can be addressed proportionately, and removing the £275 property-damage threshold. This could assist people using the same phone, computer or storage account for family life and work. It would not make every form of business interruption or lost commercial profit compensable. Nor would it mean that a product’s own repair cost is automatically recoverable under the CPA: the consultation maintains important exclusions relating to damage to the defective product itself and particular integrated components.[51]
Separately, the Consumer Rights Act 2015 already imposes contractual standards for digital content supplied within its scope, including satisfactory quality, fitness for purpose and conformity with description. Its section 46 provides a specific remedy for damage caused to a consumer’s device or other digital content where the relevant conditions are met, including a requirement concerning the trader’s failure to exercise reasonable care and skill. These rights are not identical to CPA product liability: the parties, statutory elements, remedies and policy differ. The proposed reforms would add a potentially important route to recovery for qualifying personal injury and other damage without making those existing digital consumer rights redundant.[52]
The EU comparison: convergence in scope, divergence in proof
Directive (EU) 2024/2853 provides an important comparative framework. It expressly includes software within its product definition regardless of delivery method, recognises various forms of digital product control and addresses disclosure and presumptions. Member States must transpose it by 9 December 2026, and its new regime applies to products put on the market or into service after that date. At the date of this essay that deadline had not yet passed. The Directive is EU law; it does not itself amend the United Kingdom’s Consumer Protection Act 1987 following Brexit. The Law Commission is considering a domestic legislative response rather than announcing that EU rules have already taken effect in England and Wales.[53]
The similarities are substantial. Both approaches recognise the inadequacy of relying solely on tangible goods. Both grapple with software updates, digital supply chains, modified products, the costs of strict liability and injuries which emerge after a considerable delay. The EU rules also contain an express exclusion for free and open-source software developed or supplied outside commercial activity. The Commission’s preferred approach reaches a comparable policy issue through existing commercial-supply conditions rather than recommending a general, separate open-source carve-out.
The difference in evidence is more striking. Article 9 of the Directive introduces a mechanism for disclosure of relevant evidence, subject to conditions and safeguards. Article 10 sets out situations in which defect or causation may be presumed, including particular breaches of product-safety requirements, obvious malfunction and circumstances in which complexity creates excessive difficulty and the claimant shows the relevant element is likely. Defendants retain the right to rebut the presumptions. By contrast, the Law Commission’s Consultation Question 33 asks only about presumptions of discrete factual matters, and Question 34 proposes the exceptional defect-election and risk-sharing mechanism. It does not provisionally adopt the full EU model of rebuttable presumptions of defect and causation in technically complex cases.[54]
There is a genuine choice between procedural philosophies. The EU model may give courts a direct route through complexity but risks shifting too much evidential uncertainty to producers. The Law Commission’s proposed hybrid is more restrictive and potentially more protective of defendants, but its 50 per cent mechanism is not a conventional response to a proven probability of liability. Nor does it directly solve an inability to establish specific causation. The distinctive UK approach should be justified through evidence of the practical costs and outcomes of each mechanism, not through the assumption that lesser burden-shifting must always be fairer.
The wider field of online platforms also requires care. The EU Product Liability Directive and the United Kingdom’s Online Safety Act operate alongside different regulatory frameworks and constitutional arrangements. A technology business may have compliance obligations under several regimes, but those duties are not interchangeable with the elements of a civil product claim. Neither foreign legislation nor the existence of detailed regulatory obligations removes the need for domestic Parliament to decide whether a particular category of digital feature should be a ‘product’ and what compensable damage it can cause.[55]
Four hypothetical claims expose the limits of the proposals
The legal effect of the reform is most easily understood by separating three questions: would the challenged functionality be a product; might that product be defective; and can the claimant prove that it caused compensable damage? The following examples are illustrative, not accounts of decided cases or predictions that particular claims would succeed.
A hazardous medical recommendation
A company supplies an AI tool marketed specifically to provide medication-management recommendations to patients. It gives dangerously incorrect guidance in a foreseeable use case, despite a known risk that could have been materially reduced through an available safeguard. A patient reasonably relies on the guidance and suffers physical injury. Under the proposals, the system’s software would be a product irrespective of cloud delivery. Defect might be established through unsafe design, inadequate warnings or the feasible risk reduction that was not adopted. The claimant would still need to establish the content of the recommendation, appropriate comparator evidence, the injury and causation. A generic disclaimer may be relevant but would not necessarily neutralise an unsafe product sold for high-stakes use.
