Transcript [00:03] Hello everyone and welcome to Fedsock [00:06] Forums. Today, September 22nd to 2026, [00:09] we are delighted to host a conversation [00:12] titled large liel models. When do AI [00:15] hallucinations become defamation? My [00:18] name is Lily Fremier and I am a [00:19] secondyear law student at Washington and [00:21] Lee University. I also serve as a [00:24] student liaison to the Federalist [00:25] Society's communications and technology [00:28] practice group. As always, please note [00:31] that all expressions of opinion are [00:33] those of the experts on today's call as [00:36] the Federal Society takes no position on [00:38] a particular legal or public policy [00:40] issues. Our moderator, Schlommo Clapper, [00:44] is the founder and CEO of Learned Hand, [00:46] the first and only generative AI [00:49] platform purpose-built for the [00:51] judiciary. Learned Hands technology [00:54] supports judges and court staff across [00:56] the full arc of a case from filing [00:58] through drafting and is currently [01:00] deployed with the Michigan Supreme [01:02] Court, the Superior Court of Los Angeles [01:04] County, and courts in 10 states serving [01:07] over 20 million Americans. [01:09] Before founding Learned Hand, Schlommo [01:11] clerked for the honorable Steven J. [01:13] Manashi on the US Court of Appeals for [01:16] the Second Circuit, litigated complex [01:18] commercial disputes at Quinn Emanuel [01:21] Urkart and Sullivan, and worked as a [01:23] deployment strategist at Palunteer [01:25] Technologies in both the commercial and [01:28] government sectors. Earlier in his [01:30] career, he was a behavioral researcher [01:32] in Dan Arieli's lab at Duke University. [01:36] Schlommo received his JD from Yale Law [01:38] School where he won the Thomas I. [01:41] Emerson Prize for his writing and holds [01:43] a BS in economics from the Wharton [01:46] School and a BA in European history from [01:48] the University of Pennsylvania, both [01:50] Sumu Cumlude. [01:52] One note for our audience before we [01:54] begin. If you have any questions [01:56] throughout the webinar that pertain to [01:58] the topic at hand, please submit them [02:00] through the Zoom's Q&A feature located [02:03] at the bottom of your screen, not the [02:05] chat, so that we'll be able to access [02:07] them later in the program. With that [02:09] note, I will turn things over to our [02:11] speakers and to start the conversation. [02:13] Schlommo, thank you so much for joining [02:15] us and for moderating today's [02:16] discussion. [02:18] >> Thank you, Lily. Um hopefully I will not [02:20] be guilty of any liel intentional [02:23] otherwise throughout the conversation. [02:25] So to to to introduce our speakers I [02:28] actually asked Gemini to um write uh a [02:33] brief intro of Professor Litzky and [02:35] Professor Volik for the large liel [02:38] models panel. So let's uh professors you [02:41] can let us let me know if if we're going [02:43] to add a tenth lawsuit [02:45] uh to the queue. All right. So um [02:47] professor Liss Halitzky is the Raymond [02:50] and Miriam Erlick eminent scholar chair [02:52] in US Constitutional Law at the [02:54] University of Florida Levven College of [02:55] Law and former dean of the University of [02:58] Missouri School of Law. A leading [02:59] national authority on the intersection [03:01] of tort law and the first amendment. Her [03:02] scholarship focuses on media law, free [03:04] expression and digital spaces and [03:07] internet anonymity. She serves as a [03:09] co-reporter for the ALI ongoing [03:12] restatement of the law third tors [03:13] defamation of privacy positioning her at [03:15] the forefront of efforts to adapt [03:16] traditional liable principles to modern [03:19] technological disruptions. Uh professor [03:21] is that substantively correct? [03:24] >> It is [03:26] >> great. All right. Professor Volik is the [03:29] uh professor Volik is the Thomas M. Cyel [03:32] Senior Fellow at the Hooer Institution [03:34] at Stanford University in the Gary T. [03:35] Schwarz distinguished professor of law [03:37] ammeritus at UCLA school of law. One of [03:40] the mostly most widely cited first [03:41] amendment scholars in the country. He [03:43] pioneered the scholarship on AI [03:44] generated defamation with the seminal [03:46] paper large liable models liability for [03:48] AI output. He is the founder of the [03:50] legal blog blog the Volic Conspiracy, [03:52] founder and editor and chief of the [03:54] journal of free speech law and the [03:55] author of leading case books on first [03:57] amendment law. Um, Professor Volik, is [04:00] that also [04:01] >> so far so good? Perfectly accurate. [04:02] >> So far so good. [04:04] >> Okay. Fantastic. So to to start us off [04:08] uh professor VA, can I turn to you? [04:09] Could you could you share uh with the [04:12] audience uh the facts of Keen and sort [04:14] of situate this into the uh in in into [04:18] the other in the context of the other [04:20] lawsuits that have uh uh surfaced. [04:23] >> Uh yeah. So uh um the uh case we're [04:27] talking about right now is just from [04:29] about a week or so ago Keem v Google [04:31] although it is one of nine cases that [04:33] have been filed over uh over alleged [04:36] liable by AI in US courts uh over the [04:40] last three years. Um, so, uh, the [04:42] plaintiff apparently goes by Jimmy Jimmy [04:45] Keane. Um, [04:48] is in 1997 he had pleaded guilty to a [04:52] drug distribution conspiracy, was [04:54] sentenced to 10 years in prison. Uh, but [04:56] then, uh, agreed to work with the FBI as [04:59] a prison informant to help gather [05:00] evidence against the suspected serial [05:02] killer who was who's also in prison. Uh, [05:06] and, uh, he did that. uh and as as a [05:09] result he got a pardon. Um and his [05:12] conviction was expuned. Uh uh some [05:16] people whose convictions are expuned [05:18] work really hard to kind of keep them [05:19] keep them quiet. In his case, he [05:21] published a memoir recounting this [05:23] experience was sub subsequently adapted [05:25] in an Apple TV miniseries titled [05:27] Blackbird and as best I can tell has led [05:29] a blameless life since. Um well uh maybe [05:34] this was a blameless life, but he has [05:36] been blamed. uh by uh he says Google AI [05:40] overviews [05:42] uh which uh uh at various times in the [05:46] middle of 2025 [05:48] uh uh said various things in response to [05:50] queries such as Jimmy Keane net worth um [05:54] and the chief thing that they said [05:56] according to the complaint there's [05:57] another twist to it which we'll get to a [05:58] little later on is that he was serving a [06:01] life sentence including uh ser life [06:04] sentence without parole including for [06:06] the murders of three women that is [06:10] apparently totally hallucinated. [06:14] Note what it's not the it's not that [06:16] there is some error online [06:19] that was then um echoed by Google AI [06:24] overview the way if you do a Google [06:25] search and there's a incorrect [06:27] information on some page Google search [06:29] might give you a snippet of that. Um so [06:32] some people say well the problem with [06:34] these uh uh with these um uh uh AI u uh [06:39] large language models is garbage in [06:41] garbage out. No this is a different kind [06:43] of geigo. This is gold in garbage out. [06:46] As best we can tell all of the [06:47] information about him was quite [06:49] accurate. Despite that, the allegation [06:51] is and of course at this point it's just [06:53] allegations uh that uh uh um uh the [06:58] allegation is that uh Google uh Google's [07:01] AI output made these assertions. And by [07:03] the way, note it's just as a result of a [07:05] normal Google search. It's not somebody [07:07] specifically asking for AI. It's just [07:09] the AI overview AI output. It's