AI Liability: No Regulation, No Caselaw? No Problem For English Law Says The UKJT

The UK Jurisdiction Taskforce (UKJT), a UK Ministry of Justice backed initiative, has published a legal statement on liability for non-deliberate harm caused by artificial intelligence (AI) systems under English law, following a public consultation on the issue. The focus of the statement is on the ability of English law to provide sufficient redress for harms caused by the use of AI, given the novel issues that may arise from such use.

Summary

• The UKJT considers that English common law is sufficiently flexible to address AI-related harm, with courts expected to adapt existing doctrines incrementally by analogy. Contract will likely be the primary route for allocating AI liability, with negligence likely to provide the main framework where contractual allocation is absent.
• Duties and standards of care will be highly fact-specific across the AI supply chain, with potential liability implications for foundation model developers, application developers and deployers, and professionals using AI tools.
• Evidential and causation issues will be central to AI disputes potentially making record-keeping, due diligence, disclaimers, transparency and human oversight important practical safeguards.
• Potential liability may arise for negligent or even fraudulent misstatement by AI chatbots where such statements are made by or on behalf of a legal person or adopted by that person.
• Some of the UKJT’s findings may be deemed surprising, including that software updates could support an assumed duty of care by an application developer, that professionals may sometimes be negligent for failing to use AI, and that developers could face liability where they have created a “source of danger”.

The UKJT statement adopts a deceptively simple and all-encompassing definition of AI as “a technology that is autonomous”. It notes that key AI-specific difficulties in cases involving harm caused by AI, deriving primarily from the autonomous nature of AI, may include: 

• the unpredictable relationship between inputs and outputs
• the opacity of reasoning and
• limited user control over outputs.

The UKJT also identifies practical sources of uncertainty, including complex AI supply chains (involving data providers, foundation model developers, application developers, and users), potential scale and diffusion of harm, rapid system operation and the interaction between human and AI decision-making.

The UKJT ultimately emphasises the long-established flexibility of English common law, concluding that such difficulties will be no greater than those posed by novel technologies and developments that the courts have dealt with before. The UKJT considers that existing common law doctrines should be sufficient to determine liability for harm caused by AI; the courts will look to precedent and will incrementally extend the existing legal position through analogy. 

In this article, we seek to summarise the (lengthy) UKJT paper, drawing out some key themes, and then conclude with comments on some of what might be seen as more surprising conclusions. We will focus on liability in tort, rather than the allocation of risk under contract. We have also not sought to address the analysis in the paper regarding liability in defamation. 

Basis for liability

The UKJT identifies two routes to liability for AI harm: 

• responsibilities voluntarily assumed, most often under contract or 
• liability imposed by law, including under the law of negligence. 

In practice, the UKJT expects contract to be the primary basis on which liability is allocated among participants in the AI supply chain. The extent of liability in such cases will turn on the agreed contractual allocation of risk, including warranties, indemnities, limitations and exclusions, subject to the ordinary controls on freedom of contract. 

Negligence

Where contract does not govern liability for AI harm, the UKJT suggests that negligence will most likely provide the relevant framework. The UKJT considers that there is no conceptual reason why negligence principles cannot be applied to AI harm, although the outcome will be highly fact-sensitive. 

Negligence at common law requires a claimant to establish: 

• that a duty of care is owed 
• that there was a breach of that duty of care due to a failure to meet the required standard of care 
• that the breach caused the loss alleged and 
• that the loss was sufficiently foreseeable. 

Duty of care

The fact that a defendant (such as a professional) was using AI as a tool may not make much difference to whether or not they owe a duty of care to a third party. The UKJT uses the example of a radiologist reviewing medical imaging scans, who will owe a duty of care to the patient whether they use AI to assist with that review or simply review with the naked eye. 

However, the question of whether a duty is owed is likely to become more acute as focus moves up the supply chain. The developer of a narrowly targeted application—such as the tool used by the radiologist to review imaging scans—may owe a duty if it knew, or should have known, that errors in its output were likely to exist, to be difficult to detect by human radiologists and to cause harm.

The UKJT notes that the application developer (or any other person in the supply chain) may be more likely to have assumed a duty of care where they have previously taken active steps to push out updates to their software to address risks of harm.

In contrast, the UKJT considers that it is unlikely that a duty of care would be owed by a foundation model developer for harm resulting from unforeseeable or insufficiently tested uses of their general-purpose models. In particular, the UKJT comments that, on the facts, it may be difficult to show that a particular use of a foundation model was foreseeable, given it could be put to an extremely wide range of uses. 

This has ramifications for application developers and deployers. Absent a contractual (re)allocation of risk, liability for negligence is more likely to attach to these persons. This also highlights the issue for these parties of being in the “squeezed middle”. An application developer (for example) may seek to disclaim liability in its downstream customer agreements but may face a delta risk if it is unable to hedge this liability in its agreements with its upstream providers (such as foundation model developers). 

