How does the use of AI interact with tort liability? As AI becomes more integrated into the economy and business relationships, this will become an increasingly important question.
The UK Jurisdictional Taskforce (UKJT) recently published its final ‘Legal Statement on Liability for AI Harms’ (the UKJT Statement), which analyses how existing English law principles may apply to non-deliberate harms caused by AI. The UKJT Statement’s overall conclusion is that the flexibility of English common law means it is already well equipped to answer many questions regarding liability for AI related harms and rejects the premise that the absence of an AI specific liability regime leaves a vacuum.
In this article, we draw on some of the issues discussed in the UKJT Statement to consider how liability could arise in relation to the use of AI tools in professional negligence and negligent misrepresentation, by reference to two scenarios in the corporate world: an investment manager’s provision of services to a client and an M&A transaction. We conclude by considering what practical steps businesses can take to manage the risks of liability to third parties in negligence arising from the use of AI.
Threshold legal concepts
We first cover some key legal principles which inform how liability may arise in negligence from the use of AI tools.
AI does not have legal personality
Under English law, liability can only attach to legal persons, i.e. an individual or corporate entity. As AI is not recognised as a legal person under English law, an AI tool cannot be held directly accountable for harms caused by it.
Vicarious liability is a legal concept which holds that in certain circumstances, a person who has a degree of control over another can be held liable for the other person’s acts, such as an employer being liable for acts of an employee. Because AI is not a legal person, vicarious liability cannot apply directly to AI tools. So, for example, the developer of an AI tool cannot be held vicariously liable for the acts of an AI tool.
How may liability attach for AI non-deliberate harms?
Liability for acts or failures by an AI tool must attach to a legal person. Generally, this will be through either: (1) an existing formal legal relationship between the parties, such as under a contract; or (2) a situation in which the law imposes liability on a party regardless of whether the party has chosen to assume liability.
Vicarious liability for a person’s use of AI could apply under this analysis. For example, if an employee uses AI in the course of their duties in a manner which is negligent and causes loss to a third party, their employer may be vicariously liable to the third party for any torts the employee commits.
As noted in the UKJT Statement, the law of negligence is the tort the most likely to be relevant to AI non-deliberate harms. Negligence is a broad and flexible form of tort law, which can apply in many situations and in both business-to-business and consumer contexts.
Negligence: The key principles
At a high-level, establishing liability in negligence involves:
- A duty of care, which arises where there is a relationship of proximity between two persons and it is fair, just and reasonable to impose a duty. A duty of care exists in many professional contexts, for example between a financial adviser and their client, or between a lawyer and their client.
- Failure to meet the standard of care required by the duty: the duty of care required will be to act as a reasonable person in such a position would have acted in the circumstances, and is judged based on the facts, often with consideration given to industry standards to determine reasonable practice. For a negligence claim, the tortfeasor’s conduct must fall below this standard.
- The tortfeasor’s negligence caused the alleged loss, including both factual and legal causation: To establish factual causation, any breach of the required standard of care must be a substantial cause of the loss, that is to say that “but for” the breach, the loss would not have occurred. Once factual causation is met, legal causation must also be established by determining whether there was an intervening act between the negligent conduct and the loss such that the chain of causation is broken. If no such act has occurred, the negligence will be the cause of the loss.
- The loss was sufficiently foreseeable to allow recovery: This extends to any type of damage that is reasonably foreseeable at the time that the duty of care is breached. The key element is whether the type of damages is foreseeable, which means this may extend to high levels of loss that were not in themselves foreseen, so long as the general category or type of damages was foreseeable.
- Liability for economic loss (i.e. where the loss does not flow from physical damage) will generally only arise when there is a “special relationship” between parties who have voluntarily assumed responsibility to the other, such as where there is a contract (such as a professional relationship) or negligent misstatement (where a representee is entitled to rely on a statement by a representor, regardless of whether there is a contract).
In practice, the requirement for a special relationship in order to recover economic loss means that in a commercial context, AI related negligence claims would most likely arise where there is a contract between the parties (such as professional negligence) or there has been a negligent misstatement.