just the [07:11] AI overview that is normally presented. [07:14] Um uh so he sued uh Google moved to [07:18] dismiss and the court uh uh allowed the [07:22] case to go forward. Um so uh let me [07:25] quickly go through the key points in uh [07:28] in uh Google's uh argument then mention [07:31] one point that really wasn't discussed [07:32] and I think for for for good reason. Um [07:35] uh so first Google argued well there's a [07:38] disclaimer that says well AI overviews [07:41] are only starting points and that uh it [07:44] warns that it may include mistakes and [07:46] the court says well it's not clear [07:48] whether a disclaimer was actually [07:50] present. It's a motion to dismiss. We're [07:53] going to assume the facts alleged in the [07:55] complaint and here there was no such [07:57] allegation. Um this having been said, I [07:59] should add um that uh disclaimers by [08:02] themselves are generally not enough in [08:05] liable uh litigation. Um uh now if a [08:09] disclaimer is this is all fiction, this [08:11] is all a joke, this is parody, well that [08:14] might be enough because that says [08:15] essentially this is not true. But a [08:17] disclaimer that says this might not be [08:18] true is generally not enough. Classic [08:20] example, we don't think of it a [08:22] disclaimer, but it is the phrase rumor [08:24] has it. Let's say I say rumor has it [08:26] that Schlommo Clapper Well, I'm not even [08:29] gonna bring this up. Who knows? [08:30] [laughter] I will trade trade on it. [08:33] Rumor [08:33] >> I have a I have I have dozens of [08:35] witnesses here. [08:37] >> Right. Exactly. [08:39] >> Exactly. Um and uh if I say that and it [08:43] turns out especially turns out there is [08:45] no such rumor but even if there is the [08:47] rumor is false I could still be liable. [08:50] Um uh now it may decrease the damage in [08:53] some respect. In certain situations it [08:55] may be relevant. There is a special [08:57] rumor privilege but that's for very [08:59] specific kinds of uh uh repetitions of [09:02] rumors. Uh but basically the law says [09:04] look uh uh disclaimer uh uh doesn't [09:09] doesn't eliminate the uh um the uh the [09:13] the damage to reputation. again may may [09:15] diminish it in some measure but doesn't [09:17] eliminate it because while people know [09:18] that some rumors are false they also [09:21] know some rumors are true likewise with [09:23] Google uh uh with any disclaimer that [09:25] Google might have provided I think that [09:27] by itself would not be enough a second [09:29] point uh is that Google said well users [09:33] could have easily verified that the [09:34] overviews contain false information [09:36] there was a link uh to a Wikipedia [09:38] article people clicked on it they would [09:39] see that that this is that this wasn't [09:41] so uh but but that's not enough uh to [09:45] eliminate defamation liability. Simply [09:47] saying, well, for more read this, and [09:49] then for the very few people who do read [09:51] this, read the follow the link, they'll [09:54] realize that it's false, but lots of [09:55] people won't follow. Um there was also [09:58] another case. Remember I said this is [10:00] one of nine cases. Um uh of them, some [10:03] are pending, two have settled, one of [10:05] them following arbitration. Um uh uh [10:09] another case in Delaware uh trial court [10:13] has also allowed the case to go forward [10:16] also denied a motion to dismiss but [10:17] there was a Georgia case where initially [10:19] a motion to dismiss was denied but then [10:21] summary judgment was granted called [10:23] Walters v open it was the first such [10:25] large liel models case um and uh uh the [10:28] court uh in that case uh uh stressed [10:32] that uh there wasn't just a disclaimer [10:34] on the part of open AI it's that that [10:37] the that Chad GPT actually [10:40] said things like uh uh like uh um that [10:44] uh I I can't provide the desired [10:46] information. Uh this is not accessible [10:49] to me. Eventually it did do that but uh [10:52] there were lots of very specific signals [10:55] to the person doing the searching who as [10:57] it happens actually [10:59] I think personally knew the target the [11:01] the kind of the target of the liel and [11:03] therefore didn't believe uh the the [11:05] allegations. Um so there there were lots [11:08] of uh uh uh signals that were present [11:11] that made it was not just a disclaimer [11:13] just quite clear uh in context that this [11:16] was an error. And the court says, well, [11:18] for that reason, Walters is not really [11:21] on point here because at least the [11:22] allegations in the complaint are that [11:24] there were no such such uh such powerful [11:26] signals there. Um uh now uh um [11:32] as we know uh the the question of [11:36] whether something is is false and [11:38] whether something uh uh something is [11:41] damaging to reputation is only the first [11:43] step in liel cases. If the plaintiff is [11:46] a public figure, then famously New York [11:49] Times E. Sullivan and its progeny say [11:51] that there can only be liability based [11:53] on a showing of actual malice. Actual [11:56] malice of course is not actually malice. [11:58] It means knowing or reckless falsehood. [12:01] Well, um what does uh uh what does that [12:05] mean in the AI context? Uh where in a [12:08] sense, what does it mean for an AI to [12:10] know something much less be reckless [12:12] about it? Uh and by the way uh the the [12:14] court uh the court treated uh uh for for [12:18] purposes of this motion um uh treated [12:20] Keen as a limited purpose public figure [12:23] uh because of commentary on his own life [12:25] story which he has of course uh uh [12:28] broadcast in the actual life story is [12:31] broadcast in his memoirs. So um the [12:33] courts said and I think this is this is [12:36] correct as a matter of liel law although [12:38] query whether it makes policy sense. The [12:41] court said the real focus is on uh the [12:44] mental state of the company not the [12:47] program which may not have a mental [12:48] state may not have a men's in the sense [12:50] of mind men's rhea guilty mind it [12:54] doesn't have a mind but the company [12:57] employs people who do and uh keen [13:00] contends that he directly notified [13:02] Google on three occasions it was [13:03] publishing false information about him [13:05] and despite this notice Google continued [13:07] to publish similar defamatory [13:09] statements. Uh so that the court said is [13:12] sufficient allegation of actual malice [13:14] that it that it was aware the company [13:17] was aware and didn't take steps to to [13:19] block that. Um uh [13:22] uh one more uh thing remember I [13:24] mentioned that the com that the claims [13:27] about him put out by uh AI overview were [13:30] mostly uh about allegations again false [13:34] allegations that uh that he was serving [13:36] a life sentence. Well, he was also um uh [13:39] um uh one of the allegations also he was [13:42] convicted of drug trafficking and he [13:44] says well that too is lielist because in [13:46] fact he was convicted of a felony [13:48] conspiracy to distribute cocaine. [13:50] That's not exactly drug trafficking. [13:52] Court said close enough. Close enough. [13:55] That's a legal term. Uh no legal term is [13:58] substantial truth. Uh but uh um uh that [14:01] is in fact a defense in liel cases or [14:03] rather let's say be more precise an [14:06] argument for non-lability in liel cases. [14:08] Um that uh uh that uh the uh if the [14:13] statement is substantially true if it [14:15] has the same gist or sting as the truth [14:18] doesn't matter if it may be technically [14:20] legally not completely accurate. Uh so [14:23] that's just a reminder that if there is [14:25] possible uh even if there is possible [14:27] liability, there's still