The UKJT also discusses where a party may owe a duty of care in its use of AI in relation to the negligence or wrongdoing of a third party. The obvious example is vicarious liability for the acts of an employee who misuses AI, but a party may also owe a “non-delegable duty” to particular parties to protect them against harm, such as the duty of an employer to provide a safe system of work for its employees, and the duty of a hospital to preserve the safety of its patients.

The UKJT notes that an NHS Trust which uses a third-party AI diagnostic tool could be liable for harm caused to a patient by a fault in the tool even if the Trust has taken all reasonable steps to ensure patient safety. It would have to seek contribution to that liability from the third-party developer of the tool under the Civil Liability (Contribution) Act 1978.

Finally, a party may be liable for the misuse of AI by another where it has created the “source of danger” which allowed the misuse or otherwise assumed a responsibility for it. The UKJT concludes that it is “not impossible” that a developer or other party in the AI supply chain might be held responsible for creating a source of danger by developing an AI model or system without suitable safeguards which it knows to be capable of certain types of harm if misused. However, it considers that “such cases are likely to be rare”, and that the more general purpose the AI system, the more extreme the circumstances would have to be for liability to arise. 

Standard of care

The standard of care will depend on what a reasonable person in the defendant’s position should have done. Where the defendant is a professional, the question is whether they have acted in a way that a reasonable body of the profession would have acted. This will change over time.

In the case of AI use, this will need to account for the rapid developments in AI and common practices for its use in a particular industry or profession. The courts will likely look to industry guidance, expert evidence, model selection and testing in assessing the relevant standard of care. Noting that the standard will be fact-specific, the UKJT predicts that a professional may be negligent for using AI inappropriately, using an unsuitable model, failing to conduct due diligence, failing to test or validate outputs, or, in some circumstances, failing to use AI where a competent professional would have done so. 

While acknowledging the context-specific nature of assessing whether an applicable standard of care has been breached, the UKJT does make some helpful general observations on the use of AI by professionals: 

• Failure to conduct proper due diligence on an AI system, particularly if it is new or innovative, or from an untested provider, and to check that it is appropriate for the task to be carried out, is likely to be a breach of duty. 
• Someone using AI to provide services for others should have a sufficient understanding of it. 
• A failure to be transparent with a client as to where AI has been used may be a breach of duty. 
• Putting a client’s confidential information into an AI system which does not sufficiently secure the confidentiality of that information is likely to be negligent. 
• A failure to exercise oversight over the AI system’s outputs (i.e., to put a “human in the loop”) is likely to be negligent. 

In circumstances where a foundation model developer or an application developer is liable, the partial defence of contributory negligence may reduce damages where the person harmed has themselves been at fault and that fault contributed to the harm suffered. The UKJT notes that this is more likely to arise in commercial contexts, for example where a user fails to follow instructions or warnings supplied with an AI system, or fails to intervene when errors first emerge. 

Strict liability

The UKJT notes that, in the absence of negligence or deliberate fault, those involved in the development, deployment and operation of AI systems generally will not be liable to parties with whom they have no contractual relationship. However, there may also be strict liability where AI is incorporated into a physical product (e.g., autonomous vehicle) and causes death, personal injury or damage to private property.

However, the UKJT notes that this route is likely to be narrow: the Consumer Protection Act 1987 will generally apply only where the AI forms part of a physical product, and not where the relevant AI is supplied purely as software or as a service. On July 31, 2025, the Law Commission announced that it intends to review the status of “pure software” under this law. As with many laws applicable to AI, clients should monitor developments in this area as part of their regulatory horizon scanning.

Causation—evidential issues

The UKJT recognises that disputes involving AI will likely experience material evidential challenges, particularly where the autonomous nature of AI and its opacity make it difficult to establish why a system produced a particular decision. This will likely cause difficulties with predicting or explaining events through traditional concepts of cause and effect.

In causation analysis, the UKJT suggests that the ordinary “but for” test will often remain workable, but difficulties may arise where qualities inherent in AI systems make it difficult or impossible to establish a counterfactual; for example, where evidence has not been gathered or retained, or cannot be reconstructed because of the autonomous nature of the system. 

The statement suggests that English law is capable of responding to such evidential gaps, including by drawing inferences or treating a claimant’s evidence more benevolently where the defendant’s failure to record or retain relevant information contributed to the uncertainty. It also notes that expert evidence and experimentation, including simulations comparing different inputs or guardrails, may assist in proving causation on the balance of probabilities. 

The UKJT also points towards the ability of the courts to come up with alternative approaches where there is a “systematic impossibility of proving causation”, for example, that taken in Fairchild.¹ In that case, it was impossible to prove “but for” causation due to the limits of scientific knowledge. 

However, the court found that the causation requirement was met by a defendant’s breach materially increasing the risk of injury being suffered. While this approach has not been applied outside a personal injury context, the UKJT notes that some argue it should not be so limited.

The UKJT notes the evolution of the technology to autonomous AI agents, where the role of the human user is to set tasks at a higher level, with agents completing sub-tasks with minimal human input. Multi-agent systems, or agents operating at higher levels of autonomy, can display unintended and unpredictable behaviour. The UKJT acknowledges that at higher levels of autonomy, courts may find it difficult to impose liability on AI providers without development of existing tort law norms.

