Are chatbots capable of defamation?

As I and many other commentators have been explaining for several months now, the machine learning-based systems which drive ChatGPT and similar chatbots are not yet trustworthy. By design, they are inherently (and some say incurably) predisposed to produce non-factual information. Thus, a proper way of using such chatbots today presupposes that you must fact-check all their outputs.

Obviously, this means that if a ChatGPT-like bot confabulates some non-facts and this results in something resembling defamation, a person who knows how these chatbots work should not perceive their output as defamatory, because it’s just an accidental result of a statistics-driven wordplay.

But there are those who don’t see things the same way.

An Australian regional mayor, apparently, has recently been one of those concerned with a chatbot’s potential to tarnish his reputation by confabulating several non-facts, in particular that he allegedly went to prison. As a result, the mayor demanded the chatbot provider to rectify the situation.

Without striving to carefully estimate how this particular case may develop and not attempting to provide any specific legal advice, below I will delineate a few considerations which I think should be relevant for cases of such kind in general (sections 1-4), along with some practical notes for AI providers (section 5).

1. Fact of publication

As a first step, the court deciding on a defamation case involving a ChatGPT-like bot might be willing to consider whether a defamatory “publication” occurred. For the purposes of defamation law, that usually means a communication of an untrue reputationally-damaging allegation to a third party – someone other than the alleged victim of defamation.

In a chatbot context, that will require the alleged victim to prove that defamatory communication concerning him or her appeared in a chat session initiated by a third party. Normally, that would not be very difficult to ascertain in a typical situation where an alleged victim becomes aware of the incident through third party-provided screenshots and social media coverage.

2. Provider’s authorship or acquiescence

While it will depend on the jurisdiction, most realistically chatbot providers might be sued for defamation based on acquiescence: that is where such providers fail in reasonable time to filter out potentially defamatory chatbot statements after having been informed by the aggrieved party of the fact.

Besides acquiescence (e. g. if there was none, that is, the provider promptly filtered out potentially defamatory statements as soon as it was given notice), another possible basis to sue for defamation would usually be authorship or editorial negligence, where a mere fact of statements publicly appearing would suffice.

In ChatGPT-style chatbot cases, however, this might be much trickier to argue, as obviously a confabulation appearing in a chat box is usually not authored or supposed to be reviewed by any particular person, rather, it is a result of statistics-driven computational processes.

3. Average reader’s perception

To be considered defamatory, the publication needs to be perceived as such by third parties. The respective legal analysis may differ depending on which nation’s law applies to the case. By way of example, the English law is said to require an objective test of “how the average reader would understand the words” (Raifeartaigh, at 766).

That is, in a potential chatbot defamation case in England, the court will need to establish whether regular people can’t help but treat all outputs from a sophisticated word salad tosser on par with conscious human expression capable of reputational damage.

In an ideal world, the latter would be impossible.

In such a world, people who interact with a ChatGPT-like bot would clearly understand the following:

1)?????The outputs of such a chatbot are just statistics-driven word salad some parts of which some people occasionally find true and/or useful but which should never be taken at face value.

2)?????These outputs are neither pre-determined nor fully predictable nor post-reviewed by any moral agent (person) on behalf of the chatbot provider. If any of this was the case, the chatbot would not have the same efficiency and utility.

3)?????The chatbot itself obviously lacks its own moral agency.

As a consequence of points 2 and 3 above, the words produced by the chatbot, considered in the context of the chat interface, are not attributable to any moral agent. But, according to Peter Strawson, people as moral subjects normally always reserve reactive attitudes such as blame and resentment for other subjects with moral agency and normally take only objective attitudes towards objects such as chatbots (Mason, at 3-4):

Imagine, however, that you were to regard all of someone’s behaviour as if it were like the involuntary physical and verbal tics characteristic of Tourette Syndrome and, so, divorced from his attitudes toward you and his judgments about how to behave toward you. Imagine, that is, that?you were to regard his behaviour never as expressions of his will but as mere happenings – much as you regard the weather. You would thereby take toward him the attitude that is Strawson’s concern in speaking of the wholly objective attitude. To take the wholly objective attitude toward someone is to cease to regard him as a moral agent.

