Libel via Language Models

Abstract

This paper explains how developers and users of large language models (LMs), may be treated by English and Canadian libel law. LMs could be economically significant, and the liability environment they exist in will affect where they are developed, who accumulates wealth from their development, and who bears the burdens of any negative consequences of their development. Understanding the existing liability environment allows both developers and policy-makers to make informed decisions — about which jurisdiction to offer services in and what to prioritise, for the former, and about whether the existing law serves desired policy ends, for the latter. LMs also raise challenging legal issues, because they undermine common-sense assumptions that are baked into existing legal doctrines.

Although the discussion may have broader implications for tort law generally, this paper focuses on the doctrine and theory of libel. Legal problems lie lurking in those doctrinal weeds and are helpfully revealed by the contrast between Canadian and English law. Minor jurisprudential differences in decisions from yesteryear may have significant consequences if they are followed when text is generated by LMs, rather than by people.

Keywords: Generative AI, Large Language Models, Defamation, Libel, Tort Law