all the normal [14:29] liable defenses available to the uh to [14:32] the company such as substantial truth. [14:34] Uh so that's an important point. Let me [14:36] let me mention something about something [14:38] the court didn't talk about and that's [14:40] 47 USC section 230. Uh and I think it [14:43] was correct uh uh uh not to talk about [14:46] it. Um uh so uh as we know Google has [14:50] the protection of section 230. Google in [14:53] its main business of search. Uh so [14:55] section 230 says that an online service [14:58] provider an online content provider is [15:00] not going to be liable for information [15:01] provided by others. That's why Google [15:04] can provide links to various pages and [15:06] not be responsible for any falsehoods at [15:10] those links in those pages or any [15:12] falsehoods in the little snippets that [15:14] it outputs from those pages. the [15:16] traditional Google search results. Uh [15:19] but the whole point of generative AI is [15:22] that it's generative is that it [15:23] generates things. Here it wasn't that [15:26] Google was uh being held responsible for [15:29] information provided by another content [15:31] provider. The lawsuit seeks to hold it [15:34] responsible for information provided by [15:35] itself by its own software. Uh so there [15:38] section 230 is not going to is not going [15:41] to be a protection. I think there's a [15:43] pretty broad consensus among uh law [15:45] professors and probably most kind of [15:47] impartial lawyers who consider the issue [15:49] that that is indeed so. So that's the [15:52] that's the picture of uh big picture of [15:54] Keen. I think on balance this is again [15:57] under current law a correct uh decision. [16:02] Um and uh it'll be interesting to see [16:04] how things proceed after discovery of [16:06] summary judgment. Be interesting to see [16:08] maybe Google will settle. Who knows? Uh [16:10] but but that's the big summary. I'd love [16:13] to hear what Lissa um has to say. I just [16:15] should uh uh should stress I'm not sure [16:17] this was in the biography, but in some [16:19] respects some the most important uh [16:22] thing about Lisa. There are many [16:23] important things about Lissa, but for [16:24] our purposes, she is the co-porter [16:28] for the American Law Institutees [16:29] restatement of tort's third defamation [16:32] and privacy which she's currently [16:35] working on with together with Robert [16:36] Post. Uh and so she is uh uh um one of [16:43] the leading uh academics in the country [16:46] dealing with liable law. Not just the [16:47] way we academics usually deal with it by [16:49] writing articles, but she's actually uh [16:52] co-drafting this uh uh this restatement [16:55] which as of course as lawyer uh [16:58] listeners know the restatements have [16:59] been tremendously influential and [17:01] continue to be tremendously influential. [17:02] I'm not sure the restatement will have [17:04] anything to say on AI as such. uh but [17:07] she is she is certainly the expert on [17:09] this subject. [17:13] Lissa, take it away. Okay. So, um, one [17:16] of the interesting things here about the [17:18] Keen case that distinguishes it from the [17:20] earlier case, Walters versus OpenAI, [17:22] which I had the good fortune to write [17:24] about in, uh, one of my articles called [17:27] inevitable errors, is that this case is [17:32] it does illustrate that there are going [17:35] to be inevitable errors in the form of [17:37] hallucinations. And we know from the [17:40] computer science literature that given [17:43] how AI is constructed, we can't [17:46] completely eliminate those. But both [17:49] this case and the Starbucks case that [17:51] we're going to talk about in just a [17:52] minute involve what happens after that [17:56] because both of the plaintiffs tried to [17:59] notify Google numerous times. uh Jimmy [18:04] Keane three times, Starbucks numerous [18:06] times, and then after notice, Google did [18:11] not take sufficient steps, or at least [18:13] so the allegation goes, to stop the [18:17] defamatory hallucinations. [18:20] And these are very serious allegations [18:22] in in both cases. I mean, the allegation [18:24] that you were a murderer, it's easy to [18:25] see how the hallucination happened. This [18:28] was a guy who uh was a confidential [18:30] informant about a serial killer. And it [18:33] conflated information about the ser [18:35] serial killer with information about [18:37] Jimmy Keane. And given if you know [18:39] anything about how AI works is probable [18:42] tokens or words that go together. Um it [18:46] it just conflated those two topics that [18:48] were often combined together in sources. [18:51] Uh but the problem here in these cases [18:55] and what helps make out a case that the [18:58] companies did have a a men's rhea that [19:01] makes them culpable is that they got the [19:04] notice and did nothing. And so even [19:06] though I don't want to generalize too [19:08] much from these cases because they're [19:10] just on motions to dismiss, they're at [19:12] very preliminary stages and the standard [19:15] for getting past a motion to dismiss is [19:18] really pretty low. uh especially in [19:21] state court in Delaware where one of [19:22] these cases comes, [19:24] >> it just has to be reasonably conceivable [19:26] allegations to get past the motion to [19:28] dismiss. But what they're telling us [19:31] about the company, the AI companies [19:34] generally is you better have a mechanism [19:37] for people to give notice that you're [19:40] getting outputs that are defamatory [19:42] falsehood. And when they give that [19:44] notice, you better take some kind of [19:46] action to limit the subsequent um [19:51] production of defamatory falsehoods. [19:59] Wow. So many so many rich areas for [20:02] follow-ups. Um [20:05] so um you know, Professor Lisky, I'm I'm [20:07] glad that you mentioned the Starbucks [20:08] case. Do you uh want to share a little [20:11] bit about that case and how it's [20:13] different uh uh from uh from uh from [20:17] Keen? [20:18] >> Okay. So, the the um the Keen case is in [20:22] federal district court in Illinois. The [20:24] Starbucks case is in state court in [20:26] Delaware. And if you may remember, um, [20:28] some of the important defamation [20:30] litigation with the, uh, US Dominion, [20:33] the voting machine company, uh, took [20:35] place in Delaware and resulted in one of [20:37] the largest defamation settlements, uh, [20:39] in, in fact, I think it is the largest [20:41] defamation settlement in history. Um, [20:44] so, uh, in Delaware state court, to get [20:48] past a motion to dismiss, you only need [20:51] to make reasonably conceivable [20:53] allegations because it's an you have to [20:55] put the defendant on notice of the [20:58] claims you're going to be making. And [20:59] so, you have to keep that in mind is the [21:01] judge at every turn said, "I'm going to [21:04] give Starbucks the benefit of the doubt [21:06] on all of these." Whether he can prove [21:08] these facts all in discovery is an open [21:11] question. [21:12] So Robbie Starbucks uh characterizes [21:15] himself as an activist and one of the [21:18] issues that he has been an activist on [21:21] is anti-Dei uh initiatives. So he has [21:26] spoken out firmly about about diversity [21:28] and inclusion initiatives and he was a [21:31] music producer. He's fairly well known. [21:33] He's out there on the on the internet. [21:35] So he is a fairly public person as well. [21:40] But what he is arguing is that the AI [21:44] told him that it was engineered to [21:48] produce lies about him deliberately. [21:51] And so one of his arguments in his [21:53] lawsuit is based on what the AI itself [21:57] told him that it it was basically uh out [22:01] on a campaign of character [22:02] assassination. And if you read