John C.P. Goldberg

A.I. in the form of Large Language Models (LLMs) is altering the ways in which we work, learn, and live. Along with their many upsides, an already familiar downside of LLMs is their propensity to “hallucinate” – that is, respond to factual queries with predictions or guesses that are false yet proffered as true.¹ And some of these hallucinations are not merely false but defamatory. For example, if one were to query an A.I. program: “Of which crimes has Professor X of ABC Law School been convicted?,” it might respond with a fabricated list of offenses. When defamatory hallucinations occur, who faces (or should face) liability, and on what terms? In Large Libel Models? Liability for AI Output Eugene Volokh lays out with great care a detailed roadmap for answering these questions.

Much of Professor Volokh’s article is devoted to considering and rejecting grounds for supposing that creators and operators of A.I. models enjoy blanket protection from defamation liability. First among these is an argument based on the tech industry’s best friend: Section 230 of the federal Communications Decency Act (“CDA 230”). As courts have interpreted it, this statute confers broad immunity on internet platforms for defamatory content created by third parties that they host or provide.² As such, Volokh persuasively argues, CDA 230 will typically be of no help to A.I. companies, because, even if their models have been trained on third party texts, it is the programs, not a third party, that generates the defamatory content.

Volokh also defuses the suggestion that, because LLMs use predictive algorithms, and are designed for coherence rather than truth, readers will discount the accuracy of their output roughly in the way they would discount the validity of a document they know to have been created by monkeys using typewriters. While early LLM technology perhaps was expected to spit out a lot of nonsense, current versions are taking the world by storm precisely because they are increasingly reliable and have come to be seen as such. As Volokh points out, the mere fact that A.I. users are aware of the possibility of hallucinations does not distinguish them from newspaper readers who are aware that papers sometimes publish false stories. Both are entitled presumptively to treat a given item as factual. Relatedly, under extant law, the presence in an A.I. terms of use of a disclaimer that directs users to verify independently the accuracy of its outputs will not defeat liability if, given the context, a reasonable reader would believe the information being offered is factual.

Others have suggested that the treatment of A.I. under copyright law generates an argument against defamation liability. If one maintains, as some courts have held, that A.I.-generated works are not copyrightable by program creators or users, then it might seem to follow that these same entities cannot be deemed “publishers” of A.I. generated defamatory content. (In defamation law, a “publisher” is, roughly, anyone who communicates a defamatory statement about person P to at least one person other than P.)  But this conclusion doesn’t follow. The question of who can claim to own a creative work is different from the question of who is responsible for the circulation of a statement. As Volokh aptly observes, it has never been the case that the publisher of a libelous statement can avoid liability simply by establishing that the statement is not eligible for copyright protection.

Volokh next considers these issues at a more granular doctrinal level. One bit of good news for companies that create or operate LLMs, he maintains, is that courts are not likely to deem N.Y. Times v. Sullivan’s actual malice requirement met simply because the company deployed a program while aware that it will occasionally generate falsehoods. But, as Volokh also notes, even this protective cloak has limits. For example, if a public figure who claims to have been defamed by an A.I. model notifies the creator or operator of the hallucination and provides compelling evidence of falsity – and if the company can alter the model’s weights to fix the problem yet does nothing in the face of such evidence – a jury probably would be entitled to find actual malice.

Meanwhile, under Gertz v. Robert Welch, Inc., private figures suing for a statement on a matter of public concern and who can prove actual injury need only establish negligence to prevail on their defamation claims. Of course, if public figures can prevail in cases in which an A.I. company has been apprised of false and defamatory program output, then, a fortiori, so can these plaintiffs. Because of Gertz’s relaxation of the actual malice rule, moreover, some private figures might be able to prevail even when the defendant was not provided with specific information apprising it of the falsehood of the statement. Analogizing to cases of manufacturers selling negligently or defectively designed products that physically injure consumers, Volokh suggests, for example, that defamation liability might attach if the victim of an A.I.-generated libel proves negligence in the design or operation of the program (such as an unreasonable failure to include a technologically viable self-correcting function).

