The legal repercussions arising from AI use are manifold – this much is evident from the UK Jurisdiction Task Force (UKJT)’s 130-page Legal Statement on Liability for AI Harms (“UKJT Statement”) under the private law of England and Wales published in July. One aspect of particular interest is liability that arises when AI tools make false statements, and the available legal defences in those circumstances.
Previously, we have considered how defamatory statements generated by AI may not necessarily have access to the honest opinion defence (see our video on AI, defamation and the question of opinion in our AI Watch video gallery), by virtue of the fact that to rely on the defence, the defendant must have actually held the opinion. AI tools are not conscious and so cannot hold opinions.
The UKJT’s statement seeks to address this and other thorny issues relating to defamation and AI.
False statements made by AI chatbots
The UKJT recognises the broad range of uses for AI chatbots, ranging from information and advice to emotional support. It points out that AI-generated statements are often believable and wrong at the same time. This combination can create significant harm from users or affected third parties relying on those statements. In the case of defamation, it is the subject of the false statement that may suffer harm from others relying on it.
Though AI systems are anthropomorphised due to their ability to interact with users in a way that feels like talking to a real person, the UKJT reiterates that the systems themselves have no legal personality. This makes the matter of recourse in relation to false statements generated by AI complicated – is action to be taken against the user who first generated the statement, or against the AI tool provider (and if so how far back in the chain do you go)? The answer is likely context based.
Defamatory liability
The UKJT Statement emphasises the “unique features” of defamation, and its historical context spanning centuries, with no analogue to other branches of tort. The UKJT acknowledges that the multiple defamatory defences were created and developed with the human publisher in mind, making the use of some (like the abovementioned honest opinion defence) difficult where an AI tool has been used without human review prior to publication. The substantial truth defence, which is an objective assessment, would appear to be available in respect of any statement, regardless of whether it was generated by a human mind or AI tool.
So, who is the actionable party in the context of AI-generated defamatory statements? The UKJT observes that where a user produces some words and incorporates them into their own writing, there are no special difficulties in relation to the liability of the user. The AI tool is merely another source of words that the user adopts as the author, editor or publisher. The complications arise where the tool generates defamatory statements which are then read directly by another person, without the user’s direct involvement. In that context, the UKJT recognises that liability is not limited to a single party. Various participants in the AI supply chain, whom the UKJT Statement refers to as “AI Supply Chain Actors”, may each independently attract liability as a “publisher”. Each AI Supply Chain Actor is to be assessed separately, depending on its role and level of involvement.
A person will be considered a publisher of a defamatory statement if they have participated or authorised its publication – passive facilitation will not suffice. This principle was illustrated in Bunt v Tilley [2006] EWHC 407 (QB), where it was set out that providers of access to the internet are not publishers of posts made by their subscribers, nor are operators of telephone systems in respect of slanderous calls. The same principle has been applied to search engine providers in respect of the automated publication of “snippets” of third-party websites (where the search engine provider has not yet been put on notice of any defamatory statements). The UKJT observes that this same principle could arguably apply to an AI Supply Chain Actor in respect of a tool that generates new statements based on third party content.
Whether or not the AI Supply Chain Actor incurs liability will be highly fact specific – the statement considers the following scenarios:
- Where the AI Supply Chain Actor exercises any form of manual review over the AI tool’s output before publication, they are more likely to be considered a publisher on the basis of their editorial role.
- Developers or suppliers who choose to deploy an AI tool to publish statements directly to the public will likely be deemed publishers, as this is a case of “robo-journalism”. However, the position might be less clear where the chatbot or search tool is presented to users as something apart from the Developer or Supplier’s own content.
- It could be arguable, depending on the facts, that some AI Supply Chain Actors should not be considered publishers unless and until they are on notice of publication of the statement complained of. The key factor here becomes the actor’s effective control over the publication – it follows that a foundation model developer with no involvement in publication beyond licensing the use of their model could be deemed a publisher to the extent they are on notice of a defamatory statement emanating from their model and if they have sufficient practical ability to control the model’s output, so as to prevent further publication of the statement.
The impact of warnings on meaning and serious harm
The UKJT Statement notes that context and mode of publication are central to the meaning a court will ascribe to potentially defamatory words. Where it is made clear to readers that a statement was generated by an AI tool, this may affect that meaning. In the context of a defamation claim, publications must, in any event, be read as a whole – meaning that surrounding explanatory text, such as warnings about a tool’s tendency to “hallucinate”, may influence the meaning ultimately attributed to the words, as may hyperlinked sources, depending on how they are presented.