what he [22:05] says it said about him, you can kind of [22:07] see why he might believe that because [22:10] the allegations were in fact quite [22:12] serious. Um, it said he was a a risk to [22:16] children. He had a confirmed criminal [22:18] record. He had sexually harassed [22:21] multiple women. His nanny had accused [22:23] him of abuse and neglect. Um, it created [22:26] a ficti fictitious ex-wife, [22:30] uh, which is a special detail here. And [22:34] um it even said he had been in the adult [22:37] film industry and had uh used his power [22:41] in the adult filming industry to [22:43] influence people to do things they [22:44] weren't comfortable with. So the the [22:48] number and variety of defamatory [22:50] allegations it made about Starbucks were [22:53] were many. By the way, his name does not [22:56] have an S on the end. It's Starbucks, [22:57] not Starbucks. Uh so so he says he [23:02] started by trying to get Google's [23:03] attention by posting on X and linking to [23:07] Google executives and they did in fact [23:10] he says reach out to him based on that [23:12] and then he he gets in contact uh with [23:14] somebody within Google. [23:17] He says that she tried to remedy they [23:19] she apologized tried to remedy his [23:21] problem but when she couldn't she ended [23:23] up resigning. Um but the the point here [23:27] is that there were multiple defamatory [23:30] allegations there. There were multiple [23:34] attempts to give notice to Google. [23:37] Google evidently seemed to [23:41] promise that they'd do something but in [23:43] fact never did. And he says that as a [23:46] result of this he actually had a death [23:48] threat that somebody got arrested for. [23:51] He had people come up to him and ask him [23:54] about the allegations. His children [23:56] received the allegations. [23:58] And so, uh, Google Google came into the [24:02] complaint and said it needed to be [24:03] dismissed for three reasons. It said [24:06] number one is that there's no actual [24:10] malice here. Or it didn't quite do them [24:12] in this order, but anyway, the the most [24:14] interesting one is it says there's no [24:16] actual malice here. And the judge [24:18] said,"Well, it seems to us that if you [24:23] did nothing, [24:24] then that could be reckless disregard [24:27] for the falsity of your AI's output. [24:30] Uh, and if you did something and no, I'm [24:34] sorry. If you did absolutely nothing, [24:36] then it's intentional that you you maybe [24:39] were intentionally defaming him." And [24:40] for purposes of the complaint, it [24:42] credited that the AI said, "We're doing [24:45] this deliberately, but it said if you [24:48] tried to fix it and it didn't work," [24:50] then that was reckless disregard for the [24:53] falsity of the defamatory content. Now, [24:57] I have to tell you, I think the judge is [24:59] stretching a little bit the concept of [25:01] reckless disregard. Some might say that [25:03] maybe that's just gross negligence. [25:05] Maybe it's not like I'm putting my head [25:07] in the sand. Um but but it's a fine line [25:11] you're walking there. And so the judge's [25:13] point is a really interesting point [25:14] about if you did something ineply, it [25:17] could be actual malice and if you did [25:19] nothing, it's actual malice. Um the [25:22] Google tried to contest publication [25:24] because there were some gaps in the [25:26] complaint. Uh Starbucks didn't exactly [25:29] say to whom these uh defamatory [25:33] statements had been released. And it [25:35] does make you wonder if in Discovery it [25:38] comes out that there was a pattern of [25:41] searches by Starbucks or his lawyer or [25:44] somebody associated with Starbucks to [25:46] try to produce as many defamatory [25:48] statements as possible because that it [25:50] wasn't clear in the complaint to whom it [25:52] was released. Um, one of the interesting [25:55] things though is Google said, "Well, [25:56] then it wasn't published because if it [25:58] just went to you, it has to go to a [26:01] third party in order to be published for [26:03] purposes of defamation." And here's [26:05] where he got the um the Gemini ad well [26:10] first it was barred and then Gemini but [26:11] the Google admissions uh came into play. [26:14] Uh Google told him it had released [26:17] information about him to 2.8 million [26:20] people [26:23] and and again so so when Google is is [26:26] fighting the publication element they [26:28] say well you can't believe Google. [26:31] [laughter] [26:33] So the judge says, "Well, for purposes [26:36] of a motion to dismiss, we are going to [26:38] believe Google." And it's not [26:41] necessarily true that 2.8 [26:43] million people received defamatory [26:46] information about Mr. Starbucks, but at [26:48] least one third party did based on that. [26:51] And so we're going to let him go on uh [26:53] sufficient evidence of publication to [26:55] keep going. And then finally, there was [26:58] an argument about damages. And this is [27:00] really a topic we've been working on in [27:02] the restatement. We're deep in the weeds [27:04] on uh the damages portions of the [27:06] restatement these days. And so this was [27:09] of particular interest to me. But there [27:12] was a conflict [27:14] uh between Tennessee law uh and Delaware [27:18] law as to how much information you had [27:21] to prove to prove that you were injured [27:23] by the defamatory statement. And under [27:25] Delaware law, you didn't have to prove [27:27] very much. And under Tennessee law, you [27:29] did have to prove some degree of actual [27:31] injury from the defamatory statement. [27:34] But the court said it didn't matter [27:35] which law applied because under any [27:38] circumstances, Mr. Starbucks had shown [27:41] emotional distress, particularly from [27:43] these people coming up to him and his [27:45] children and um the threats against his [27:48] family that he said were intensified [27:50] after the shooting of Charlie Kirk. And [27:53] so therefore, this information getting [27:54] out to the public could lead to more [27:56] threats to his family. So the judge said [27:59] there's enough information showing that [28:01] you've suffered real damages to let this [28:03] case go forward. U but again, I think [28:06] these these cases together, like you [28:08] might be inclined to say, well, they're [28:11] just motions to dismiss. Who knows [28:13] what's really going to be proved in [28:15] discovery. I seriously doubt if in [28:18] discovery they're going to show that 2.8 [28:21] 8 million people received a defamatory [28:23] statement about Robbie Starbucks. [28:25] Seriously doubt that. Uh but what they [28:28] are sending a clear signal on is if you [28:32] are a generative AI company, you better [28:35] get your notice and cure system, [28:38] you know, [28:40] working efficiently because this is a [28:42] scenario that's going to keep happening. [28:46] >> Oh my goodness. So so so many questions. [28:48] So the first is I want to pick up on [28:51] professor Liski what you just mentioned [28:52] about the notice and cure system. [28:56] My understanding of these these models [28:59] is that every time it's kind of like a [29:00] like this game of Russian roulette, [29:03] right? And you sort of have these sort [29:04] of probabilistic models and they're uh [29:07] their different ways. I mean there's [29:08] it's it's quite complex. [29:12] There are sort of two potential cures [29:14] here. Number one is you presumably [29:15] professor Volik your remedy is [29:17] post-processing filters in a lookup [29:18] table. Is that [29:20] >> is that is that correct? [29:21] >> That is one thing that I can just based [29:24] I was a computer programmer for years [29:27] back in the 80s. So we used to write [29:29] programs by hammering them out on clay [29:32] tablets. Uh but it's the