Libel via Language Models

Authors

Peter Wills, Centre for Governance of AI

Document Type

Article

Abstract

This article 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 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 article 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.

Creative Commons License

This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.

Citation Information

Wills, Peter. "Libel via Language Models." Osgoode Hall Law Journal 62.1 (2025) : 153-199.
DOI: https://doi.org/10.60082/2817-5069.4099
https://digitalcommons.osgoode.yorku.ca/ohlj/vol62/iss1/5

References

1. JD (Harvard), LLM (Cantab), DPhil (Oxon). This article was written while I was a DPhil Candidate and a Research Scholar at the Centre for Governance of AI. I owe great thanks to the anonymous reviewers for their thorough and thoughtful critiques. Responding to those critiques materially improved this document. James Goudkamp, Rebecca Williams, Horst Eidenmuller, Tom Melham, Michael Collins, Owain Evans, Mary Angela Rowe, the participants in the UCL-Oxford Symposium on Law, Technology, and Fundamental Rights, the participants in the Oxford Future of Technology and Society Discussion Group, and the editors of the Osgoode Hall Law Journal all also provided valuable comments and criticism of different aspects of this article. Kimberley Paradis, with the support of the Centre for the Governance of Artificial Intelligence, provided assistance. I thank them all.

2. See Byron Kaye, “Australian Mayor Readies World’s First Defamation Lawsuit over ChatGPT content” (5 April 2023), online: Reuters www.reuters.com/technology/australian-mayor-readies-worlds-first-defamation-lawsuit-over-chatgpt-content-2023-04-05 [perma.cc/8WT9-JUME]; Pranshu Verma & Will Oremus, “ChatGPT Invented a Sexual Harassment Scandal and Named a Real Law Prof as the Accused” (5 April 2023), online: The Washington Post www.washingtonpost.com/technology/2023/04/05/chatgpt-lies [perma.cc/L6ZV-VJMX]; Chris Stokel-Walker, “‘AI Cannot be an Excuse’: What Happens when Meta’s Chatbot Brands a College Professor a Terrorist?” (25 August 2022), online: The Daily Dot www.dailydot.com/debug/meta-chatbot-blender-marietje-schaake [perma.cc/8X8E-4XZD]; Alex Hern, “Facebook Translates ‘Good Morning’ into ‘Attack Them’, Leading to Arrest” (24 October 2017), online: The Guardian www.theguardian.com/technology/2017/oct/24/facebook-palestine-israel-translates-good-morning-attack-them-arrest [perma.cc/UT88-LW3T].

3. I use “English law” to refer to both English and Welsh law. I use “Canadian law” to refer to that of the common law provinces and territories. Except where specifically noted, I treat Canadian libel law as unitary.

4. See Lachaux v Independent Print Ltd, [2019] UKSC 27 at para 1 [Lachaux UKSC]; Peter Wills, “Care for Chatbots” (2025) 58 UBC L Rev 525 (discussing the treatment of negligent misrepresentation).

5. See Daniel M Ziegler et al, “Fine-Tuning Language Models from Human Preferences” (8 January 2020), online (pdf): arxiv.org/pdf/1909.08593 [perma.cc/F7NF-R2PQ].

6. See Amanda Askell et al, “A General Language Assistant as a Laboratory for Alignment” (9 December 2021) at 8-9, online (pdf): arxiv.org/pdf/2112.00861 [perma.cc/LB63-LZP4]; Long Ouyang et al, “Training Language Models to Follow Instructions with Human Feedback” (4 March 2022) at 68, online (pdf): arxiv.org/pdf/2203.02155 [perma.cc/76EZ-XJJN].

7. See e.g. Hugo Touvron et al, “LLaMA: Open and Efficient Foundation Language Models” (27 February 2023), online (pdf): arxiv.org/pdf/2302.13971 [perma.cc/4AW9-JU3V] (publishing model weights albeit ones originally protected by a licence).

8. See Askell et al, supra note 6 at 6-7.

9. See Yuntao Bai et al, “Constitutional AI: Harmlessness from AI Feedback” (15 December 2022), online (pdf): arxiv.org/pdf/2212.08073 [perma.cc/MSH7-ZKDW].

10. See Amelia Glaese et al, “Improving Alignment of Dialogue Agents via Targeted Human Judgements” (28 September 2022) at 6-7, online (pdf): arxiv.org/pdf/2209.14375 [perma.cc/Y4PN-J8V6].

11. It may not be accidental in all respects, of course; and where the legally relevant respects are not accidental, this change is less important.

