New Statement on Liability for AI Harms

What happens when someone suffers loss because of an AI chatbot’s mistake? The UK Jurisdiction Taskforce (UKJT) explored this question – and others around liability for non-deliberate harm – in its recent 130-page Legal Statement on Liability for AI Harms. While risk allocation will often be covered by a contract (between the AI provider and user, for example), these issues become more difficult where no contract exists.

Loss caused by AI use

Without contractual protection, the tort of negligence often determines who bears liability for non-deliberate harm. In many cases, the usual requirements (duty of care, breach of duty and resulting harm) will be easily applied. The UKJT Statement considers issues that are less straightforward. 

• Vicarious liability: While a person can be vicariously liable for someone else’s wrongs (such as an employee), an AI tool is not a person and so vicarious liability for the tool’s actions does not apply. That said, an employer could be vicariously liable for its employee’s actions when using AI. 
   
• Professional liability: Professionals must exercise reasonable skill and care, but what that looks like changes over time. The UKJT notes that professionals could be negligent for using AI inappropriately, using unsuitable models or failing to validate outputs, and it highlights the importance of human oversight. Perhaps surprisingly, it also suggests that professionals could be liable for failing to use AI where a competent professional would have done so, even though AI is just one of many tools a professional could use.
   
• Strict liability: Unless a contract says otherwise, the risk of harm arising without fault typically lies where it falls. But strict liability could arise for death, personal injury or damage to private property caused by a defective product. For AI, this is probably relevant only where AI is integrated into physical products (such as household appliances).
   
• Causation: A person will usually only be liable for loss that arose “but for” their acts or omissions, and only if no intervening act breaks the chain of causation. For example, the UKJT thinks it unlikely that a developer or deployer will be liable for misuse by a bad actor, unless the AI was obviously dangerous or the person should have prevented its misuse. By contrast, a developer or deployer could be liable for harm caused by AI acting autonomously, unless unforeseeable.

False statements by chatbots

Negligence is not, however, the only tort which may be relevant. When considering who bears responsibility for a chatbot’s false outputs (where this is not allocated via a contract), the UKJT Statement discusses: 

• Negligent misrepresentation: Tortious or statutory liability for negligent misrepresentation may arise where a false statement is made by a person or on their behalf. Whether a person is responsible for a chatbot’s output depends on the facts. The UKJT suggests that, when assessing negligence, the focus should be on careless acts that permitted the output – like decisions in design, testing and deployment – rather than the output itself.

  A person could be liable where a chatbot is held out as communicating on their behalf or where they represent (expressly or impliedly) that its statements are true. For example, a business may represent that information and facilities on its website (including chatbots) are authorised by the business, and a court would probably treat the statements as being made by or on behalf of that business. But a person simply passing on information as AI-generated output – about which they have no knowledge or belief – probably has no responsibility for its accuracy.

• Deceit: The tort concerns false statements made without belief in their truth, intending that they be believed and causing the recipient to believe them true. The UKJT suggests a person will be liable for AI-generated representations only if they had sufficient knowledge of the representation’s parameters and intended it be acted on by others, or if they were reckless as to its truth.
   
• Defamation: In the AI context, liability for defamation depends on whether a person “published” its output. The UKJT suggests liability could arise where there was human review or where AI was deployed to publish statements publicly in the course of business. While liability for “robo-journalism” seems straightforward, identifying the publisher may be harder in other scenarios.

Comment

The UKJT Statement focusses on tort and shows that tortious liability for AI harms can be resolved through incremental extension of existing law rather than wholesale change. While autonomous AI tools present novel challenges, the UKJT rightly notes that tort law is flexible enough to accommodate new developments.