Large Libel Models? When Do AI Hallucinations Become Defamation?

WEBVTT
1
00:00:01.590 --> 00:00:06.190
Welcome to FedSoc Forums, a Federalist Society podcast. I'm Elizabeth Slattery,
2
00:00:06.349 --> 00:00:09.880
Vice President of Practice Groups. Be sure to subscribe wherever
3
00:00:09.910 --> 00:00:13.419
you get your podcasts so you won't miss an episode. Now,
4
00:00:13.480 --> 00:00:15.359
please enjoy today's conversation.
5
00:00:17.739 --> 00:00:23.019
Hello, everyone, and welcome to FedSoc Forums. Today, September 22nd, 2026,
6
00:00:23.019 --> 00:00:27.940
we are delighted to host a conversation titled Large Libel Models,
7
00:00:28.339 --> 00:00:33.270
When Do AI Hallucinations Become Defamation? My name is Lily Freemeyer,
8
00:00:33.409 --> 00:00:35.520
and I am a second-year law student at Washington and
9
00:00:35.549 --> 00:00:39.119
Lee University. I also serve as a student liaison to
10
00:00:39.140 --> 00:00:44.740
the Federalist Society's Communications and Technology Practice Group. As always,
11
00:00:44.899 --> 00:00:47.939
please note that all expressions of opinion are those of
12
00:00:47.979 --> 00:00:51.570
the experts on today's call, as the Federalist Society takes
13
00:00:51.630 --> 00:00:57.070
no position on particular legal or public policy issues. Our moderator,
14
00:00:57.210 --> 00:01:00.530
Shlomo Clapper, is the founder and CEO of Learned Hand,
15
00:01:00.920 --> 00:01:05.939
the first and only generative AI platform purpose-built for the judiciary.
16
00:01:06.760 --> 00:01:10.439
Learned Hand's technology supports judges and court staff across the
17
00:01:10.540 --> 00:01:13.760
full arc of a case, from filing through drafting, and
18
00:01:13.819 --> 00:01:17.549
is currently deployed with the Michigan Supreme Court, the Superior
19
00:01:17.579 --> 00:01:20.549
Court of Los Angeles County, and courts in 10 states,
20
00:01:20.769 --> 00:01:25.659
serving over 20 million Americans. Before founding Learned Hand, Shlomo
21
00:01:25.739 --> 00:01:27.480
clerked for the Honorable Steven J.
22
00:01:27.519 --> 00:01:28.920
Menasche on the U.S.
23
00:01:28.980 --> 00:01:33.230
Court of Appeals for the Second Circuit, litigated complex commercial
24
00:01:33.269 --> 00:01:37.629
disputes at Quinn, Emanuel, Urquhart, and Sullivan, and worked as
25
00:01:37.670 --> 00:01:42.230
a deployment strategist at Palantir Technologies in both the commercial.
26
00:01:41.890 --> 00:01:42.790
And government sectors.
27
00:01:44.010 --> 00:01:47.189
Earlier in his career, he was a behavioral researcher in
28
00:01:47.269 --> 00:01:51.769
Dan Ariely's lab at Duke University. Shlomo received his JD
29
00:01:51.829 --> 00:01:54.950
from Yale Law School, where he won the Thomas I.
30
00:01:55.069 --> 00:01:58.269
Emerson Prize for his writing and holds a BS in
31
00:01:58.329 --> 00:02:02.010
economics from the Wharton School and a BA in European
32
00:02:02.049 --> 00:02:05.489
history from the University of Pennsylvania, both summa cum laude.
33
00:02:07.040 --> 00:02:09.699
One note for our audience before we begin, if you
34
00:02:09.740 --> 00:02:12.439
have any questions throughout the webinar that pertain to the
35
00:02:12.479 --> 00:02:15.669
topic at hand, please submit them through the Zoom's Q
36
00:02:15.669 --> 00:02:18.379
& A feature located at the bottom of your screen,
37
00:02:18.879 --> 00:02:21.129
not the chat, so that we'll be able to access
38
00:02:21.169 --> 00:02:24.289
them later in the program. With that note, I will
39
00:02:24.330 --> 00:02:27.849
turn things over to our speakers and to start the conversation. Shlomo,
40
00:02:28.009 --> 00:02:30.050
thank you so much for joining us and for moderating
41
00:02:30.090 --> 00:02:30.830
today's discussion.