We consider two scenarios below to explore how liability may attach to AI non-deliberate harms, one relating to professional negligence and one to negligent misrepresentation / negligent misstatement.
Professional negligence scenario: An investment manager makes investment decisions based on an incorrect AI output
Professionals generally owe a duty to perform their obligations with reasonable skill or care both under contract and tort (i.e. in negligence) and breach of their obligations may result in economic loss to their clients. Many investment professionals now frequently use AI tools in the course of their services to clients (and indeed, the use of AI tools is increasingly expected). A potential claim may arise where a professional relies upon the erroneous output of an AI tool to provide their services, in a way which causes their advice to be negligent.
Consider the following scenario:
An investment manager has been appointed to manage the investments of a private client based on a discretionary investment mandate. The investment manager uses an AI tool purchased from a third-party developer to analyse potential investments and assess them against the client’s investment objectives and risk appetite. The investment manager uploads the documentation of the client’s objectives and risk appetite to the AI tool. The output of the AI tool is used to assist the manager in ensuring that investments selected pursuant to the discretionary mandate are within the client’s risk appetite and consistent with the client’s objectives.
Following a significant market downturn, the client’s portfolio loses a large proportion of its value. The client makes a complaint to the investment manager. It conducts a review which identifies that some poorly performing investments were made in a way which arguably was not consistent with the client’s mandate and their objectives and risk appetite.
The investment manager identifies that this arose from investments identified originally by the AI tool, which interpreted one aspect of the client’s risk appetite in a different way to how the client now argues it should have been interpreted. The investment manager argues that, to the extent any investments were made outside the mandate, this arose from an error in the programming of the AI tool which it says it was unaware of and is the fault of the developer.
- As a starting point, the AI tool itself cannot be liable. The client may bring a claim against the investment manager in negligence for negligent breach of its duties. The client will need to show that a reasonably competent investment manager would not have acted as the investment manager did. The client would also have a claim for breach of contract under their contractual terms and conditions or investment management agreement with the investment manager (and in practice claims would always be made for breach of contract as well as negligence).
- Negligence will always be fact specific. The extent to which AI tools are used and relied upon by professionals and the level of oversight that should reasonably be exercised is likely to change over time, particularly if AI becomes more integrated into workflows. A relevant consideration could also be transparency around the use of AI. If an adviser or service provider has disclosed that their advice or services rely on AI and clearly identifies areas where they have not conducted detailed checks, potentially this could limit the extent of their liability.
- However, courts are unlikely to permit professionals to excuse liability entirely due to reliance on AI on a blanket basis, because clients rely upon professionals to provide specialist services. Further, guidance issued by regulators on AI often highlights the need to check AI outputs. Accordingly, disclosure of the use of AI (absent some very clear contractual limitation of liability) is unlikely to be sufficient to excuse a failure to exercise skill and care. In addition, the expected level of oversight of an AI tool will also be fact specific.
- The negligent use of AI may give rise to challenges in relation to causation and loss. Tort law depends on the use of counterfactuals in order to test “but for” causation. As AI becomes more integrated into workstreams, it may become more difficult to envisage a counterfactual where AI was not used, or was used in a different way. This is particularly the case with the use of agentic AI (i.e. AI systems that can generate outputs with a limited need for human supervision).
- In this example, if AI had not been used or was used correctly, how if at all would investment decisions have been made differently and would any loss have occurred? While these are difficult issues, complex counterfactuals and thorny issues of causation and loss are common in tort and something that the courts are used to grappling with. Further, in a professional scenario, the courts are likely to find a way to ensure that professionals are ultimately responsible for their advice.
- In this scenario, the investment manager might also claim against the developer of the AI tool. In such a claim, a key issue would be the scope of duty the developer has towards the investment manager under the relevant contract (and also potentially in tort) and whether there was in fact an error in the programming of the AI tool, or instead, for example, had the investment manager configured the AI tool incorrectly. Note that claims based on negligent programming in relation to AI are potentially complicated by the fact that, as outlined in the UKJT Statement, the functioning of an AI system develops through autonomous processes such as machine learning (and also the programming of an AI system is likely to have been undertaken at least in part using AI).