Accordingly, normally people should consider whatever is thrown at them during a chatbot session as “mere happenings”, such as involuntary verbal tics of a person with a Tourette Syndrome (that is, not blameworthy even if defamatory in form). Or, as an equally apt example, on par with unconscious utterances of someone in his sleep – as given in Green’s restatement of speech acts theory:

For instance, while talking in my sleep I might say, “I hereby promise to climb the Eiffel Tower,” without thereby making any promise. [N]either a sentence, nor even the utterance of a sentence, is sufficient on its own for the performance of a speech act, be it a promise or some other. […] The agent in question must not only make her intention to undertake a certain commitment manifest; she must also intend that that very intention be manifest.”

As suggested by Tiersma (at 303), to produce a defamatory effect, one needs to commit an act of accusation, which is done by successfully communicating an “utterance that intentionally attributes responsibility to someone for a blameworthy act”. In a case of chatbot utterances, there is no moral agent to whom we could honestly attribute the intention to accuse, which again brings us to the essential point: as long as we remain within the confines of the chat interface, the only thing we can observe coming from a ChatGPT-like bot are “mere happenings”, similar to the Tourette Syndrome-induced verbal tics or unconscious ramblings of someone in a state of deep sleep.

4. Users’ automation bias

But, as it happens, we do not necessarily live in an ideal world where the analysis from the previous section will be necessarily upheld by a court ruling.

Quite on the contrary, we live in a world where Xerxes allegedly had the sea whipped as punishment, where people in general are prone to all kinds of bias (including automation bias), and where this bias matters for the purposes of assessing the average reader’s perception as a matter of defamation law.

It appears that even following all explanatory coverage of the recent months, some members of the public (hopefully, a dwindling minority) still take the outputs of ChatGPT-like chatbots at face value, or at least fear that others might do so.

To me, this is a clear example of automation bias: treating a raw, unverified and unedited output of a knowingly hallucination-prone system as if that output should be perfect and error-free.

And it also appears, at least anecdotally, that people who grew up around simpler, engineered computing systems are likely to be especially at risk for that kind of bias. They have a hard time comprehending that modern machine learning-based software can be as much or more useful than “good old-fashioned” computing systems, but sometimes also unpredictable and untrustworthy by design.

As a result, it might well be possible that a court deciding on a defamation case will establish that average people in a particular jurisdiction indeed perceive the particular chatbot’s outputs as capable of causing reputational damage.

5. AI providers: practical steps to take

Minding the above factors, those who develop and deploy ChatGPT-like bots and similar generative AI technology cannot exclude the risks of liability by merely explaining the technical limitations and relying on disclaimers as well as contractual “service provided as-is” terms.

Reasonable further steps for the generative AI providers may include:

1)?????Managing user expectations. Providers should not overstate the accuracy and potency of their products, such as by insinuating that they’re just a step away from artificial general intelligence. Instead, the providers should proactively manage user expectations. Thus, the providers will reduce their exposure to all sorts of litigation, including not only fancy chatbot defamation suits but also more mundane claims based on deceptive advertising.

2)?????Implementing a notice and takedown procedure. Providers of AI chatbots and similar generative AI systems should establish a policy and a working procedure to enable the collection and prompt execution of public requests for filtering out defamatory and other objectionable information from AI system outputs. Failure to establish or promptly follow such procedures is an actual legal and reputational risk.

3)?????Implementing an adequate risk management framework. Commercial AI system providers should reconsider the practices where they maximise profits while externalising risks. Such AI system providers should consider upgrading from outdated corporate-centric risk management by adopting new instruments such as ForHumanity AI Risk Management Framework, which take a broader and more societally conscious approach to managing AI-specific risks.

4)?????Soliciting diverse inputs and multistakeholder feedback. Making sure that negative consequences of commercial operations are not dumped in an unmitigated fashion onto the wider society is an essential part of building and maintaining sustainable business. To that end, when designing, developing and commercialising their AI systems, providers need to reach beyond a traditional narrow range of stakeholders and solicit feedback which is sufficiently diverse and reflective of a wider society in which those systems operate.

If you’d like to find out more about AI risk management, check out my Responsible AI webinars. As my newsletter subscriber, you’re eligible for a 15 percent discount (use promo code NEWS).