42
00:02:32.129 --> 00:02:32.689
Thank you, Lily.
43
00:02:33.729 --> 00:02:37.530
Hopefully I will not be guilty of any libel, intentional
44
00:02:37.569 --> 00:02:41.939
or otherwise, throughout the conversation. So to introduce our speakers,
45
00:02:42.000 --> 00:02:48.099
I actually asked Gemini to write a brief intro of
46
00:02:48.840 --> 00:02:52.780
Professor Olitski and Professor Volek for the large libel models panel.
47
00:02:52.840 --> 00:02:57.159
So let's, professors, you can let me know if we're
48
00:02:57.180 --> 00:03:03.189
gonna add a 10th lawsuit All right. So, Professor Larissa
49
00:03:03.229 --> 00:03:06.729
Lidsky is the Raymond and Miriam Ehrlich Eminence Scholar Chair
50
00:03:06.770 --> 00:03:09.270
in U.S. Constitutional Law at the University of Florida Levin
51
00:03:09.310 --> 00:03:11.849
College of Law and former dean of the University of
52
00:03:11.889 --> 00:03:14.729
Missouri School of Law, a leading national authority on the
53
00:03:14.770 --> 00:03:17.349
intersection of tort law and the First Amendment. Her scholarship
54
00:03:17.389 --> 00:03:20.870
focuses on media law, free expression in digital spaces, and
55
00:03:20.990 --> 00:03:25.629
internet anonymity. She serves as a co-reporter for the ALI
56
00:03:25.870 --> 00:03:28.810
ongoing restatement of the law, third torts, defamation and privacy,
57
00:03:28.889 --> 00:03:31.379
positioning her at the forefront of efforts to adapt traditional
58
00:03:31.409 --> 00:03:34.599
libel principles to modern technological disruptions.
59
00:03:35.479 --> 00:03:38.020
Professor, is that substantively correct?
60
00:03:38.819 --> 00:03:39.099
It is.
61
00:03:40.000 --> 00:03:40.240
Great.
62
00:03:40.689 --> 00:03:45.879
All right. Professor Volokh is the Thomas M.
63
00:03:46.340 --> 00:03:49.259
Seibel Senior Fellow at the Hoover Institution at Stanford University
64
00:03:49.280 --> 00:03:52.240
and the Gary T. Schwartz Distinguished Professor of Law Emeritus
65
00:03:52.319 --> 00:03:55.379
at UCLA School of Law. One of the most widely
66
00:03:55.439 --> 00:03:57.759
cited First Amendment scholars in the country, he pioneered the
67
00:03:57.780 --> 00:04:01.580
scholarship on AI-generated defamation with his seminal paper, Large Liable Models,
68
00:04:01.620 --> 00:04:05.159
Liability for AI Output. He's the founder of the legal blog,
69
00:04:05.219 --> 00:04:08.370
The Volokh Conspiracy, founder and editor-in-chief of of the Journal
70
00:04:08.389 --> 00:04:10.930
of Free Speech Law and the author of leading case
71
00:04:10.969 --> 00:04:12.210
books on First Amendment law.
72
00:04:13.610 --> 00:04:17.089
Professor Wallach, is that also? So far, so good. Perfectly accurate.
73
00:04:17.110 --> 00:04:17.410
So good.
74
00:04:18.449 --> 00:04:19.740
Okay, fantastic.
75
00:04:20.000 --> 00:04:23.439
So to start us off, Professor Wallach, can I turn
76
00:04:23.459 --> 00:04:28.120
to you? Could you share with the audience the facts
77
00:04:28.180 --> 00:04:32.639
of Keene and sort of situate this into the other,
78
00:04:33.170 --> 00:04:36.370
in the context of the other lawsuits that have surfaced?
79
00:04:37.540 --> 00:04:38.560
Yeah, so.
80
00:04:40.040 --> 00:04:42.779
The case we're talking about right now is just from
81
00:04:43.100 --> 00:04:46.180
about a week or so ago, Keen v. Google, although
82
00:04:46.220 --> 00:04:49.139
it is one of nine cases that has been filed
83
00:04:50.220 --> 00:04:52.579
over alleged libel by AI in U.S.
