AI hallucinations: a defamation claim waiting to happen

“I'm sorry, but I cannot engage in any form of defamation or provide negative statements about any individual, including Liam Tolen or anyone else”.

When expressly asked to defame a named individual, the above is an entirely sensible, reasonable response for a generative AI tool such as ChatGTP to produce. But express programming to avoid a defamation elephant trap is a probably on par with, Hans Brinker saving his village from flooding by blocking the hole in the dam with his finger.

It certainly does nothing to overcome the problem widely referred to as an AI Hallucination. IT giant and AI pioneer, IBM describes such incidents as “a phenomenon wherein a large language model […] perceives patterns or objects that are non-existent or imperceptible to human observers, creating outputs that are nonsensical or altogether inaccurate.”

AI Hallucinations range from the innocuous to the ridiculous, but the reputationally dangerous output from one of these AI fever dreams lies somewhere in between. It would be immediately obvious that something had gone wrong if a chatbot suggested that spiderman actor Tom Holland was responsible for the assassination of JF Kennedy.

But what if a mayor of a small Australian district was wrongly cited as having been imprisoned in relation to an actual foreign bribery scandal, in which he was in fact the completely innocent whistle-blower, against whom no charges were ever brought?

That is the position in which Brian Hood found himself. Unsurprisingly, Hood engaged lawyers to send what amounted to a Letter before Claim although it is not thought that proceedings have been issued.

Under English Law

To establish a claim in defamation, it must be established that there was a publication, to a third party, of a statement that would make the ordinary person think worse of them. Generally, there is no need to establish malice or even an intention to defame so the fact that a defamatory statement has been generated by a computer programme is not likely to present an immediate barrier on the basis of a lack of consciousness.

The claimant would also have to establish that they have been named or can be identified from the statement and, to establish if there was any defamatory meaning, the statement would have to be read as a whole – referred to by the court as taking the “bane and antidote” together. So perhaps a cautious algorithm could be “trained” to avoid saying something negative without steeping it in caveats and cautionary language in a way that a litigation averse journalist might.

Using the mayoral case study, it is taken as read that words imputing that an individual has committed a criminal offence would give rise to an action for defamation. It seems overwhelmingly likely that the English Courts would have no trouble in determining a statement alleging criminality, composed by AI would be defamatory if that same statement was considered defamatory when composed by a human author.

Artificially defamatory?

The claimant would of course still have to establish serious harm. This may be AI’s first line of defence. This battleground may include arguments around the potentially restricted audience an AI hallucination may have reached. It may well be that the defamatory statement would only be generated on a very specific question being asked or potentially there may be data available to the AI defendant to know precisely how many times the defamatory answer has been generated. That said, with the rise of AI, its ability to amplify the dissemination of information may ultimately be an aggravating feature.

In the absence of evidence of a substantial audience having been reached, an AI defendant may invite the court to conclude that likely limited dissemination of the defamatory statement would not make the “game worth the candle”. This is the language adopted by the court when striking out claims where there has been a defamatory statement but where there was not enough at stake to litigate over it. This is likely to be a heavily fact driven analysis. In the mayor’s case, evidence that the AI hallucination was brought to the mayor’s attention by his constituents is likely to mean that the serious harm test would be satisfied.

Omnipresent AI

By its nature, AI will likely bring an international element with it. This is an issue likely best examined via the lens through which the court considers defamatory statements published on the internet. When it comes to online defamatory publications, the court has made clear that there is no legal presumption of substantial publication within the jurisdiction. claimants will need to evidence that the material in question was accessed and downloaded to a sufficient extent within the jurisdiction. The AI defendant may well have the upper hand here if threatened with claims in a jurisdiction where the defamatory statement can be shown to have had insufficient engagement.

Continuing to look through the internet lens, it is arguable that the AI defendant may benefit from defamation’s short (one year) limitation period coupled with the single publication rule introduced in 2013. Prior to 2013, each time a website publication was accessed, the clock on limitation “reset” such that an article, which at first went unnoticed but for whatever reason came to prominence after two years, could still be actionable.

Given the single publication rule, if the AI defendant could establish it has first published the defamatory answer over a year ago, regardless of whether the claimant knew about it, it may find itself with an arguable limitation defence. Undoubtedly, this will be a nuanced argument before the court.

The AI Factor

It's not however, all weighted in favour of the AI defendant to which not all of those defences available to a human defendant will be open to it. Take for example, the defence of honest opinion. Honest opinion fails in circumstances in which it can be shown that the defendant did not in fact hold the opinion in question.

The subjective element of that test is likely to prove problematic for an algorithm which on the face of it is incapable of holding opinions. However, perhaps the journalist, republishing an AI hallucination can rely on its AI origins as evidence that it was a reasonable opinion to hold.

Ultimately, the AI defendant will not of course be the algorithm or any anthropomorphic manifestation of it. It will in fact be the company which owns and controls the software or the author / publisher which uses AI as its source without fact-checking.

Responsibility

It is evident that AI is capable of producing defamatory content so the central question is likely to be: will the English Courts be inclined to hold someone liable for a defamatory AI hallucination? The answer to that question is likely informed by how the court (or rather, the Judiciary) is approaching AI in more general terms.

Court of Appeal Judge, Lord Justice Birss recently told the Law Society in a speech: “I asked ChatGPT 'can you give me a summary of this area of law? … and I put it in my judgment. It’s there and it’s jolly useful… I’m taking full personal responsibility for what I put in my judgment, I am not trying to give the responsibility to somebody else”.

Echoing this sentiment of responsibility, the Master of the Rolls, Geoffry Vos has said: “Judges do not need to shun the careful use of AI … [b]ut they must ensure that they protect confidence and take full personal responsibility for everything they produce.”

As such, whether it’s the company behind the algorithm or a less than diligent journalist with a penchant for copy and paste, provided the publication itself is defamatory, the court is unlikely to have trouble in holding the “author” responsible despite the defamatory statements origins as an AI hallucination.

Liam Tolen is a senior associate at Ashfords LLP