12. An API defines how one piece of software is supposed to be available to interact with other pieces of software.

13. One might expect there to be some contracts between Developer and the various Code Intermediaries and possibly even the controller of the ultimate interface and Human, and for these contracts to provide indemnities for defamation liability. I set aside the contractual overlay of who will actually bear the cost of tort liability; I focus on whom the law would hold responsible absent a contract. This pre-contractual analysis will be relevant for LMs developed or released by less sophisticated entities (possibly including open-source LMs), as well as in the event of insolvency of contractual parties that provide an indemnity.

14. See especially Erika Chamberlain, Karen Eltis & Raymond E Brown, Brown on Defamation: Canada, United Kingdom, Australia, New Zealand, United States (Carswell, 2024). This leading Canadian text on defamation flits lightly over jurisdictional lines. Indeed, tort law generally shows this kind of spriteliness. See also John Bell, “The Relevance of Foreign Examples to Legal Development” (2011) 21 Duke J Comp & Intl L 431 at 437.

15. James Goudkamp, “Rethinking Fault Liability and Strict Liability in the Law of Torts” (2023) 139 Law Q Rev 269 at 276.

16. Grant v Torstar Corp, 2009 SCC 61 at para 28 [Grant].

17. See e.g. Hodgkinson v Economical Mutual Insurance Co, 2003 CanLII 36413 at paras 23-24 (ONCA) [Hodgkinson].

18. See e.g. Huth v Huth, [1915] 3 KB 32 at 38-39 (CA); Hodgkinson, supra note 17 at para 23. See also Graxdy v McNichol, 1931 CanLII 548 (MBCA).

19. Defamation Act 2013 (UK), s 1.

20. See Introduction, above.

21. See SG v JC, 2001 CanLII 3041 at paras 19-20 (ONCA).

22. Bou Malhab v Diffusion Métromédia CMR Inc, 2011 SCC 9 at paras 35-36.

23. Ibid.

24. See Cassidy v Daily Mirror Newspapers Ltd, [1929] 2 KB 331 at 354 (CA).

25. See Hilary Young, “But Names Won’t Necessarily Hurt Me: Considering the Effect of Disparaging Statements on Reputation” (2011) 37 Queen’s LJ 1 at 8.

26. Richard Parkes et al, eds, Gatley on Libel and Slander, 13th ed (Sweet & Maxwell, 2022), s 3.016 [Gatley].

27. Ibid, ss 3.030, 3.036.

28. Emily M Bender et al, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” (Paper delivered at the Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, Virtual Event, Canada, 3-10 March 2021), online (pdf): dl.acm.org/doi/pdf/10.1145/3442188.3445922 [perma.cc/35GQ-44U9], DOI: https://doi.org/10.1145/3442188.3445922.

29. See Seth C Lewis, Amy Kristin Sanders & Casey Carmody, “Libel by Algorithm? Automated Journalism and the Threat of Legal Liability” (2019) 96 Journalism & Mass Communication Q 60, DOI: https://doi.org/10.1177/1077699018755983.

30. See Adam Slavny, “The Normative Foundations of Defamatory Meaning” (2018) 37 Law & Phil 523 at 524-25, DOI: https://doi.org/10.1007/s10982-017-9321-5. Slavny terms this the “Non-Moralised View” of defamatory meaning.

31. See e.g. Morgan v Odhams Press Ltd, [1971] 1 WLR 1239 at 1252 (HL (Eng)), Morris of Borth-y-Gest LJ; Oriental Daily Publisher Ltd v Ming Pao Holdings Ltd, [2012] 15 HKCFA 59 at para 145, Neuberger LJ; Randall v Weich, [1982] BCJ 862 at paras 20, 23-24 (BCSC) [Randall]; Kohuch and Loewen v Wilson, (1988) CanLII 4916 at para 151 (SKKB). For a discussion of this phenomenon, see also Young, supra note 25. From a more theoretical perspective, this approach accords with what Slavny terms “[t]he Moralised View.” See supra note 30 at 524-25.

32. Supra note 19, s 1.

33. See Lachaux v Independent Print Ltd, [2015] EWHC 2242 at 59-60 (QB) [Lachaux QBD]. The Court in Lachaux UKSC adopted Warby J’s interpretation in Lachaux QBD. See supra note 4 at para 20.

34. 1998 CanLII 4208 at paras 18-20 (BCSC), rev’d on other grounds 2000 BCCA 155. See also Loukas v Young, [1968] 3 NSWR 549 (NSWSC Austl).

35. See e.g. Gill v Maciver, 2022 ONSC 1279 at para 206; Penfold v Westcote (1806), 2 Bos & Pul (NR) 335 at 335 (Ct Com Pl UK); Bennette v Cohen, [2005] NSWCA 341 at para 51.