Given his conclusion that there are meaningful prospects of defamation liability in the A.I. domain, Volokh is moved to consider whether extant law ought to be changed in light of one or more of several concerns: that this liability will be crushing; that it will cause A.I. creators and operators to change their programs in ways that make them less useful; or that it will disadvantage new entrants relative to established entities that can better absorb the cost. Here the main possibility he entertains is the establishment of a new qualified privilege, expanding out from the privilege that many jurisdictions apply to reports of suspected child abuse, that would protect from liability anyone who makes a good faith effort to provide what is understood by recipients to be preliminary information that requires further verification before being acted upon. Although avowedly “not a cheerleader for the American tort liability system” (P. 539), Volokh rejects this envisioned privilege as overly broad and out of sync with an important reality: recipients of A.I.-generated information often will not treat it with the circumspection with which officials are expected to treat an uncorroborated allegation of child abuse.³

As the foregoing synopsis I hope demonstrates, Large Libel Models provides a very valuable service in charting out, at the dawn of the A.I. era, the terms on which defamation claims for A.I. hallucinations are likely to play out. As such, it is a must-read for practitioners and scholars working in these areas. For the most part, I find its analysis persuasive, particularly its bottom-line assessment that companies that provide A.I. using LLMs are substantially more vulnerable to defamation liability than are traditional internet platforms such as Google. I would suggest, however that the prospects for liability are in some ways less grim than Professor Volokh supposes, and will offer a different perspective on how disturbed we ought to be about the prospect of significant liability.

On the first point, much will depend on the defamation scenarios that actually occur with any frequency in the real world. A private-figure plaintiff who can prove that their job application was turned down because their prospective employer’s A.I. query generated a defamatory hallucination about them would seem to have a strong claim. By contrast, suppose that P (also a private figure) learns from their friend F that a certain query about P will generate a hallucination that is defamatory of P, but also that P does not know who among their friends, neighbors, and co-workers (if any) have seen the hallucination. It seems likely that P will face an uphill battle establishing liability or recovering meaningful compensation. Even assuming P can prove that the program’s creator or operator was at fault (assuming a fault standard applies), P is likely to face significant challenges proving causation and damages, particularly given modern courts’ inclination to cabin juror discretion on these issues.⁴ I suspect this is especially likely to be the case if the program includes – as many programs now do – a prominent disclaimer that advises users independently to verify program-generated information before relying on it. While, as noted, disclaimers do not defeat liability outright, they might well render judges (and some juries) skeptical in particular cases about causation and damages.

Apart from doctrine, one must also take account of realpolitik, as Volokh recognizes. Back in 1995, it took only a whiff of possible internet service provider liability for the tech industry to get Congress to enact CDA 230. And Volokh tells us that A.I. is already a $30 billion dollar business (P. 540). If, as seems to be the case, the political and economic stars favoring the protection of tech are still aligned, legislation limiting or defeating liability for A.I. defamation could well be on the horizon, particularly in the wake of a few court decisions imposing or even portending significant liability.

The foregoing prediction rests not only on an assessment of the tech industry’s political clout, but also on a read of our legal-political culture. For most of the twentieth century, courts and legislatures displayed marked hostility to immunity from tort liability. (Witness the celebrated abrogation of charitable and intrafamilial immunities.) Today, by contrast, courts and legislatures seem quite comfortable with the idea of immunizing actors from liability in the name of putative greater goods. Nowhere is this trend more evident than in their expansive application of CDA 230. While Professor Volokh worries about the prospect of ‘too much’ A.I. defamation liability, the more reasonable fear may be too little. Indeed, it would seem to be a bit of good news that extant tort law, if applied faithfully by the courts, stands ready to enable at least some victims of defamatory A.I. hallucinations to hold accountable those who have defamed them.

1. For example, when ChatGPT is asked to describe work being done in an academic field, it sometimes includes fictitious citations based on its prediction that the cited pieces ‘ought to’ exist. Joy Buchanan & Olga Shapoval, GPT-3.5 Hallucinates Nonexistent Citations: Evidence from Economics (June 3, 2023) available at SSRN.
2. The soundness of this interpretation is questionable. John C.P. Goldberg & Benjamin C. Zipursky, Recognizing Wrongs 319-37 (2020).
3. The article concludes by briefly considering when the operation of A.I. might generate liability for other torts, such as invasion of privacy or aiding another’s wrongful injuring of a third party. As to the latter, it observes that there is an interesting question as to whether the law should protect A.I. companies with something like the learned intermediary doctrine for any A.I. provided to professionals specifically for their use (e.g., to radiologists to help them interpret MRI results).
4. Robert D. Sack, Sack on Defamation: Libel, Slander and Related Problems §10:5:2, at 10-44 (4th ed. 2010) (discussing increasing willingness of courts to second-guess jury awards of damages); Restatement (Third) of Torts: Defamation and Privacy § 1, cmt. d (Prelim. Draft No. 3, Mar. 2023) (proposing to replace the traditional rule of presumed damages – according to which certain defamation plaintiffs who do not offer specific evidence of reputational harm can nonetheless receive a compensatory damages award set by a jury – with a new rule that would entitle such plaintiffs only to nominal damages awards).