kind of thing [29:35] you could imagine that if if you had to [29:37] if you had to design it without actually [29:39] without being being even in in the [29:41] industry you could say look there are [29:44] technologies there are simple text [29:46] processing technologies that you could [29:48] use they may be crude they may be [29:50] imperfect they may end up overblocking [29:53] in some respects and underblocking in [29:54] other respects but you've created these [29:57] amazing tools that can answer apparently [30:01] like take the uh pass the bar exam. It's [30:05] a very high level and so on and so [30:06] forth. And then if the company says, [30:09] "Oh, but to check to see if this is a [30:12] known uh hallucination, oh, that's [30:15] beyond our powers." That just doesn't [30:17] seem terribly credible. [30:22] >> One of the things that's interesting to [30:23] me here is is there are various [30:26] solutions after the fact to try to [30:28] eliminate the hallucination. And one is [30:31] the simple block list. So, [30:33] you know, I can't have the words [30:36] together. Robbie Starbucks and sexual [30:39] harassment. And as as Eugene points out, [30:42] they're pretty crude, but they're pretty [30:44] simple and effective for some things. [30:47] Now, it's easy to do an end run around [30:49] them. You could say um [30:51] >> uh the I forget what Robbie Starbucks [30:55] wife's name, but the the what is the [30:57] husband of what's her name Starbucks [31:00] accused of? and potentially you could [31:03] generate the information but the crude [31:05] block seems like an easy cheap first [31:08] step that in negligence terms you would [31:11] say it's kind of negligent not to deploy [31:13] that uh you know we talk you you uh are [31:17] named you name your company after [31:19] learned hand learned hand is famous for [31:21] the the burden of taking precautions [31:24] being less than the probability of harm [31:26] times the gravity of harm well a block [31:28] list is a simple one but here's the [31:31] question. There are other ways of going [31:33] at um filtering out material. There's uh [31:37] rag generation. So you generate the [31:39] sources so that the the reader can check [31:41] them themselves. That's another line of [31:43] defense. And then there's what what's [31:47] been called machine unlearning which I [31:49] guess is fairly expensive like changing [31:51] the weights fundamentally in a model [31:53] ends up being fairly expensive. [31:56] Might it be reasonable though in a case [31:59] if you really have a model that has [32:02] decided Mr. uh Starbucks is public enemy [32:06] number one is and is on a campaign of [32:09] character assassination. It seems that [32:12] even a relatively expensive uh process [32:15] perhaps of machine unlearning might be [32:18] warranted there. But in these cases [32:20] we're going to have to have technical [32:21] experts to tell us what's feasible and [32:23] what's not. And that standard is going [32:26] to be constantly evolving extremely [32:29] quickly as these cases get litigated. [32:32] >> Yeah. [32:34] Yeah. So that's absolutely right. uh uh [32:37] and uh uh I think that there is a lot [32:40] that the companies can do, but I will [32:42] tell you one thing that I think they [32:43] needn't do that I think the Delaware [32:46] court may have aired on is I don't think [32:49] they need to have to deal with someone [32:52] saying, "Oh, your own chatbot admitted [32:55] X." I don't think chat bots admit [32:57] things. Um uh I think uh if it were a [33:01] Google employee who said this, well, [33:03] that might be enough. But given that we [33:05] do know that uh that chatbots not just [33:09] sometimes make mistakes, humans [33:10] sometimes make mistakes, but that [33:12] chatbots don't actually [33:15] report based on what we as humans know [33:18] of as knowledge of information [33:21] uh but have these algorithms that are [33:23] based on kind of the frequency of words [33:25] being located together and such. I just [33:27] don't think that their output should be [33:30] seen as reliable evidence, not just for [33:32] purposes of evidence law, but even for [33:34] purposes of uh of uh what is a plausible [33:37] allegation. Now, as as Lissa pointed [33:40] out, um Delaware uh requires I believe [33:46] the standard is whether something is [33:47] that's alleged is conceivably true as [33:52] >> conceivable. Yeah. [33:53] >> Right. as opposed to plausible, which is [33:56] the federal standard. So maybe that's [33:58] the difference there. But I do think [34:00] it's a mistake uh to uh for for courts [34:03] to rely on allegations that are based [34:06] just on this is what uh uh this is the [34:09] output that the chatbot produced. Yeah. [34:11] >> Yeah. I I agree. I found that troubling [34:13] even at the complaint stage because it's [34:16] such an unreliable first off it's a [34:19] plaintiff's report of what the AI said [34:21] to the plaintiff. It's a very unreliable [34:23] piece of information. But I think one of [34:26] the larger themes we're seeing across [34:30] society here is the idea that you can [34:33] just wave your hands and say technology [34:35] and get a get out of jail free card. I [34:38] think that era is definitively over. [34:41] People are very suspicious of big tech [34:43] platforms. They're very suspicious of [34:46] AI. They're not they're, you know, uh [34:49] especially when you have these AI [34:51] doomers out there saying it's there's a [34:52] 10% chance it's going to kill us all [34:54] within [34:56] I forget how long, 10 years, 20 years. [34:59] >> U [35:01] people are just not accepting that AI [35:03] should get a free pass to do whatever. [35:06] >> Yeah. We're not we're not gonna have [35:07] like a redo of section 230 for the [35:09] social media platforms. Um if we have [35:12] time I want to come back to how these [35:14] questions can be included under the [35:16] umbrella of section 230 and whether [35:18] maybe whether a link to one of them [35:20] would would do it. One one question [35:22] before that though is [35:25] professor Lzky this and professor Volik [35:27] based on your sort of conversations [35:28] whether it's a lookup table I mean I can [35:30] imagine I mean Gemini's answers are [35:32] speedy quick if you use them they're so [35:34] so quick and even like a lookup at [35:37] Google scale like I'm not sure that [35:38] would probably double the latency but [35:40] they might decide that it's worth right [35:43] maybe not maybe not [35:44] >> lookups are pretty quick these days I [35:46] mean this is you're not going to have to [35:48] do a sequential search through the [35:51] Even back in the 80s, you could have [35:53] like a hash table. You it's it's not [35:56] going to slow things down that much [35:57] >> regardless is that there's there's like [35:59] an escalating level of technical [36:00] complexity. Like theoretically, a [36:02] product council at Google could say, [36:03] "Okay, we'll start with a lookup and [36:05] then we we can escalate to something [36:07] else and then we can go to machine [36:08] learning and we can cue those every two [36:11] months when the model's trained, etc." [36:12] And have that's probably the shape it's [36:14] going to take. Another approach would be [36:15] a little more cold-blooded and say, [36:17] "Look, there are nine cases over three [36:18] years, [36:19] >> right? Is this is this is this just the [36:21] cost of doing business? Yeah. [36:23] >> One one thing that struck me is I would [36:25] have expected with all the AI [36:28] >> in in sort of my line of work, we see [36:30] sort of the AI lawsuits coming through [36:31] the roof and we're trying to help courts [36:33] handle that sort of deluge. I'm shocked [36:36] that there haven't been more [36:38] >> right [36:39] >> defamation cases against the models. [36:42] What