Negligent misrepresentation/negligent misstatement scenario: An AI-generated contracts summary is relied upon in an M&A transaction and turns out to be erroneous
Large language models such as chatbots are increasingly used to conduct analysis and generate outputs in the form of natural language. Ultimately, LLMs are reliant on the data available to them and can hallucinate (and sometimes have a tendency to be definitive in the views they produce). This gives rise to an important question: Can an AI user be liable for negligent misrepresentation or negligent misstatement where an AI-generated statement is communicated to a third party, relied upon and causes that third party to suffer loss?
Consider the following scenario.
A buyer becomes aware of a rare opportunity to buy a business that they have been interested in for a long time. The seller indicates that to close a deal, the transaction must be negotiated and completed over a very condensed time period. As a result, negotiations are handled directly between the principals. The buyer decides to conduct high level only due diligence, given the time pressure and because they know the line of business very well.
The target’s main subsidiary engages in trading with counterparties, based on hundreds of long-term contracts which have been amended and supplemented many times over the years. The buyer is familiar with the largest counterparties, but to conduct a review of even a sample of the contracts would be very resource intensive and challenging in the time period. For its internal purposes, when it decided to commence a sale process, the seller generated a summary of the contracts using an LLM-based AI tool. The seller provides the AI summary to the buyer on the buyer’s request shortly before the deal is signed. The buyer knows that it is AI-generated. Unknown to the buyer and the seller, the AI tool misinterpreted certain contractual clauses relating to change of control and termination rights, which under some of the larger contracts, give customers the right to terminate following a change of control of the parent.
The buyer is known in the market in which the business operates as being focussed on the bottom line and driving a hard bargain with customers. Following the acquisition, several counterparties reassess their commercial position and decide to terminate their contracts with the target’s subsidiary relying on change of control clauses in the contracts, which means that significant business is lost. The buyer conducts a review of the contracts to assess its legal strategy. It discovers that the AI generated summary of the contracts provided in the due diligence was incorrect in certain respects, in particular that several of the largest long term contracts can be terminated based on a change of control. It concludes that it significantly overpaid for the business given the probable loss of these contracts.
We first consider negligent misrepresentation, i.e. a cause of action where a false pre-contractual statement is made carelessly (i.e. non-deliberately) which induces a party to enter into a contract and causes it to suffer loss.
-
As it is not a legal person, the AI tool cannot be liable for negligent misrepresentation and cannot be the legal maker of the statement.
-
The buyer would need to claim against the seller for negligent misrepresentation. To establish liability, a threshold issue is that the statement that the representee (i.e. the buyer) was induced by must have been made by or on behalf of the representor (i.e. the seller).
- As noted in the UKJT Statement, there is currently no English authority on whether an AI tool can make a statement by or on behalf of a legal person in negligent misrepresentation, but the analysis that a court would take would be based on the facts.
- For example, did the seller present the AI generated contracts summary as being a document put forward on its behalf? In other words, even if the buyer knew that the contracts summary had been generated by AI, did the seller make an implied or express representation that it was made on the seller’s behalf? Or did the seller simply pass on the AI contracts summary to the buyer, making it clear that it assumed no responsibility for it?
- This is likely to be an area where the facts would be disputed between the parties. The seller may also rely on any relevant limitations of liability in the transaction document.
Suppose that the buyer is aware that the seller is unable to meet a claim for damages (for example, if it is an SPV with no assets). Could the buyer look to the developer of the LLM, arguing that it is liable in negligent misstatement? This is a tort which can apply where the defendant owed the claimant a duty to not carelessly cause the type of harm suffered. It can sometimes apply in similar situations to negligent misrepresentation, but does not require the statement to induce a party to enter into a contract and to be made by the other party to the contract.