36. French v Smith, [1923] 3 DLR 902 at 903 (ONSC).

37. 1998 CanLII 13925 at paras 1, 6 (SKKB) [Roth]. Cf Ralston v Fomich, 1992 CanLII 1652 (BCSC) [Ralston]. In Ralston, the defendant calling the plaintiff a “sick son of a bitch” was found to be defamatory. Ralston should not win any contest against Poznekoff or Roth because the court applied the wrong legal test. Ralston was a bench trial, but the court applied the test for whether a matter should go to a jury (whether the words were “capable of a defamatory sense”), not the single meaning rule. See Lawson v Baines, 2011 BCSC 326 at para 40, aff’d 2012 BCCA 117 (for the correct approach in such situations).

38. Randall, supra note 31 at para 23.

39. Young, supra note 25 at 13.

40. Pizzagate involved a man who, believing a variety of alt-right conspiracy theories, strolled into a pizza joint armed with a semi-automatic rifle, anticipating that he would uncover a pedophile ring. See Amanda Robb, “Anatomy of a Fake News Scandal” (16 November 2017), online: Rolling Stone www.rollingstone.com/feature/anatomy-of-a-fake-news-scandal-125877 [perma.cc/JWE2-Y36B].

41. See e.g. Clare LE Foster, “Truth as Social Practice in a Digital Era: Iteration as Persuasion” (2022) 38 AI & Society 2009 at 2009-10, DOI: https://doi.org/10.1007/s00146-021-01306-w (summarizing some of the literature on the effect of repeating a false statement).

42. See “The Social Foundations of Defamation Law: Reputation and the Constitution” (1986) 74 Cal L Rev 691, DOI: .

43. Ibid.

44. Ibid at 715.

45. Ibid at 706.

46. Ibid at 711.

47. Ibid at 713.

48. See Soliman v Bordman, 2021 ONSC 7023 at para 151.

49. Ibid at paras 21-23.

50. Ibid at para 28.

51. Ibid at para 26.

52. Ibid at para 203.

53. Ibid at para 207.

54. Ibid at para 199, citing Hill v Church of Scientology of Toronto, 1995 CanLII 59 at para 182 (SCC) [Hill]. The Court in Hill itself relied on an earlier edition of Gatley (see supra note 26 at paras 592-93).

55. See supra note 37.

56. See Introduction, above.

57. Danah Boyd & Kate Crawford, “Critical Questions for Big Data” (2012) 15 Information, Communication & Society 662 at 663, DOI: http://doi.org/10.1080/1369118X.2012.678878.

58. See Crookes v Newton, 2011 SCC 47 at para 16 [Crookes].

59. See Emily B Laidlaw & Hilary Young, “Internet Intermediary Liability in Defamation” (2018) 56 Osgoode Hall LJ 112 at 115-16, DOI: https://doi.org/10.60082/2817- 5069.3389.

60. Crookes, supra note 58 at para 18.

61. Fairfax Media Publications Pty Ltd v Voller, [2021] HCA 27 at para 32, Kiefel CJ, Keane JJ & Gleeson JJ.

62. Crookes, supra note 58 at para 18.

63. Expressed under the definitions of “[b]roadcasting” and “publish” in Manitoba’s statute; “broadcasting” in Ontario’s statute; and “broadcasting” and “publication” in Newfoundland’s statute. See The Defamation Act, RSM 1987, c D20, s 1; Libel and Slander Act, RSO 1990, c L12, s 1(1); Defamation Act, RSNL 1990, c D-3, ss 2(a), 2(e).

64. See generally Bunt v Tilley, [2006] EWHC 407 (QB) [Bunt]; Metropolitan International Schools Ltd v Designtechnica Corp, [2009] EWHC 1765 (QB) [Designtechnica]; Tamiz v Google Inc, [2012] EWHC 449 (QB).

65. See Tamiz v Google Inc, [2013] EWCA Civ 68 at paras 25-26 [Tamiz CA].

66. Bunt, supra note 64 at paras 22-23; Tamiz CA, supra note 65 at para 26.

67. Bunt, supra note 64 at para 23 [emphasis in original]; Designtechnica, supra note 64 at para 36.

68. See Bunt, supra note 64 at paras 36-37. Although Bunt did rely on Emmens v Pottle and McCleod v St Aubyn, the former relates to a publication defence (what would now be the innocent dissemination defence) considered after acknowledging prima facie liability while the latter relates to contempt of court, not defamation. See Emmens v Pottle, (1885) 16 QBD 354 (CA); McCleod v St Aubyn, [1899] AC 549 at 357 (PC St Vincent). See also Bunt, supra note 64 at paras 22-23.

69. See Crookes, supra note 58 at para 21.

70. See Tamiz CA, supra note 65 at paras 25-26.

71. See Kent v Postmedia Network Inc, 2015 ABQB 461 at paras 70-72. But see Weaver v Corcoran, 2015 BCSC 165 at paras 179-82, rev’d on other grounds 2017 BCCA 160 at para 102 [Weaver CA].