84
00:04:52.639 --> 00:04:57.889
Courts over the last three years. So the plaintiff apparently
85
00:04:57.910 --> 00:05:01.769
goes by Jimmy. Jimmy Keen is, in 1997, he had
86
00:05:02.110 --> 00:05:03.430
pleaded guilty.
87
00:05:03.350 --> 00:05:09.269
To a drug distribution conspiracy, was sentenced to 10 years
88
00:05:09.329 --> 00:05:13.170
in prison, but then agreed to work with the FBI
89
00:05:13.230 --> 00:05:15.629
as a prison informant to help gather evidence against the
90
00:05:15.670 --> 00:05:21.290
suspected serial killer who was also in prison. And he
91
00:05:21.310 --> 00:05:24.209
did that. And as a result, he got a pardon.
92
00:05:25.449 --> 00:05:31.389
And his conviction was expunged. Some people whose convictions are
93
00:05:31.439 --> 00:05:33.860
expunged work really hard to kind of keep them keep
94
00:05:33.879 --> 00:05:34.399
them quiet.
95
00:05:34.459 --> 00:05:36.519
In his case, he published a memoir.
96
00:05:36.480 --> 00:05:39.959
Recounting this experience was subsequently adapted in an Apple TV
97
00:05:40.279 --> 00:05:41.660
miniseries titled Black Panther.
98
00:05:42.420 --> 00:05:44.079
And as best I can tell, he's led a blameless
99
00:05:44.139 --> 00:05:44.699
life since.
100
00:05:46.180 --> 00:05:50.439
Well, maybe this was a blameless life, but he has
101
00:05:50.480 --> 00:05:58.589
been blamed by, he says, Google AI overviews, which at
102
00:05:58.629 --> 00:06:03.970
various times in the middle of 2025 said various things
103
00:06:04.009 --> 00:06:06.850
in response to queries such as Jimmy Keen net worth.
104
00:06:08.170 --> 00:06:09.889
And the chief thing that they said.
105
00:06:09.949 --> 00:06:12.029
According to the complaint, there's another twist to it, which
106
00:06:12.069 --> 00:06:13.970
we'll get to a little later on, is that he
107
00:06:13.990 --> 00:06:19.389
was serving a life sentence, including a life sentence without parole,
108
00:06:19.680 --> 00:06:24.560
including for the murders of three women. That is apparently
109
00:06:24.660 --> 00:06:26.000
totally hallucinated.
110
00:06:28.079 --> 00:06:31.120
Note what it's not. It's not that there is some
111
00:06:31.300 --> 00:06:32.420
error online.
112
00:06:33.279 --> 00:06:38.870
That was then echoed by Google AI overview. The way
113
00:06:38.930 --> 00:06:42.069
if you do a Google search and there's incorrect information
114
00:06:42.129 --> 00:06:44.370
on some page, Google search might give you a snippet
115
00:06:44.389 --> 00:06:48.430
of that. So, some people say, well, the problem with
116
00:06:48.439 --> 00:06:56.209
these AI large language models is garbage in, garbage out. No,
117
00:06:56.240 --> 00:06:59.220
this is a different kind of Geico. This is gold in,
118
00:06:59.279 --> 00:06:59.899
garbage out.
119
00:07:00.319 --> 00:07:02.500
As best we can tell, all of the information about
120
00:07:02.540 --> 00:07:03.459
him was quite accurate.
121
00:07:04.250 --> 00:07:07.050
Despite that, the allegation is, and of course at this
122
00:07:07.089 --> 00:07:17.379
point it's just allegations, that Google's AI output made these assertions.
