Redefining Defamation: Establishing Proof of Fault for Libel and Slander in AI Hallucinations
When Internet searches return wildly inaccurate or defamatory results generated by artificial intelligence (AI), who should hold the blame for the damage done to the subjects’ reputations? Decisions resulting from recent and ongoing defamation lawsuits concerning AI, such as Battle v. Microsoft Corp. and Walters v. OpenAI, indicate severe oversight regarding the potential severity of damages caused by AI hallucinations to the reputations of both private individuals and public figures. In order to better protect such individuals from future harm and encourage AI developers to resolve such errors in AI-generated content, new standards must be created to establish proof of fault beyond negligence and actual malice. AI models’ lack of legal personhood demands the creation of a hybrid standard of liability, in which developers may be held at fault for the use of unreliable sources in training AI models and the distribution of false information through their products.
When one inputs a question into any large language model (LLM) or search engine that provides AI-generated responses, there is always a risk that the result, which is based on data gathered from the Internet, will not be factually correct. These inaccuracies, commonly referred to as AI hallucinations, can be detrimental to the subject’s reputation and livelihood, particularly when the content of the hallucination is damaging or misleading. For this reason, several public figures have taken legal action against major technology companies and AI firms for defamation, alleging libel and slander on account of the spread of misinformation via their products. In Walters v. OpenAI, radio host Mark Walters sued ChatGPT developer OpenAI over an AI hallucination contained in an erroneous summary of another lawsuit generated by the chatbot. [1] The hallucination incorrectly stated that Walters, who was not involved in the lawsuit, had been charged with embezzlement. Ultimately, the court failed to find proof of fault on OpenAI’s part, either by negligence or actual malice, due to the administration of sufficient warnings concerning potential inaccuracies by the developers, consequently granting summary judgment in the defendant’s favor. This decision is indicative of the inability of current defamation standards to properly address and provide compensation for what would otherwise be viewed as negligent or malicious behavior involving LLMs.
Similarly, in Battle v. Microsoft, Embry-Riddle Adjunct Professor and Battle Enterprises CEO Jeffery Battle sued Microsoft over an AI-generated search result produced by the company’s search engine, Bing. [2] The result in question contained an AI-generated summary conflating the identities of Battle, an aerospace professor and businessman, with that of another Jeffrey Battle, a Taliban-affiliated terrorist and felon. Similar to Walters’, Battle’s case depends on his ability to prove at minimum negligence by Microsoft, or a failure on the company’s part to exercise reasonable care with regard to the information distributed, for the harm done to his reputation by this error. Due to the rigor of the existing standards, such arguments have been historically difficult to achieve, especially if the court determines that Battle is to be treated as a public figure, which would only further heighten the requirements for establishing proof of fault and detract from the feasibility of his case. However, this dispute is currently in the process of being resolved out of court, in accordance with a late 2024 court ruling compelling the case to arbitration, meaning the pending outcome will not be bound to the previous decisions.
With so little precedent in use regarding this incredibly new sect of defamation law, looking back on historic decisions in defamation law is a valuable step in gaining a better understanding of how such cases should be handled moving forward. The outcome of past cases have helped courts determine how public figures should be distinguished from the rest of the population in terms of the protections guaranteed by current legislation, as well as by what metric liability may be imposed for libelous statements. New York Times v. Sullivan (1964) is a landmark case in defamation law, which set the standard by which liability for libel may be assigned. In the case, the New York Times appealed a previous court decision that had awarded Public Safety Commissioner L.B. Sullivan damages for falsehoods contained in an advertisement published by the Times. [3] The court determined that the Times could not be held liable for the errors, as Sullivan was unable to prove the Times’ prior knowledge or recklessness regarding the inaccuracy of the information.
Since then, case law has aimed to define the boundaries for situations in which entities are potentially liable for the creation and distribution of libelous material. The outcome of New York Times v. Sullivan set the precedent for the use of the aforementioned actual malice standard, or the legal requirement that necessitates prior knowledge or recklessness with regard to defamatory statements. In a subsequent case, Gertz v. Welch (1974), attorney Elmer Gertz filed libel action against Robert Welch Inc. for framing him as a communist in their publication. [4] The outcome of this case showcased the distinction in standards for public figures compared to ordinary citizens and private individuals, the latter of which should be afforded more protection. Both decisions demonstrate how especially difficult it can be to establish proof of fault for public figures, whose reputations are perhaps most at risk of suffering harm by AI hallucinations. Furthermore, the unique qualities of AI and LLMs, particularly their lack of legal personhood, or capacity to engage in legal activity as an independent entity, generate additional complications for the attribution of libel and slander.
As the courts attempt to place the recent cases regarding AI hallucinations and defamation within the longstanding precedent of fault-based standards, the need for more specific legislation regarding this issue becomes apparent and calls for urgent address. If the standards currently set in place, such as the actual malice and recklessness tests, are upheld and applied to contemporary cases, it would be difficult to establish proof of fault by technology companies or developers for any incorrect statements made or generated by their product as a result of AI hallucination. [5] As a non-human entity, whether or not an AI model is capable of producing output that could be considered in the conversation of liability for libel and slander presents conflict amongst legal scholars. [6] Plaintiffs cannot sue AI models, which in these cases are the actual creators and distributors of the libelous material, thus creating a more convoluted trajectory for assigning fault. This ambiguity raises a question about whether technology companies and developers whose products generate output using AI should be regarded as responsible for the content of such output in the same way that traditional news or media organizations would be held accountable for the content they publish. To do so, the standard by which proof of fault is established should be adjusted to accommodate these differences. While the ability to assign accountability for the errors in AI-generated responses output by their products will depend on the courts’ upcoming decisions in these foundational cases, the current precedent indicates a need for a modified negligence standard and a reimagined definition of defamation to accommodate the nuances of modern technology.
Edited by Yoona Lee and Andrew Chung
[1] Walter v. OpenAI, LLC, 1:23-cv-03122 (N.D. Ga.).
[2] Battle v. Microsoft Corporation, 1:23-cv-01822 (D. Maryland).
[3] New York Times Company v. Sullivan, 376 US 254 (1964).
[4] Gertz v. Robert Welch Inc., 418 US 323 (1974).
[5] Leslie Y. Garfield Tenzer, “Defamation in the Age of Artificial Intelligence,” NYU Annual Survey of American Law (2023).
[6] Eugene Volokh, “Large Libel Models? Liability for AI Output,” Journal of Free Speech Law 489 (2023).