An ordinary chatbot supplies an inaccurate factual answer
A general-purpose chatbot invents a historical reference in response to a broad research question. The user spends time correcting it and misses a deadline, but suffers no personal injury or qualifying property damage. Treating the AI system as a product would not transform the mistake into an actionable CPA claim. The system may not have failed the objective safety standard in that context, and the loss complained of may fall outside the CPA’s damage categories. Contractual terms, consumer rights or professional duties of a human who adopted the output might be relevant. The example demonstrates why product classification must not be confused with a guarantee of factual accuracy.
A recommender and prolonged psychological injury
A teenager develops a clinically recognised psychiatric condition after repeated exposure to harmful recommendations allegedly driven by a platform’s optimisation model. The recommendation feature could fall within the proposed product category, and the claim might identify inadequate protection or a materially unsafe design. But the litigation would have to distinguish the algorithm’s effect from prior vulnerability, material sought by the user and other relevant circumstances. It would need evidence of actual exposure and clinically supported causation. The possibility of a claim is not proof of legal responsibility. Here the proposed evidence rules and preservation of platform data may determine whether the case can meaningfully be tried.
A security update damages a connected device
A manufacturer remotely changes the software operating a domestic connected product, and the updated system creates a new dangerous operating condition which damages other private property. The relevant questions include whether the modified product failed the safety standard, whether the change was within the original producer’s control, whether statutory defences are available and when the long-stop began. Under the proposals a sufficiently substantial modification could restart the limitation clock. The example illustrates why retaining post-sale control may justify continuing responsibility, while the same analysis would not automatically treat an unrelated third-party alteration as the manufacturer’s defect.
Access to justice requires more than a new definition
The principal normative attraction of the consultation is that it allocates responsibility by reference to unsafe functionality and commercial control rather than technological form. The danger is that the benefits may remain primarily theoretical. An injured consumer may lack funds to retain technical experts, may be unable to obtain historic system data, and may face multiple possible defendants. Psychiatric-injury litigation may require expensive clinical evidence, difficult individual causation analysis and evidence from a platform that has strong commercial reasons to resist disclosure. Statutory liability without a practical route to establish its elements is an incomplete remedy.
The Commission’s proposals to facilitate substitution of mistakenly named or identified corporate defendants address one aspect of that problem. They do not solve insolvency, the cost of expert testimony, the absence of records or the procedural challenge of claims involving many users with similar but not identical histories. The consultation explicitly seeks information about the economic effects of reform, including costs to consumers and businesses, insurance and equality impacts. Those issues should inform the design of procedural safeguards as much as the scope of substantive liability.[56]
Targeted obligations concerning record preservation, relevant model versions, documented safety interventions and disclosure would strengthen the proposal. Their proper scope should be risk-based: a consumer photo filter should not necessarily face the same record-keeping burdens as an automated safety-critical system. Privacy and the interests of other users must be protected, and the law should not presume that an AI provider can supply a humanly comprehensible explanation of every internal calculation. But a developer should ordinarily be able to disclose what model it supplied, the material safeguards it deployed, what it knew about a relevant class of risk and which changes occurred before or after an alleged injury. Courts can regulate confidentiality and relevance without accepting an across-the-board claim that technological opacity defeats scrutiny.
Effective legal design also depends upon proportionality in litigation. The likely cost of expert investigation should not exceed any conceivable recovery for ordinary consumers. Representative and group proceedings may reduce repeated costs where genuinely common product defects are alleged, but they do not eliminate individual causation in psychiatric claims. Early neutral technical assessments, staged disclosure, narrower issues trials and appropriate cost management may prove valuable. These are suggested procedural responses, not enacted changes contained in Consultation Paper No 279. They ought to be considered alongside substantive reform if Parliament intends the extension of liability to result in actual compensation rather than symbolic recognition.
Should software developers face strict product liability?
A defensible answer is yes, but only for a suitably defined class of unsafe product functionality and with carefully calibrated evidential and substantive limits. There is no good reason why software which causes physical injury should escape the safety regime solely because its code was downloaded or executed remotely. Nor is there a principled basis for exempting a company’s own recommendation architecture from scrutiny merely because the output consists of user-generated material. In both situations, the producer has engineered a system with identifiable purposes, settings and risks. The inquiry should focus on the safety of those choices and their consequences, not the physical form in which they are delivered.