gives? Why have there been so few? [36:44] And uh [36:45] >> yeah, and it's personal for me because I [36:47] wrote the article and each case is an [36:49] opportunity for someone to cite it and [36:52] people are people are maliciously [36:55] designedly refusing to file these [36:58] lawsuits that make my article more [37:00] important. [37:01] >> You you might you might need to have [37:02] some interns sort of jin up some [37:04] lawsuits on the side. [37:04] >> There you go. [37:06] >> Um I mean, Professor Lisky, you also [37:09] sort of predicted there would be a new [37:10] form of bar trade. So, do either of you [37:12] have, you know, [37:13] >> I I did I talk about that about, you [37:17] know, if you just sit there searching [37:19] yourself, doing vanity searches over and [37:21] over until it says something defamatory [37:22] about you, if that's enough to generate [37:24] a case, it's it's truly problematic. And [37:27] so to that extent I do think we need a [37:30] little bit of a of a technological [37:32] understanding in defamation law that [37:34] that you know a single hallucinatory [37:37] output to a single person who's looking [37:39] for it is maybe not the basis for a good [37:42] case. [37:43] >> Doesn't that doesn't that relate to the [37:45] publication requirement? meaning because [37:46] I think as sort of a naive approach I [37:50] think that that strikes me as the the [37:53] strongest argument meaning you at least [37:55] for allegations you need to show over [37:57] many different IP servers with different [37:59] VBNs or etc to show that this is [38:01] something that's innate to the model and [38:03] is getting published not only as a [38:06] matter of liability not just as a matter [38:07] of damages. So [38:09] where do you predict that courts will [38:11] end up on on the publication [38:13] requirements? [38:15] So you know the standard rule in [38:17] defamation is that it only needs to be [38:20] published to one third party. Now that [38:22] may if it only was published to one [38:24] third party it may affect the quantum of [38:26] damages. The danger here is if it is [38:30] published to one third party first off [38:32] they can republish it. And so the harm [38:36] could come from the repetition of the [38:37] defamatory falsehood by that one third [38:39] party if if they're a reporter for [38:41] example and they rely unthinkingly [38:44] without verification on that [38:45] information. Uh but the other thing is [38:48] if one third party and the judge in in [38:50] uh Starbucks seemed to think about this [38:53] if it's published to one third party how [38:56] many other third parties that you don't [38:58] know about also received it. And what's [39:01] unique about these cases is we may be [39:03] able to have a record of that. So to the [39:06] extent that that we have logs that we [39:08] can see who searched for Robbie [39:10] Starbucks and what were they told when [39:12] they searched for Robbie Starbucks, we [39:14] might be able to uh you know unlike an [39:17] ordinary defamation case really very [39:20] accurately verify the the primary extent [39:24] of the audience. Of course, they might [39:25] have said it to other people. Um, but [39:28] but there's a tool here for really um I [39:32] don't think the publication element will [39:34] change, but I do think our demand that [39:36] the damages be real might be affected by [39:39] these cases. [39:46] >> Is there is there not an argument though [39:48] that like one chat results I mean even [39:51] the word chat seems to be that it's [39:54] meaning [39:55] It's almost like an an individual email. [39:58] I would say it's more akin to that. I [40:00] guess what is the standard for public [40:04] publication? [40:05] >> Publication is one person. [40:09] >> Classic illustration of uh of liel a [40:12] letter. [40:13] >> Yeah. send a letter to someone saying [40:15] this person you're about to marry [40:19] >> that's enough [40:20] >> criminal record or abused me or whatever [40:22] else or I mean that is a hornbook [40:25] illustration especially from back in the [40:27] day uh when uh uh I was about to say [40:31] when people were pickier about whom they [40:32] read that's not right but but when when [40:35] the kinds of allegations may have been [40:37] somewhat different but these days one [40:39] thing that's very common is job [40:40] references if somebody says u uh If if [40:43] if someone asks me, "Oh, should I I hire [40:47] so and so?" I say, "No, uh we fired him [40:50] because he was stealing from petty [40:52] cash." Uh that is publication. Even [40:55] though it's said to one person, that is [40:56] a wellestablished principle of [40:58] defamation law. [41:00] >> It's it's even publication if it's [41:01] foreseeable a piece of snail mail uh [41:05] that goes to their executive assistant. [41:08] If you know they're going to open it [41:09] first, if it's reasonably foreseeable [41:11] that that third person is the person who [41:13] gets it, that's publication. So [41:15] publication is really a very low [41:17] standard, [41:18] >> right? But plus also one other thing if [41:20] you think about it, uh how do webs say [41:23] something is put on a website, we think [41:25] that's quintessential publication. Well, [41:26] how do websites work? Well, one [41:30] user at a time, one set of messages at a [41:33] time. If even if 10,000 people visit the [41:37] website and see something, each one is [41:39] getting just a message in response to a [41:41] query. Uh that doesn't keep it from [41:44] being publication. So um yeah, [41:47] >> I mean one thing that the question [41:49] implicates is maybe uh one model of [41:52] thinking about these is thinking about [41:54] them like distributors which are not [41:56] treated like traditional publishers. So, [42:00] uh, you know, a a a a conduit for [42:03] information [42:05] and so it's they're not responsible for [42:08] the information until they're put on [42:10] notice of the contents and then fail to [42:12] do something. So, that's that's one [42:14] argument is they're that they're they're [42:17] not really like a newspaper in the way [42:20] they're putting out information. They're [42:22] more like a telephone or telegraph until [42:26] notice occurs or a bookstore. [42:28] >> Right? So I can imagine that being being [42:30] a rule that's adopted. One thing I [42:33] should stress is that they really are in [42:36] many ways like traditional publishers in [42:38] that they're the ones who actually [42:40] create the output. Uh they being the [42:43] company, the company creates the output. [42:45] So traditional. So one reason why [42:47] bookstores and news stands are treated [42:50] as distributors and even under the old [42:53] liable law which was quite friendly to [42:55] strict liability even in certainly [42:57] negligence liability requires [43:00] distributors were only liable if they [43:02] really were on notice that something was [43:03] defamatory is you can't expect a [43:05] bookstore owner to read every book um uh [43:09] uh that's on the shelves. You can't [43:11] respend expect a news stand operator to [43:14] read every issue of of a newspaper that [43:18] that arrives in the morning. Uh uh but [43:22] uh but I think people would say you can [43:24] expect companies to at least be to to at [43:27] least have mechanisms for making sure [43:30] that what is after all created by them [43:33] and let out by them into the stream of [43:35] commerce is uh uh is accurate. By the [43:38] way, one thing we haven't talked about [43:40] yet, but I'm sorry, Lissa. [43:42] >> Yeah, just one point on that though. I [43:44] mean, given the way generative AI works, [43:46] can you expect there never to be a [43:48] defamatory hallucination? [43:50] >> Right. So, I don't think you expect [43:52] never to be. [43:53] >> So, so here's my