72. See Menear v Miguna, 1996 CanLII 8214 (ONSC), rev’d 1997 CanLII 4432 (ONCA) (to allow jury to decide).

73. See e.g. Designtechnica, supra note 64 at paras 49-51; Niemela v Malamas, 2015 BCSC 1024 at para 107 [Niemela]; Duffy v Google Inc, [2015] SASC 170 at para 205.

74. See Murray v Wishart, [2014] NZCA 461 at para 144 [Murray].

75. See Tamiz CA, supra note 65 at paras 25-26.

76. See Shanthakumar v WST Media Group Inc et al, 2021 ONSC 2802 at paras 19-21 [Shanthakumar].

77. Defamation Act 1996 (UK), c 31, s 1(1).

78. Ibid, s 1(3).

79. See supra note 58 at paras 20-21.

80. See Laidlaw & Young, supra note 59 at 118 (stating “it is too soon to declare the [secondary publication doctrine] dead in Canadian law”).

81. See e.g. Baglow v Smith, 2015 ONSC 1175 at paras 175-96 [Baglow]; Pritchard v Van Nes, 2016 BCSC 686 at para 113.

82. See Thorpe v Boakye, 2022 ONSC 7176 at paras 56-62 [Thorpe]. Thorpe comes the closest but does not analyze whether the broadcasting provisions apply to internet intermediaries. A 1981 report of the Law Reform Commission of British Columbia considers that the phrasing in the Ontario statutes may capture cable television (their preoccupation) but might also be overbroad. See Report on Cable Television and Defamation (Law Reform Commission of British Columbia, 1981) at 9-10.

83. See Laidlaw & Young, supra note 59 at 118.

84. See Hill, supra note 54 at para 176.

85. Ibid; Crookes, supra note 58 at para 48, McLachlin CJ and Fish J; Tamiz CA, supra note 65 at para 34; Weaver CA, supra note 71 at para 86.

86. Grant, supra note 16 at para 76. See also Lewis v Daily Telegraph Ltd (1963), [1964] AC 234 at 260, Reid LJ, 283-84, Devlin LJ (HL (Eng)) [Lewis].

87. Ibid at para 114.

88. See Laidlaw & Young, supra note 59 at 118; Designtechnica, supra note 64 at para 54.

89. See Baglow, supra note 81 at paras 175-96.

90. See Davison v Habeeb, [2011] EWHC 3031 at para 41 (QB); Tamiz CA, supra note 65 at paras 27-35.

91. See Thorpe, supra note 82 at paras 158-60, 167.

92. See Trkulja v Google Inc LLC & Anor (No 5), [2012] VSC 533 at paras 16-18 [Trkulja VSC]; Google Inc v Trkulja, [2016] VSCA 333 at paras 348-49, 352 (albeit as a secondary publisher), 372 (in obiter dicta) [Trkulja VSCA], rev’d on other grounds (but approved as to this finding); Trkulja v Google LLC, [2018] HCA 25 at paras 38-40 [Trkulja HCA].

93. See Trkulja VSCA, supra note 92 at paras 348-49, 352, 372; Trkulja HCA, supra note 92 at paras 38-40.

94. Supra note 64 at paras 50-51.

95. Trkulja VSC, supra note 92 at para 27.

96. Tamiz CA, supra note 65 at paras 21-23; Niemela, supra note 73 at paras 103-107.

97. See Murray, supra note 74 at para 130 (noting that Facebook commenters could be directly sued). Compare Baglow, supra note 81 at para 196 (recognizing the impracticality of suing anonymous commenters).

98. See Introduction, above.

99. See Trkulja HCA, supra note 92 at paras 38-39.

100. See Designtechnica, supra note 64 at paras 51, 53.

101. Ibid at paras 50-51.

102. See e.g. Lewis, supra note 86 at 284, Lord Devlin; Gatley, supra note 26, s 12.017.

103. See Tamiz CA, supra note 65 at paras 27-35.

104. Weaver CA, supra note 71 at para 86.

105. I thank an anonymous reviewer for this scenario.

106. One might think the result would differ in British Columbia, where “defamatory words in a broadcast are deemed to be published.” Libel and Slander Act, RSBC 1996, c 263, s 2. In that context, “broadcast” refers only to analog broadcasts. Ibid, s 1.

107. Supra note 77, s 1(3)(d).

108. See Part I(A)(2), above.

109. See Automated and Electric Vehicles Act 2018 (UK), c 18. This scheme recalls s 4, describing the immunity of an insurer for accidents caused by software modifications that the insurer had forbidden.