At the same time, the case for reform should not rest on a technological version of absolute liability. A software product is not automatically defective because it makes a mistake, a person suffers distress, or a foreseeable hazard occurs. The statutory requirement of entitled safety must continue to evaluate all circumstances, including the benefits and inherent limits of the technology, its target users, risk information, reasonable precautions and relevant alternatives. The seven proposed principles can provide useful structure, provided they are applied as genuine tests of product safety rather than an invitation to reconstruct a negligence case against every developer.
The most pressing safeguards can be expressed as five propositions. First, Parliament should define functional software and platform design separately from pure expressive content, while recognising that content may be evidence of how an algorithm operated. Secondly, the legally responsible defendant should be identified through development, presentation and meaningful control, with room for more than one party to be liable. Thirdly, compensable psychiatric injury should remain clinically meaningful and causally attributable, while the artificial limitations of sudden shock and physical proximity should not automatically govern product claims by closely connected indirect victims. Fourthly, evidential rules should address technical information held only by providers without simply reversing every burden of proof. Fifthly, limitation and defences should reflect material post-supply changes and continuing producer control without imposing perpetual support obligations.[57]
Of those issues, the Commission’s novel defendant-election proposal merits the closest scrutiny before enactment. A default starting point of 50 per cent damages is a significant departure from the conventional all-or-nothing proof of defect. Its appeal lies in preventing an absence of accessible evidence from always favouring the party controlling the product. But fairness requires a stronger account of why equal loss-sharing is the correct default, how the threshold of exceptional difficulty will be applied, and what happens when the critical uncertainty concerns causation instead. Reforms which require preservation and disclosure of relevant facts, and permit limited presumptions where one party has a demonstrable information advantage, may more directly address the practical problem without weakening conceptual coherence.
There is also a danger of substituting civil litigation for safety governance. The Product Regulation and Metrology Act 2025 provides a modern statutory framework for product regulation, while the Online Safety Act 2023 establishes distinct regulatory responsibilities for online services. Product liability should complement rather than displace regulation, enforcement and professional responsibility. The better objective is a system in which designers have incentives to prevent reasonably avoidable harm, regulators can act before injury becomes widespread, and injured individuals have a credible remedy where a defective product has actually caused compensable damage.[58]
Conclusion
The Law Commission’s October 2026 consultation identifies a genuine mismatch between the structure of the Consumer Protection Act 1987 and the present market for software and AI systems. The distinction between a dangerous algorithm embedded in a physical product and the same algorithm supplied through an online service is increasingly hard to defend. So too is an assumption that platforms cannot create product risks through their own designed recommendation and engagement mechanisms. On these points the case for modernisation is strong.
The harder issue is translating that extension into fair liability. Hastings remains a warning that strict liability is not absolute: the claimant must establish an objectively defective level of safety, actionable harm and causation. AI systems are often adaptive, interconnected and technically opaque, and psychiatric injury may develop in circumstances involving many interacting causes. These features challenge the ascertainment of defect and the feasibility of individual proof; they do not make those legal elements dispensable. The Commission’s seven proposed defect principles would bring useful structure, while its proposals concerning producer control, updates, recognised psychiatric harm and extended long-stop periods respond to real gaps. The exceptional-evidence election between rebutting defect and sharing damages is more debatable and deserves careful testing against less disruptive procedural alternatives.
The preferable legal position is therefore conditional liability based on unsafe design and function, proven causal injury and fair access to the relevant evidence. A system which cannot be made safe for its advertised high-risk purpose should not escape responsibility simply because its developer cannot predict every individual output. Equally, a generally useful system should not incur product liability merely for a lawful communication, an unavoidable technological limitation or a harm it did not cause. Reform should distribute the costs of demonstrably unsafe digital products without creating a universal insurer for the hazards of modern communication. Whether the Law Commission can achieve that balance will depend not only on what Parliament eventually calls a product, but on how it defines the boundaries of defect, causation, responsibility and proof.
References and primary sources
Numbered references link to legislation, judgments and the original Law Commission consultation. Pinpoint paragraph references identify the passages relied on; the essay distinguishes current law from provisional proposals and exploratory consultation questions. All online material checked to 6 November 2026.
- Law Commission, Product liability project (consultation launched 8 October 2026); Law Commission, Product Liability, Consultation Paper No 279 (2026), executive summary and ch 1.