thing, and this is [43:54] something we haven't much talked about [43:56] because both of the cases we talked [43:57] about, Keen and Starbucks, involved [43:59] basically people who were least limited [44:00] purpose public figures. Uh, what about [44:02] for private figures? The official [44:04] standard is negligence. at least if you [44:06] can prove actual compensatory damages. [44:09] Um uh there is an very interesting case [44:12] in Minnesota uh which being litigated [44:15] now in Minnesota trial court which both [44:17] of us have written about called LTL. Um [44:21] uh and it involved a company a uh solar [44:25] uh solar panel installation company uh [44:28] that alleged I think correctly alleged I [44:30] think I I I repeated the query and saw [44:34] the same results at one point. uh that [44:36] uh Google was uh uh was outputting false [44:39] statements about it and says look here [44:41] we actually have people talk bring [44:43] receipts. Well, we have receipts or very [44:45] close to it. We have emails from people [44:47] saying we're not doing business with you [44:49] anymore. After all, you guys are being [44:52] investigated by the state attorney [44:53] general for consumer fraud. They're not [44:55] being investigated. Nobody had said they [44:58] were being investigated until Google uh [45:00] uh output that. So, they're saying [45:02] negligence. Well, um, one, I think, [45:06] close analogy, and it's just an analogy. [45:08] It's not that the law would actually [45:10] apply here by its own terms, but I think [45:12] it's a helpful analogy, is design defect [45:15] liability for products, right? Um, so [45:18] let's say it was a self-driving car, and [45:20] let's say it caused not reputational [45:21] injury, but physical injury. Design [45:24] defect liability is basically negligence [45:26] liability. Uh, because you're only [45:28] liable, uh, the defendant is only liable [45:31] if the design was unreasonable. How do [45:32] we tell if it's unreasonable? Well, it [45:34] depends if there were cost effective [45:37] mechanisms for solving that. Again, the [45:38] learn at hand uh formula. Uh and [45:42] Schlommo, I think this responds a little [45:43] bit to to to your point about well, what [45:46] if it just slows things down too much? [45:48] Well, it's kind of like with with car [45:51] manufacturer. If you say you could have [45:53] installed this new feature uh that would [45:56] have avoided this accident, the car [45:57] manufacturer says it would have cost [45:58] $50,000. it would have increased the [46:00] price of the car by $50,000 would made [46:02] it unaffordable to people. The law at [46:05] least in theory says okay fine that is [46:07] it was reasonable for you not to include [46:09] it. But if you avoided including a $5 [46:12] feature well then in that case it was [46:14] unreasonable potentially for you. So I [46:17] think um something similar would happen [46:19] here. I think they could say look we did [46:21] the best we could. We we were aware [46:23] we're putting out this thing that is [46:25] dangerous. Again, dangerous to [46:26] reputation, not not to people's bodies, [46:29] but liel law allows for recovery for [46:33] negligent harm for reputation the way [46:35] the design defect law does for negligent [46:37] harm to people's bodies and property. Um [46:39] so, uh they could say, look, we did the [46:42] best we could under the circumstances. [46:44] And the plaintiff say no, no, no, look [46:46] what else you could have done. The [46:47] result, of course, would be a battle of [46:48] the experts, uh which is never [46:51] particularly edifying. I think and [46:52] there's always questions uh of whether [46:55] the jury is going to be able to tell [46:57] which computer science professor uh is [46:59] is more accurate here but that's just [47:02] the way we do things in our legal system [47:04] in the context of product liability and [47:07] another in the context of medical [47:08] malpractice liability and the like. Uh [47:11] and you know things go on it's sort of [47:14] seen as the least bad of the option. So [47:16] maybe that's the way things might play [47:18] out. What do you think? There's already [47:20] precedent for that in the Walters versus [47:21] OpenAI summary judgement because there [47:24] the judge credited experts that said we [47:26] designed this, you know, in in the we [47:29] did we used every available safety [47:31] mechanism we could uh to try to [47:34] eliminate this kind of hallucination and [47:36] to warn and and so forth. So the [47:38] warnings will come into play as well. [47:41] what kinds of disclaimers, what kinds of [47:42] warnings, what kinds of design choices [47:46] are sufficient to protect against these [47:48] harms, [47:49] >> right? Uh yeah, I think that that's [47:51] absolutely right. And I do think that [47:52] that Lissa, what you said is I think [47:54] quite important. Disclaimers won't get [47:56] the company off the hook uh [47:59] categorically, but they surely are [48:01] relevant in various uh various contexts. [48:04] And sometimes [48:06] the way you do things more reasonably or [48:08] maybe the mo the only reasonable thing [48:10] you could do is to have a prominent [48:12] disclaimer or something like that. Uh um [48:16] so so so those they're certainly [48:18] relevant to the analysis even if they're [48:20] not. [48:21] >> Well, and the other thing that's [48:23] relevant to the analysis is including [48:24] links to sources that have the [48:26] information for the user to verify and [48:29] then you may get some negligence on the [48:31] part of the prompter or user if if [48:34] there's easy verification available and [48:36] they choose not to take advantage of it. [48:38] For example, [48:39] >> I'm sorry. Go on. Go on, Lissa. Sorry. [48:41] Well, I was going to say I think a [48:42] journalist who uses information from AI [48:45] without verifying it very likely could [48:49] have actual malice in repeating that [48:51] statement because, you know, lawyers [48:54] right now are finding out using AI [48:57] information without verification can can [48:59] get you sanctioned and and cause uh [49:01] serious professional harm. [49:03] >> Right. So, so this may be some one thing [49:05] that maybe Lissa and I do disagree in [49:07] some measure on here. Um, so one thing I [49:10] want to point out and I think Lissa and [49:12] I would agree on this as a as a [49:13] descriptive matter, although maybe we [49:15] have different sense of the implications [49:16] of this is uh if somebody carelessly [49:22] relies on AI output in just not doing [49:26] business with someone or in cutting off [49:28] a contract or something like that, we [49:30] might loosely call them negligent, but [49:32] there's no there's no way that the uh [49:36] defamed person can recover from them. [49:38] There's no tor of negligent failure to [49:41] do business with someone, right? Uh or [49:45] uh uh somebody sent a letter to my [49:49] fiance and she broke it off with me and [49:51] it was really foolish of her to believe [49:52] that I'm going to sue her for negligent [49:54] failure to marry me. Right. uh uh uh so [49:58] uh there there wouldn't there wouldn't [50:00] be any recourse for a defamed person uh [50:05] against some against people who believed [50:07] however foolishly believed the [50:09] defamation but with regard to somebody [50:12] who then republishes it Lissa I think [50:15] you're right that there that there um [50:18] might be in principle some course but [50:21] let me ask a couple of things first of [50:22] all let's say that that a reporter [50:26] thoughtlessly, [50:28] literally thoughtlessly in the sense of [50:30] without thinking about there might be an [50:33] error, however foolishly, however [50:35] carelessly, but thoughtlessly [50:37] republishes something that he or she saw [50:39] in a uh in a AI output. Would that be [50:44] actual malice? I think it could be gross [50:45] negligence, but I don't think it would [50:47] be actual malice. [50:49] >> You know, I guess it depends on what you [50:52] think the reliance is. Is it akin? I [50:55] often talk about conscious to my [50:56] students about conscious disregard of of [50:59] of falsity as the ostrich defense. You [51:02] put your head in the sand and you avoid [51:06] knowing the truth. [51:07] >> Right? But I'm thinking of non-concious [51:09] disregard. Just somebody gets up on the [51:11] stand afterwards and says, "Look, I [51:14] realize it was totally idiotic for me to [51:17] to to just believe it, but I just zoned [51:19] out." [51:20] >> You could argue it's gross negligence. I [51:22] think it's at least arguable it would be [51:24] reckless disregard for the truth [51:26] >> and and you know as time goes on the [51:30] veilance on that may change [51:32] >> right so here let me ask you another [51:33] question and this is a this is about [51:35] section 230 so section 230 won't provide [51:38] protection for uh the uh for the AI [51:42] company I think we we agree on that [51:44] because they're the ones who generate [51:45] the output [51:46] >> they generate it yes [51:47] >> but but um section 230 does provide [51:52] protection for online speakers. So, this [51:55] would only be for an online reporter who [51:58] merely pass along information provided [52:01] by other online speakers. [52:05] Um uh so I wonder if and maybe this is [52:08] just a defect with section 230, but I [52:10] wonder if the reporter if the reporter [52:12] publishes only electronically and not on [52:15] paper in in your scenario. The reporter [52:17] who foolishly uh relies on the AI um [52:22] output might actually have a section 230 [52:24] immunity because he's just passing along [52:26] whatever uh whatever he got online. [52:30] >> You know, that's a good question. And I [52:31] I think on a very technical reading of [52:34] 230, maybe you could get there. I don't [52:36] think that's what 2:30 was meant to do, [52:38] >> right? [52:39] >> And um you know, as a policy reading of [52:42] 230, you probably shouldn't go there. [52:45] >> But I think you're right. On a very [52:47] technical reading, they could argue that [52:49] 230 absolved them because it's it's not [52:51] userenerated content, but it's content [52:54] generated by another. [52:56] >> Right. Right. Right. So there is a [52:58] there's a science fiction novel that I [53:01] love excellent one probably one of my [53:03] favorites by professor sorry by an [53:06] author who happened to be a computer [53:07] science professor Verer Vji he died a [53:09] few years ago it's called a fire upon [53:11] the deep uh and uh it's set in the far [53:16] future but it there is this interstellar [53:20] network that was clearly based on the [53:22] early internet in fact if anything the [53:24] preodern so-called Usenet that but there [53:27] was a line that people would use there [53:30] which just like a proverb a far future [53:32] proverb that was they don't call it the [53:34] net of a million lies for nothing [53:37] [laughter] [53:38] so so boy has that proven true to the [53:40] internet generally and now seems to be [53:42] AI by the way before we go uh Schlommo I [53:45] just wanted to mention one thing for the [53:46] benefit of our readers I'm sure you are [53:48] well acquainted with this but learned [53:50] hand learned wasn't the judge's first [53:54] name [53:55] >> the first name was Billings, which for a [53:58] lawyer, how much better for a male [54:00] lawyer, for a female lawyer, the best [54:02] name is just Sue, that's very common. [54:05] But for a male lawyer, Billings, how [54:07] much better can you get than that? I [54:09] guess when he became a judge, he had to [54:11] call learn it because, you know, he [54:13] didn't bill by the hour. [54:14] >> Truly, it's his parents who are most [54:16] farsighted in this, [54:18] >> though. his uh I I wonder what was going [54:21] in the hand family table with him and [54:24] his uh cousin Augustus Noble. [54:28] >> So um okay well uh any other final [54:32] comments uh before uh we wrap up what [54:35] was I think a very uh uh informative uh [54:39] panel. There are a few questions from [54:41] the audience but I'm not sure we'll have [54:42] time to [54:43] >> Oh I'm so sorry. Yeah. [54:44] >> No, no, no. This is [54:45] >> got so carried away. Maybe time for one. [54:48] Do we have time for one? [54:49] >> All right, I'll we'll we'll be quick. [54:50] One is about um traditional actual [54:53] malice doctrine requires the relevant [54:55] state of mind of an institutional [54:56] defendant to be quote brought home [54:58] unquote to those responsible for [55:00] publication. How should that rule [55:01] operate when a corporate defendant [55:02] deploys an AI system that generates and [55:04] publishes the particular factual [55:06] assertion without any human having [55:09] composed, reviewed or perhaps even known [55:11] of that assertion before publication? Is [55:13] existing institutional art attribution [55:15] doctrine sufficient or does autonomous [55:17] AI publication create an accountability [55:19] gap requiring a new theory of [55:20] attribution? [55:23] >> I think in these cases we talked about [55:25] today it's pretty easy because you have [55:27] notice and it's what happens after [55:30] notice. It's easy enough to prove actual [55:31] malice with regard to action or inaction [55:34] after notice. I think the harder [55:36] question is applying actual malice to [55:39] cases before notice. [55:41] >> Right. And so I I think that's very hard [55:43] to do because all you would like I don't [55:47] think saying statistically it produces a [55:49] certain number of hallucinations is [55:51] enough to put you on notice of a [55:53] specific defamatory falsehood. And what [55:56] I would do is adapt the doctrine to say [55:58] that that you have to at least have [56:00] notice to have a case. But um yeah, [56:04] >> right. In actual malice cases, I I agree [56:06] entirely with with [56:07] >> in actual Yes. In actual malice cases [56:09] >> that uh that uh uh I've heard people [56:12] say, well, you you knew for sure that it [56:15] was saying some false things. Well, it's [56:16] kind of like saying you knew that your [56:19] that the cars you manufacture would kill [56:21] someone, so we're not only going to try [56:23] to sue you for negligence, we're going [56:24] to prosecute you for murder because you [56:27] you put out cars that you knew would [56:29] kill someone. All right, but that kind [56:30] of statistical knowledge can't be [56:32] enough. I think for good policy reasons [56:34] and the law generally reflects that. So [56:36] I agree entirely with Lisa on that for [56:38] negligence-based claims again private [56:40] figures or speech on matters of private [56:42] concern that uh uh that is a different [56:46] matter precisely because there could be [56:48] non actual malice liability. [56:51] >> Fabulous. Um well, thank you all for [56:54] that. And now I'm gonna uh make sure we [56:57] all order our copy of uh a fire um a [57:01] fire upon the deep as well. [57:04] >> Uh uh I I stand by that recommendation. [57:08] It it is fantastic. [57:11] >> Thank you, professors. [57:14] >> Thank you. [57:15] >> On behalf of the Federalist Society, I [57:17] want to thank our experts for the [57:19] benefit of their valuable time and [57:21] expertise today. And I want to thank our [57:23] audience for joining and participating. [57:25] We also welcome listener feedback by [57:27] email at fedsockforumsfedsock.org. [57:31] As always, keep an eye on our website [57:33] and your emails for announcements about [57:35] upcoming virtual events. Thank you all [57:37] for joining us today. We are adjourned.