110. Gaskin v Retail Credit Co et al, 1963 CanLII 245 at para 14, McGillivray JA (ONCA).

111. Bent v Platnick, 2020 SCC 23 at para 148 [Bent], cited in Peter A Downard, The Law of Libel in Canada, 4th ed (LexisNexis Canada, 2018) at para 14.2. For a civilian context, compare AB c Google, 2023 QCCS 1167 at paras 553-56 (discussing the upset one might feel from knowing defamatory text is out there, even without knowing that it has been read).

112. See Hill, supra note 54 at para 166. The SCC included in the idea of a person’s reputation the notion of how that person believes other people see them, which would suggest that one’s reputation exists in one’s own mind as well as that of others. In England, that belief would be compensable independently of reputation as a matter of “anxiety and annoyance.” Fielding v Variety Incorporated, [1967] 2 QB 841 at 855 (CA). See also McCarey v Associated Newspapers Ltd (No 2), [1965] 2 QB 86 at 107 (CA). The English approach appears more consistent with the ordinary meaning of the word “reputation.” This approach may be implicitly adopted in Bent. See supra note 111 at paras 148-49.

113. Some may disagree with this statement. For example, Slavny suggests there is a “non-instrumental” justification for defamation wherein reputation is harmed even when no negative consequences flow from it. On Slavny’s view, the “mere fact” of another’s attitude or belief changing, even when unaccompanied by conduct changing, intuitively counts as harm to reputation. See supra note 30 at 528, 530. I do not share that intuition, and so I discount the non-instrumental justification. In saying this, I do not mean to discount the impact one person’s beliefs about another’s thoughts may have on the first person. A person B who believes they have declined in the thoughts of A may suffer significant upset from this belief; but this suffering would exist independently of whether B’s belief is true. I distinguish “reputation” from “a person’s beliefs about their own reputation.” Moreover, even on Slavny’s view that the non-instrumental justification suffices to ground a defamation action, it is not necessary. Treating inclusion in an LM’s training set as publication could still be justified on instrumental grounds.

114. See Ley v Hamilton, [1935] 153 LT 384 at 386 (HL (Eng)).

115. Hill, supra note 54 at para 169.

116. Ley v Hamilton, supra note 114 at 386.

117. I infer here that text that refers to (true) cheats will also refer to those people in negative terms, and so the LM will “learn” to associate someone being called a “cheat” with other negative characteristics.

118. See e.g. Data Protection Act 2018 (UK), c 12, s 46; Personal Information Protection and Electronic Documents Act, SC 2000, c 5, Schedule 1, 4.9.5; Personal Information Protection Act, SA 2003, c P-6.5, s 25; Personal Information Protection Act, SBC 2003, c 63, s 24.

119. See Murray Shanahan, “Talking About Large Language Models” (16 February 2023), online (pdf): arxiv.org/pdf/2212.03551 [perma.cc/BJ26-G9ZT].

120. Such persons would by default be the LM developers but could conceivably include the curator of a database that was offered to an LM developer.

121. See Gatley, supra note 26, ss 11.001-11.003.

122. A defendant could, of course, use the other defences, but they would not raise novel issues. For example, a truth defence for LM-generated speech would be adjudicated the same way it is today: by summoning witnesses, taking affidavits, and using cross examination to establish what really was true at the time of the statement.

123. See supra note 19, s 3(8).

124. Ibid, ss 3(1)-3(4).

125. Ibid, ss 3(5)-3(6).

126. See UK, Ministry of Justice, Defamation Act 2013 (Explanatory Notes) at para 25, online: www.legislation.gov.uk/ukpga/2013/26/notes/division/5 [perma.cc/M48X-AZS3].

127. Ibid.

128. Spiller v Joseph, [2010] UKSC 53 at para 107 [Spiller] [emphasis in original].

129. Ibid at para 108.

130. Ibid at paras 110-12.

131. WIC Radio Ltd v Simpson, 2008 SCC 40 at paras 28, 63 [WIC Radio].

132. Ibid at para 61.

133. Ibid at para 46.

134. Ibid.

135. Ibid at paras 46-47, 50. In England, the courts resolve the public interest communication defence by treating honesty as related to the intended meaning, not the objective meaning ascertained by the courts. See Economou v de Freitas, [2018] EWCA Civ 2591 at para 95.

136. See WIC Radio, supra note 131 at paras 47, 50.

137. Copyright, Designs and Patents Act 1988 (UK), c 48, s 178.

138. Ibid, s 9(3).

139. Ibid, s 10.

140. CCH Canadian Ltd v Law Society of Upper Canada, 2004 SCC 13 at para 25.

141. See generally Carys Craig & Ian Kerr, “The Death of the AI Author” (2021) 52 Ottawa L Rev 31.

142. Ibid at 43 [emphasis in original], citing Pamela Samuelson, “Allocating Ownership Rights in Computer-Generated Works” (1986) 47 U Pitt L Rev 1185.