- As claims in negligent misstatement do not require a contractual relationship between the parties, whether a duty of care exists at all is almost always disputed. The precise way in which a court analyses whether a duty of care exists has differed across the major authorities. However, key conceptual issues are the relationship between the claimant and the defendant and whether the defendant has assumed responsibility towards the claimant.
- In this context, a key consideration would be whether the AI developer (i.e. a legal person) has assumed responsibility for the output of the LLM to third parties that rely on it, such as the buyer in this example. In general, this will not be the case. A developer does not necessarily assume responsibility for the outputs of the LLM even to the immediate user (where there may be an existing contractual relationship, if the AI tool is a specialist tool used by businesses). Therefore, an assumption of responsibility towards third parties is unlikely. One reason for this is that, as the UKJT Statement notes, the output of an LLM is not pre-determined by the developer. The outputs of LLMs are heavily dependent on the prompts used by the user. A user can also use a prompt that can require an LLM to respond in a specific way designated by the user.
- Could there be exceptions? Possibly, depending on the facts. For example, if an AI developer releases a specialist AI DD tool, which produces outputs that have some kind of express assurance of accuracy such that the developer could be held to be assuming responsibility. However, this kind of situation is likely to be rare. In most cases, there will not be this fact pattern. Developing an AI tool that can DD related outputs and making it available in the market is unlikely to be sufficient for a court to conclude that the developer owes a duty of care to third parties that rely on the output.
- Even if a duty of care was found to exist, the third party would need to establish breach of that duty (as well as causation and loss). In this example, the buyer would need to establish that specific acts or omissions caused the AI developer to be in breach of duty. However, given the extent to which AI systems can produce outputs through autonomous processes (and the impact of the user’s prompts), this may be difficult to establish.
Conclusion and practical takeaways
The modern law of negligence has existed for nearly a century, tracing back to the 1932 case of Donoghue v Stevenson. It is flexible and has incrementally evolved over decades to adapt to social, economic and technological changes.
AI tools represent a significant development for the court to grapple with when applying the principles of negligence. In particular, the fact that AI tools such as LLMs can produce complex outputs written in natural language autonomously and with little human input presents challenges in terms of how liability should be allocated. However, because AI tools do not have legal personhood, liability arising from the use of AI will always attach to individuals or corporate entities and not the AI itself.
This means that the courts are likely to allocate liability using established concepts in the law of negligence, such as the extent to which the defendant assumed responsibility or exerted control, or the extent to which the claimant reasonably relied on the defendant in relation to a statement or alleged duty of care. The ways in which liability may apply are likely to be clarified in case law in the coming years; however, liability in negligence is always highly fact and situation specific. As AI becomes more integrated, the circumstances in which a negligence claim may arise are likely to increase. The duties of care that apply will evolve over time, particularly as AI tools continue to develop and become more integrated into workflows.
What can be done practically in the meantime? Businesses should understand how AI may impact their potential liability to third parties and take steps to manage those risks. For example:
- Ensure that standard form contracts which are used for contracting with third parties are kept up to date, e.g. consider whether changes are required to the scope and limitation of liability provisions in light of the use of AI tools in the provision of services to customers.
- Recognise that the use of AI tools in providing services to third parties will generally not excuse you from liability. Conduct appropriate training to employees to ensure that this is understood in the business, including by non-lawyers that are engaging with customers.
- Manage the risk of vicarious liability if an employee uses AI tools negligently in the provision of services to a third party. Consider implementing protocols on best practice for AI use, such as verification procedures which set out appropriate processes for checking and validating AI outputs.
- Be aware that, if you rely on statements generated by AI tools in terms of decision making, pursuing recourse is likely to be less straightforward than reliance on professional advice, where there is seldom a dispute that the professional owes duties to the client.
- Review contractual arrangements with the developers/service providers of AI tools. Ensure that you understand to what extent, if at all, the developers/service providers are taking responsibility for any loss that is caused to third parties through the use of their tool by your business.
Above all, businesses should monitor legal developments in this fast-moving area of law and evolve their risk management strategies accordingly.
With thanks to Emma Torok for her assistance in preparing this article.