A similar logic applies to serious harm to reputation, which the UKJT observes will depend on the effect of the publication and, in particular, the extent to which third parties regard the AI tool’s output as reliable. Where a tool is well known to hallucinate, or the risk of inaccuracy is otherwise flagged, the resulting harm to a claimant’s reputation may be reduced and could fall below the statutory serious harm threshold – as ever, much will turn on the particular facts.
Defences
As with many defamation claims, actual or threatened, the availability of a defence is often the real battleground. The UKJT observes that several common defences assume human authorship or review and are correspondingly less available in respect of wholly AI-generated statements.
Substantial Truth
The defence of substantial truth is entirely objective, and so is unaffected by the involvement of AI in the creation of a statement.
Reporting privilege
The “reporting privilege” defence protects reports of legal and parliamentary proceedings, public meetings and governmental publications. The availability of the defence turns principally on whether the published statement is a fair and accurate copy, extract or summary of the underlying event or document. As an objective test, this is unaffected by the involvement of AI in the creation of a statement.
For reports protected by qualified (as opposed to absolute) privilege, the defence can be defeated by malice. Malice can be established where the defendant reported material they knew to be false, or were otherwise recklessly indifferent as to whether it was true or false. This test is assessed subjectively – honest belief in the truth of the statement, even if unreasonable or arrived at carelessly, will generally defeat an allegation of malice. The UKJT Statement does not consider the impact of this subjective element, including the extent to which a person’s belief in false statements generated by AI (without any checks to determine accuracy) could amount to reckless indifference.
Honest opinion
The statutory defence of honest opinion is more problematic. Although it has an objective starting point, it will be defeated where the defendant did not, in fact, hold the opinion expressed – something an AI tool, lacking legal personality or consciousness, cannot do. The UKJT confirms that this defence is therefore unlikely to be available to an AI Supply Chain Actor in respect of an AI-generated opinion.
Public interest
Similarly, the public interest defence requires a subjective, reasonable belief that publication was in the public interest – a state of mind an AI tool cannot hold. The UKJT does suggest, however, that this defence could become available to an AI Supply Chain Actor once it becomes aware of the publication, if it then reasonably believes continued publication to be in the public interest.
Hosting
Developers and hosting suppliers operating websites may also seek to rely on the defence under Section 5 of the Defamation Act 2013, by showing that they were not the operator who “posted” the statement. The UKJT clarifies that “posting” is distinct from authorship, as a person may be considered to have “posted” a statement on a website even where it was AI-generated. An affected party may overcome the defence if they can show that the operator has failed to comply with the regulations applicable to it – whether a court would consider there to be a ‘poster’ of an AI-generated statement at all, and, if so, who that would be, or whether it would consider this provision inapplicable, will be a matter for determination.
Unknowing involvement in publication
Finally, Section 1 of the Defamation Act 1996 offers a defence to those unknowingly involved in publication who took reasonable care and had no reason to believe their actions contributed to a defamatory publication. This defence, designed to protect those involved in the publication chain of hard copies does not easily translate to the AI world. The UKJT notes that knowledge of an AI tool’s propensity to hallucinate is likely to undermine reliance on this defence in many cases, although it may still assist a developer or hosting supplier who has taken active steps to minimise defamatory output.
Conclusion
The UKJT’s Statement offers a thoughtful exploration of how established defamation principles may apply to AI-generated statements, while candidly acknowledging the ambiguity that remains in this area. Considerable scope for interpretation persists around the extent to which developers and other AI Supply Chain Actors may be treated as publishers and so incur liability, and around how far the statutory defences relating to hosting and unknowing publication will ultimately extend to AI-generated content. As AI tools become further embedded in the creation and dissemination of information, these are questions that courts – and potentially Parliament – will need to grapple with.
The analysis above is specific to defamation. A person could also seek to protect their reputation via data protection law (i.e. UK General Data Protection Regulation and the Data Protection Act 2018), which provides specific rights and remedies. The analytical framework used in data protection is distinct from that in defamation. The UKJT Statement expressly excludes data protection from its scope but it also has potential to raise complex questions for the court and/or Parliament.
Article co-authored by Alice Bolitho, CMS London.