143. Théberge v Galerie d’Art du Petit Champlain inc, 2002 SCC 34 at para 30. See also Society of Composers, Authors and Music Publishers of Canada v Entertainment Software Association, 2022 SCC 30 at paras 5, 67.

144. See Spiller, supra note 128 at paras 66-67, 79.

145. See WIC Radio, supra note 131 at paras 3, 30, 50.

146. See Introduction, above.

147. WIC Radio, supra note 131 at para 53.

148. For example, an LM could be fed additional context saying, “also, X is a terrorist,” which may make responses on unrelated topics more likely to defame X.

149. See Shafi Goldwasser et al, “Planting Undetectable Backdoors in Machine Learning Models” (9 November 2024), online (pdf): arxiv.org/pdf/2204.06974 [perma.cc/C8WW-XD6Y]. Goldwasser et al suggest that some modifications may be undetectable by analyzing a Machine Learning System. The better approach may be to analyze the development process rather than the ultimate model.

150. See Defamation Act 2013, supra note 19, s 3(5).

151. Ibid, s 3(6).

152. Adam v Ward, [1917] AC 309 at 334 (HL (Eng)), Lord Atkinson.

153. See e.g. Leenen v Canadian Broadcasting Corp, 2000 CanLII 22380 at para 108 (ONSC); Puddister v Wells, 2004 NLSCTD 188 at para 25; Bird v Ontario, 2014 ONSC 2457 at para 24; Halls v Mitchell, 1928 CanLII 1 at 134 (SCC).

154. See e.g. Pullman v Walter Hill & Co Ltd, [1891] 1 QB 524 at 528 (CA), Esher MR [Pullman]; Adam v Ward, supra note 152 at 334, Lord Atkinson; Bent, supra note 111 at para 121.

155. London Association for Protection of Trade v Greenlands Ltd, [1916] 2 AC 15 at 35 (HL (Eng)), Lord Atkinson [Greenlands]; Waller v Loch (1881), 7 QBD 619 (CA). Note, however, that Brett LJ would have required that the defendant’s belief be reasonable. See Hanly v Pisces Productions Inc, 1980 CanLII 3047 at 373 (BCSC); MacDonald v Tamitik Status of Women Assn, 1998 CanLII 1227 at paras 115-17 (BCSC).

156. Supra note 26, s 15.039. Gatley provides no direct authority for this proposition, but it would seem to be an instance of the general rule that qualified privilege attaches to situations where a questioner appears to have a legitimate interest in the answer to the question. See e.g. Greenlands, supra note 155 at 461, Lord Atkinson and at 462-63, Lord Parker; compare the speech of Lord Buckmaster at 457.

157. Even if it does not augur well for Dave!

158. See e.g. Pullman, supra note 154 at 528, Esher MR; Cusson v Quan, 2007 ONCA 771 at para 39, rev’d on other grounds 2009 SCC 62.

159. Huntley v Ward (1859), 6 CB (NS) 514 at 517 (Ct Com Pl UK). See Grant, supra note 16 at para 30; Gatley, supra note 26, s 15.003.

160. Reynolds v Times Newspapers Ltd, [2001] 2 AC 127 at para 195 (HL (Eng)) [Reynolds]; Martin v Lavigne, 2011 BCCA 104 at para 36.

161. Horrocks v Lowe, [1975] AC 135 at 152 (HL (Eng)) [Horrocks].

162. Fisher v Rankin, 1972 CanLII 1535 at 713 (BCSC).

163. Compare Reynolds, supra note 160 at 202. Lord Nicholls considered that the common law standard for responsible journalism was only what the media itself espoused as appropriate (albeit without living up to that standard).

164. See e.g. OpenAI, “Best Practices for Deploying Language Models” (2 June 2022), online: openai.com/blog/best-practices-for-deploying-language-models [perma.cc/7AEM-FJ7X]. Since an earlier draft of this article, more work has been done in this space. See e.g. Markus Anderljung et al, “Frontier AI Regulation: Managing Emerging Risks to Public Safety” (7 November 2023) at 24-30, online (pdf): arxiv.org/pdf/2307.03718 [perma.cc/9RSS-SSRX]. This work identifies safety standards for frontier models (of which LMs are one type). See also Jonas Schuett et al, “Towards Best Practices in AGI Safety and Governance: A Survey of Expert Opinion” (11 May 2023), online (pdf): arxiv.org/pdf/2305.07153 [perma.cc/CA5D-2PVP]. The authors present empirical evidence of industry acceptance of certain safety measures. Among these safety measures are what are known as “capabilities evaluations,” described in greater detail by Mary Phuong et al. See “Evaluating Frontier Models for Dangerous Capabilities” (5 April 2024), online (pdf): arxiv.org/pdf/2403.13793 [perma.cc/N3FL-TETE]. See also Toby Shevlane et al, “Model Evaluation for Extreme Risks” (22 September 2023), online (pdf): arxiv.org/pdf/2305.15324 [perma.cc/YSV6-A2AV]. There is also a focus on setting out risk-aware policies for how to decide when to release new models. See e.g. UK, Department for Science, Innovation & Technology, Emerging Processes for Frontier AI Safety (Policy Paper) (DSIT, 2023), online: www.gov.uk/government/publications/emerging-processes-for-frontier-ai-safety/emerging-processes-for-frontier-ai-safety [perma.cc/A24Y-ZXVF]. Caution, however, is warranted. The burgeoning literature in this space may not address courts’ concerns. Courts handling defamation are concerned about defamation, not all the other risks that could come from models. These practices are primarily concerned with other risks, and the risk of defamation may be engaged with only incidentally.