Misstatement by AI 

The UKJT also explores the possibility of liability for harm caused by false statements made by AI chatbots. Given AI has no legal personality, any claim in relation to such statements must be directed at a legal person, such as a developer, deployer or user.

For negligent misstatement, liability will generally require that the statement was made by or on behalf of a legal person, or that the legal person adopted the statement. The UKJT considers that a developer of a foundation model or a particular application is unlikely to be regarded as “making a statement” to users of the chatbot. Instead, the focus is likely to be on the company which is using the chatbot, e.g., for customer services.

The UKJT suggests that a court is likely to treat chatbot statements as made by a business or person where the chatbot is presented as communicating with that business’s authority. Conversely, where the provider clearly presents the output as AI-generated and disclaims review of outputs, the provider may be less likely to be treated as the maker or adopter of the statement, although it may still make implied representations about the tool’s development, testing or accuracy.

Fraudulent misstatement (or deceit) raises a further complication because it requires human knowledge, belief or recklessness as to falsity, and an intention that the representation be acted upon. That requirement does not necessarily demand human review of each specific output, but it does require sufficient human involvement in the parameters, objectives or deployment of the AI communication.

The UKJT’s commentary therefore highlights the importance for deployers to: (a) include appropriate disclaimers when using AI chatbots, including by disclosing that outputs are AI-generated and may contain inaccuracies; and (b) consider the extent and process for human review and quality control of outputs. 

Comment

There are a few points that come out of the report that may be the subject of particular focus. 

As noted above, the UKJT comments that an application developer may be more likely to assume a duty of care if they issue updates to address risks that have been identified. On the face of it, this seems a somewhat surprising position with an obvious risk of perverse incentives: why should application developers risk increasing their liability exposure by taking steps to avoid harm? 

However, there is English authority on the circumstances in which a party may be regarded as having assumed a duty of care which supports such a position.² Furthermore, any perverse incentive to avoid providing helpful updates in order to mitigate against the risk of assuming a duty of care will probably be more than balanced out by the concern that a failure to provide such updates could: 

• be a breach of the standard of care that will apply should a duty of care nonetheless arise and
• increase the likelihood of a person suffering a loss (and therefore the risk of a claim arising in the first place).

Another point that has raised some eyebrows is the suggestion by the UKJT that a professional could in certain circumstances be held liable for not using AI. As the UKJT explains, in the professional’s hands, AI is just a tool and the question of whether or not it should be used is no different from that which arises in respect of any other tool. In that regard, it might be expected that, where a task can be appropriately performed with a variety of different tools, surely the choice of which tool to use should be for the judgement of the professional?

The UKJT’s example of the radiologist who fails to use the AI system that could have detected the cancer suggests that liability is likely to only arise where the choice of tool could have a major difference to the outcome, though other examples are given which would result in a less significant difference to the outcome. In any event, this statement underlines that companies should carefully consider when and how AI tools are used, both to address the risk of negligent use and possibly a negligent failure to use these tools.

Another area where there is likely to be focus is the potential for AI supply chain participants to owe a duty of care on the basis that it has created a “source of danger”. There has been a large amount of recent publicity over the release of new AI models which security agencies have suggested could massively increase cybersecurity threats. The UKJT considers that a developer of a general-purpose AI system may know that its products may be misused, but in this regard it is no different from the manufacturer of knives who knows that their products may be used in violent crime.

However, the analogy may not be as strong as the UKJT suggests, given the far more serious and widespread potential consequences of the misuse of an AI model. It is also worth noting that the English courts have indicated a willingness to entertain cases which push the boundaries of the law in this area.³

That said, given the English courts are also very reluctant to find a duty of care owed to an indeterminate class, the circumstances in which such a duty could be owed by the developer of a general purpose AI model are still likely to be very rare. Even if a duty of care is found owing, most established foundation model providers implement extensive safeguards which may demonstrate that the relevant standard of care threshold is met and they may seek to exclude or limit liability contractually (subject to the usual rules regarding exclusion clauses).

The approach that the English courts will take to liability for AI harm remains to be seen, but, for now, the UKJT’s statement provides a useful indication as to the direction of travel.

Footnotes

1. Fairchild v. Glenhaven Funeral Services Ltd [2002] UKHL 22.

2. The UKJT cites the case of Tulip Trading Ltd v. Bitcoin Association for BSV and others [2023] EWCA Civ 83, where the Court of Appeal held that it was at least arguable that the developers of open-source software which supports the trading of cryptocurrency owed a fiduciary duties towards cryptocurrency owners who had (allegedly) been the victims of cryptocurrency theft because the developers exercised a “positive” control over the software by taking active steps to update the source code where errors or bugs were detected.

3. In Begum v. Maran [2021] EWCA Civ 326, the Court of Appeal found that it was at least arguable that a shipbroker had assumed a duty of care by creating a source of danger when it sold a ship for scrap knowing that it was likely to be broken up in Bangladesh, given the dangerous working practices in the Bangladesh shipbreaking yards.