- Law Commission, ‘Law Commission proposes major overhaul of product liability law for the digital age’ (8 October 2026); Ministry of Justice consultation portal (closing date: 14 January 2027). Consultation proposals are not operative legislation.
- Consumer Protection Act 1987, s 2(1)–(5) (identity of potentially liable producers, importers, own-branders and suppliers; joint and several liability).
- Consumer Protection Act 1987, s 3(1)–(2) (objective safety standard and relevant circumstances).
- Consumer Protection Act 1987, s 4 (statutory defences); Law Commission, Consultation Paper No 279, ch 6.
- Consumer Protection Act 1987, s 5 (recoverable damage and private-property threshold), and s 45 (meaning of personal injury, including impairment of mental condition).
- Hastings v Finsbury Orthopaedics Ltd [2022] UKSC 19, particularly [15]–[21] (entitled safety, time of supply and the claimant’s burden).
- Hastings v Finsbury Orthopaedics Ltd [2022] UKSC 19, particularly [22]–[31] (the evidential treatment of revision rates and competing explanations).
- Wilkes v DePuy International Ltd [2016] EWHC 3096 (QB), particularly [68]–[105] (the statutory approach to defect, circumstances relevant to safety and warnings).
- Gee and others v DePuy International Ltd [2018] EWHC 1208 (QB), particularly [64]–[186] (defect and risk assessment); Law Commission, Consultation Paper No 279, para 4.13.
- Consumer Protection Act 1987, s 1(2) (the existing meaning of a product); Law Commission, Consultation Paper No 279, paras 2.14–2.43.
- UK Jurisdiction Taskforce, Legal Statement on Liability for AI-related Harms (July 2026), summary propositions S.2–S.8. The statement expresses expert legal opinion and is not binding case law.
- Law Commission, Product Liability, Consultation Paper No 279 (2026), paras 2.49–2.51 and Consultation Questions 1–2 (software, method of supply and AI models).
- Law Commission, Consultation Paper No 279, paras 2.52–2.76, especially the treatment of AI models and the consultation on software supplied without physical media.
- Law Commission, Consultation Paper No 279, paras 2.55–2.67 and 2.133–2.148, Consultation Questions 9–10 (digital functionality, information and expression).
- Law Commission, Consultation Paper No 279, paras 2.149–2.165, Consultation Questions 11–13 (bespoke software and the product–service boundary).
- Law Commission, Consultation Paper No 279, paras 2.90–2.124, Consultation Questions 6–7 (design components of online platforms and recommendation systems).
- Law Commission, Consultation Paper No 279, paras 2.117–2.124 (assessing the relevant platform design or component rather than treating each item of content as a product).
- Law Commission, Consultation Paper No 279, paras 2.96–2.104 and 5.41–5.49 (expression and harm); Human Rights Act 1998, Sch 1, Art 10 (freedom of expression).
- Online Safety Act 2023, especially ss 9–12 (regulatory duties in relation to relevant online services).
- Law Commission, Consultation Paper No 279, paras 2.87–2.89 (relationship between online safety regulation and private compensation).
- Law Commission, Consultation Paper No 279, paras 4.33–4.40, especially para 4.39 (the seven proposed principles for assessing whether a product is defective).
- Law Commission, Consultation Paper No 279, paras 4.43–4.53 (warnings, instructions and product presentation).
- Law Commission, Consultation Paper No 279, paras 4.54–4.63 (possible risk-reduction measures, cost and other safety considerations).
- Consumer Protection Act 1987, ss 2–3; Law Commission, Consultation Paper No 279, paras 4.98–4.104 (elements of the claim and proof).
- UK Jurisdiction Taskforce, Legal Statement on Liability for AI-related Harms (2026), summary propositions S.9–S.11 (factual and legal causation).
- Law Commission, Consultation Paper No 279, paras 4.99–4.107 (technical and evidential obstacles in product liability claims).
- Law Commission, Consultation Paper No 279, paras 4.111–4.116, Consultation Question 33 (possible presumptions about discrete factual matters; an exploratory question).
- Law Commission, Consultation Paper No 279, paras 4.135–4.140, Consultation Question 34 (exceptional difficulty and the defendant’s election between rebuttal and a provisional 50 per cent sharing mechanism).