165. Horrocks, supra note 161 at 149.

166. Botiuk v Toronto Free Press Publications Ltd, 1995 CanLII 60 at para 79 (SCC) [Botiuk]. At least in this regard, malice in the honest opinion setting differs from malice in the qualified privilege setting. Contra Creative Salmon Company Ltd v Staniford, 2009 BCCA 61 at para 32. See also Spiller, supra note 128 at paras 67-68.

167. See Horrocks, supra note 161 at 150.

168. Supra note 166 at para 79; Horrocks, supra note 161 at 150.

169. Horrocks, supra note 161 at 149-50.

170. See supra note 111 at para 136.

171. Ibid.

172. Supra note 161 at 150. See also Roberts v Bass, [2002] HCA 57 at paras 78-80, Gaudron, McHugh, and Gummow JJ; but see para 185, Kirby J.

173. Crookes, supra note 58 at para 20, citing Vizetelly v Mudie’s Select Library Ltd, [1900] 2 QB 170 (CA) [Viztelly]. See also Emmens v Pottle, supra note 68 at 357.

174. Supra note 77, s 1(1).

175. See CM Callow Inc v Zollinger, 2020 SCC 45 at paras 56-63, 67-73. The SCC describes the “opportunity for dialogue” between private common law and Quebec civil law as “particularly useful” and “arguably a special mandate for [the SCC]” (ibid at para 60). See also Wastech Services Ltd v Greater Vancouver Sewerage and Drainage District, 2021 SCC 7 at paras 108-10; Annapolis Group Inc v Halifax Regional Municipality, 2022 SCC 36 at paras 46-50.

176. See AB c Google, supra note 111 at paras 251-59, 279, 308-311, 521-24, 553-56 (awarding 500,000 CAD against Google for failing to remove a website from search results). See also Karen Eltis, “Is ‘Truthtelling’ Decontextualized Online Still Reasonable? Restoring Context to Defamation Analysis in the Digital Age” (2018) 63 McGill LJ 553 at 578-81, DOI: https://doi.org/10.7202/1066336ar (arguing for importing civilian ideas into Canadian common law defamation).

177. Vizetelly, supra note 173 at 180, Romer LJ.

178. See Designtechnica, supra note 64 at para 70.

179. Supra note 26, s 7.038.

180. See e.g. Lambert v Roberts Drug Stores, 1933 CanLII 289 at 322 (MBCA). Although Vizetelly appears to be a second case, here the defendants did not have actual notice of the warning because they failed to read a publication that gave the warning. See supra note 173 at 171.

181. See Batten v Pall Mall Deposit and Forwarding Company Ltd (1927), The Times (24 June 1927) 5 (KB) [Batten].

182. Weldon v “The Times Book” Company Ltd (1911), 28 TLR 143 at 143 (CA). See also Ross v Eason & Son Ltd, [1911] 2 IR 459 at 465 (KBD), Holmes LJ. See also Goldsmith v Sperrings Ltd, [1977] 1 WLR 478 at 488 (CA), Denning MR, dissenting (though on a point not properly before the court).

183. See Bottomley v FW Woolworth and Co (1932), 48 TLR 521 (CA).

184. See e.g. Morris v West Hartlepool Steam Navigation Co Ltd, [1956] AC 552 at 574 (HL (Eng)).

185. Ouyang et al, supra note 6 at 9.

186. See e.g. Arram Sabeti, “Teaching GPT-3 to Identify Nonsense” (25 July 2020), online: Arram Sabeti arr.am/2020/07/25/gpt-3-uncertainty-prompts [perma.cc/GP2F-SSXH]. Sabeti uses context to get GPT-3 to respond with “yo be real” when asked a nonsense question.

187. See Stephanie Lin, Jacob Hilton & Owain Evans, “TruthfulQA: Measuring How Models Mimic Human Falsehoods” (8 May 2022), online (pdf): arxiv.org/pdf/2109.07958 [perma.cc/KLU6-RGE5].

188. Glomarize: to neither confirm nor deny the truth of a matter. See US Department of Justice, “FOIA Update: OIP Guidance: Privacy ‘Glomarization’” (1 January 1986), online (blog): www.justice.gov/archives/oip/blog/foia-update-oip-guidance-privacy-glomarization [perma.cc/B6ZJ-2RY3].

189. State-of-the-art LMs appear to evaluate such matters with some accuracy. See Saurav Kadavath et al, “Language Models (Mostly) Know What They Know” (21 November 2022), online (pdf): arxiv.org/pdf/2207.05221 [perma.cc/VF8Q-M8AQ].

190. This is one component of the mechanism for producing better LMs. See Bai et al, supra note 9.

191. Technically, damage is an element of libel but is presumed and not subject to rebuttal. See Jameel v Dow Jones & Co Inc, [2005] EWCA Civ 75 at paras 32, 41.

192. [1933] All ER Rep 432 at 439 (CA). Note that I am observing the absence of an argument (for considering the scale of harm when assessing care) and inferring from that absence that the argument held no sway. Alternate readings of this case are possible.

193. See e.g. Barron MP & Anor v Vines (Rev 1), [2016] EWHC 1226 at para 21(3) (QB).

194. See e.g. Wright v Granath, [2022] EWHC 1181 at para 81 (QB).

195. [2011] EWHC 1884 at paras 66, 70 (QB) [Thornton]. See also ibid at para 116 (suggesting that the standards of responsible journalism require checking any defamatory fact).

196. See supra note 111 at para 136.

197. See Part I(A)(2), above.

198. I use the game theoretic sense of “common knowledge.” See Peter Vanderschraaf & Giacomo Sillari, “Common Knowledge” in Edward N Zalta & Uri Nodelman, eds, The Stanford Encyclopedia of Philosophy (Stanford University, 2021), online: plato.stanford.edu/entries/common-knowledge/index.html [perma.cc/56ZJ-PRP6].

199. Defamation Act 2013, supra note 19, s 1(1).

200. Whiten v Pilot Insurance Co, 2002 SCC 18 at paras 120, 125. See also A Mitchell Polinsky & Steven Shavell, “Punitive Damages: An Economic Analysis” (1998) 111 Harv L Rev 869 at 874, 921-22, DOI: https://doi.org/10.2307/1342009. England does not recognize this deterrence objective and restricts punitive damages to oppressive, arbitrary, or unconstitutional government action and to profit making from the tortious conduct. See e.g. Rookes v Barnard, [1964] AC 1129 at 1226 (HL (Eng)). Punitive damages are rarely awarded in modern-day English defamation cases. See James Goudkamp & Eleni Katsampouka, “An Empirical Study of Punitive Damages” (2018) 38 Oxford J Leg Stud 90 at 111, DOI: https://doi.org/10.1093/ojls/gqx013. By contrast, 27 per cent of defamation awards reported by Canadian courts from 2003-2013 featured punitive damages. See Hilary Young, “The Canadian Defamation Action: An Empirical Study” (2017) 95 Can Bar Rev 591 at 619, DOI: https://doi.org/10.2139/ssrn.2802787.

201. Batten, supra note 181.

202. Defamation Act 2013, supra note 19, s 4(1)(b).

203. See Serafin v Malkiewicz & Ors, [2020] UKSC 23 at paras 60-62.

204. See e.g. Lachaux v Independent Print Ltd & Anor, [2021] EWHC 1797 at paras 146-47 (QB); Hijazi v Yaxley-Lennon (Rev 1), [2020] EWHC 3058 at para 45 (QB).

205. See Sivier v Riley, [2021] EWCA Civ 713 at para 30.

206. Grant, supra note 16 at para 126.

207. See Part I(B)(2), above.

208. Grant, supra note 16 at para 126.

209. See e.g. Wallace v United Grain Growers Ltd, 1997 CanLII 332 at paras 76-77 (SCC); Johnson v Unisys Limited, [2001] UKHL 13 at para 47; Nevsun Resources Ltd v Araya, 2020 SCC 5 at para 237, Brown & Rowe JJ, dissenting.

210. See Emily B Laidlaw, “Re-Imagining Resolution of Online Defamation Disputes” (2019) 56 Osgoode Hall LJ 162, DOI: https://doi.org/10.60082/2817-5069.3390. Compare Jack Clark & Gillian K Hadfield, “Regulatory Markets for AI Safety” (11 December 2019), online (pdf): arxiv.org/pdf/2001.00078 [perma.cc/YW5H-LLD3].

211. See Owain Evans et al, “Truthful AI: Developing and Governing AI That Does Not Lie” (13 October 2021), online (pdf): arxiv.org/pdf/2110.06674 [perma.cc/98J6-TFBF], especially at 51ff.