- Law Commission, Consultation Paper No 279, paras 4.137–4.141 (remaining need to establish damage and causation, and potential adjustments to apportionment).
- Consumer Protection Act 1987, s 45; Law Commission, Consultation Paper No 279, paras 5.21–5.26.
- Law Commission, Consultation Paper No 279, paras 5.27–5.51, Consultation Questions 37–38 (directly caused psychiatric injury, recognised conditions and ordinary fortitude).
- Paul v Royal Wolverhampton NHS Trust [2024] UKSC 1 (scope of secondary-victim psychiatric-injury claims in clinical negligence).
- Law Commission, Consultation Paper No 279, paras 5.52–5.75 and Consultation Questions 39–40 (indirect psychiatric harm and the proposed treatment of close relationships).
- Law Commission, Consultation Paper No 279, paras 5.62–5.76, Consultation Questions 40–41 (relationship categories and evidential safeguards).
- Law Commission, Consultation Paper No 279, paras 3.1–3.27, including Consultation Question 16 (the producer and software developer).
- Law Commission, Consultation Paper No 279, paras 3.23–3.48 (continuing or delegated control, including software-related contributions).
- Law Commission, Consultation Paper No 279, paras 3.44–3.48 (illustrations concerning third-party software and products).
- Law Commission, Consultation Paper No 279, paras 3.30–3.43 and Consultation Questions 20–21 (substantial modification and potential responsibility of the modifying actor).
- Law Commission, Consultation Paper No 279, paras 3.49–3.77 (fulfilment providers, online marketplaces and the proposed ‘core participant’ conditions).
- Law Commission, Consultation Paper No 279, paras 7.10–7.20 (procedural questions where an appropriate corporate defendant was initially misidentified).
- Law Commission, Consultation Paper No 279, paras 2.173–2.177, Consultation Question 15 (commercial supply without an upfront purchase price).
- Law Commission, Consultation Paper No 279, paras 2.77–2.85 (open-source software and the appropriate scope of liability).
- Law Commission, Consultation Paper No 279, paras 4.64–4.71 (the safety relevance of updates and continued manufacturer control, without a general new obligation to update).
- Law Commission, Consultation Paper No 279, paras 3.30–3.43 and 4.66–4.70 (substantial modification contrasted with routine safety measures).
- Consumer Protection Act 1987, s 4(1)(e); Law Commission, Consultation Paper No 279, paras 6.24–6.48 (development risks).
- Law Commission, Consultation Paper No 279, paras 6.15–6.23 and 6.49–6.63, Consultation Questions 48 and 51–52 (post-supply safety circumstances and defences).
- Limitation Act 1980, s 11A; Law Commission, Consultation Paper No 279, paras 6.64–6.67.
- Law Commission, Consultation Paper No 279, paras 6.68–6.99, Consultation Questions 53 and 55 (time limits, substantial modifications and latent injuries).
- Law Commission, Consultation Paper No 279, paras 5.85–5.93, Consultation Questions 43–44 (corruption or destruction of data and its distinction from data disclosure).
- Law Commission, Consultation Paper No 279, paras 5.77–5.107, Consultation Questions 42 and 45–46 (property damage, mixed use and the monetary threshold).
- Consumer Rights Act 2015, s 34 (quality of digital content) and s 46 (consumer remedies for certain damage to devices and digital content).
- Directive (EU) 2024/2853 on liability for defective products, Arts 2, 4 and 22 (software and temporal application; transposition by 9 December 2026). The Directive is used comparatively, not as applicable UK law.
- Directive (EU) 2024/2853, Arts 9–10 (disclosure and presumptions); Law Commission, Consultation Paper No 279, paras 4.111–4.116 and 4.135–4.141 (the distinct proposed English and Welsh approach).
- Directive (EU) 2024/2853; Online Safety Act 2023; Law Commission, Consultation Paper No 279, ch 2 (different legal frameworks and their scope).
- Law Commission, Consultation Paper No 279, paras 7.10–7.20 and ch 8 (procedural reforms, equality and economic implications).
- Law Commission, Consultation Paper No 279, paras 4.39, 4.111–4.141, 5.52–5.75 and 6.64–6.99. The five propositions in the essay are the author’s evaluation, not verbatim Commission recommendations.
- Product Regulation and Metrology Act 2025; Online Safety Act 2023. These are relevant regulatory contexts, not substitutes for a proven civil claim under Part I of the 1987 Act.