123
00:07:17.660 --> 00:07:19.439
And by the way, note, it's just as a result
124
00:07:19.459 --> 00:07:22.300
of a normal Google search. It's not somebody specifically asking
125
00:07:22.379 --> 00:07:25.800
for AI. It's just the AI output. It's just the
126
00:07:25.819 --> 00:07:31.899
AI overview that is normally presented. So he sued, Google
127
00:07:31.930 --> 00:07:36.370
moved to dismiss, and the court allowed the case to
128
00:07:36.389 --> 00:07:40.970
go forward. So let me quickly go through the key
129
00:07:41.029 --> 00:07:46.069
points in Google's argument, then mention one point that really
130
00:07:46.120 --> 00:07:50.860
wasn't discussed, and I think for good reason. So first,
131
00:07:50.959 --> 00:07:54.660
Google argued, well, there's a disclaimer that says, well, AI
132
00:07:54.740 --> 00:07:58.509
overviews are only starting points, and that It warns that
133
00:07:58.569 --> 00:08:01.910
it may include mistakes. And the court says, well, it's
134
00:08:01.949 --> 00:08:05.100
not clear whether a disclaimer was actually present.
135
00:08:05.160 --> 00:08:06.180
It's a motion to dismiss.
136
00:08:06.870 --> 00:08:09.689
We're going to assume the fact is alleged in the complaint,
137
00:08:09.910 --> 00:08:13.189
and here there was no such allegation. This having been said,
138
00:08:13.310 --> 00:08:18.370
I should add that disclaimers by themselves are generally not
139
00:08:18.430 --> 00:08:20.829
enough in libel litigation.
140
00:08:23.230 --> 00:08:25.430
Now, if a disclaimer is, this is all fiction, this
141
00:08:25.490 --> 00:08:25.910
is all.
142
00:08:25.829 --> 00:08:28.540
A joke, this is parody, well, that might be enough,
143
00:08:28.569 --> 00:08:31.220
because that says essentially, this is not true. But a
144
00:08:31.259 --> 00:08:33.259
disclaimer that says this might not be true is generally
145
00:08:33.299 --> 00:08:33.659
not enough.
146
00:08:33.940 --> 00:08:36.539
Classic example. We don't think of it a disclaimer, but
147
00:08:36.600 --> 00:08:39.740
it is. The phrase, rumor has it. Let's say I say,
148
00:08:39.799 --> 00:08:43.139
rumor has it that Shlomo Clapper, well, I'm not even
149
00:08:43.179 --> 00:08:46.809
going to bring this up. Who knows? I will train
150
00:08:46.870 --> 00:08:47.080
on it.
151
00:08:47.789 --> 00:08:50.470
I have dozens of witnesses here.
152
00:08:50.690 --> 00:08:53.629
Right, right. Exactly. Exactly.
153
00:08:54.370 --> 00:08:58.549
And if I say that, and it turns out, Especially,
154
00:08:58.570 --> 00:09:00.629
it turns out, there is no such rumor. But even
155
00:09:00.669 --> 00:09:03.139
if there is, the rumor is false, I could still
156
00:09:03.179 --> 00:09:08.019
be liable. Now, it may decrease the damage in some respect.
157
00:09:08.080 --> 00:09:10.419
In certain situations, it may be relevant. There is a
158
00:09:10.460 --> 00:09:14.379
special rumor privilege, but that's for very specific kinds of...
159
00:09:15.269 --> 00:09:21.070
repetitions of rumors. But basically, the law says, look, disclaimer
160
00:09:23.379 --> 00:09:30.080
doesn't eliminate the damage to reputation. Again, may diminish it
161
00:09:30.120 --> 00:09:32.620
in some measure, but doesn't eliminate it because, well, people
162
00:09:32.669 --> 00:09:35.679
know that some rumors— are false. They also know some
163
00:09:35.720 --> 00:09:39.860
rumors are true. Likewise, with Google, with any disclaimer that
164
00:09:39.879 --> 00:09:41.799
Google might have provided, I think that by itself would
165
00:09:41.820 --> 00:09:46.500
not be enough. A second point is that Google said, well,
166
00:09:46.799 --> 00:09:49.399
users could have easily verified that the AI overviews contained
167
00:09:49.440 --> 00:09:50.139
false information.
168
00:09:50.179 --> 00:09:53.090
There was a link to a Wikipedia article. People clicked
169
00:09:53.129 --> 00:09:56.070
on it. They would see that this wasn't so.
170
00:09:56.909 --> 00:10:01.779
But that's not enough to eliminate the liability. Simply saying, well,
171
00:10:02.419 --> 00:10:04.500
for more, read this. And then for the very few
172
00:10:04.539 --> 00:10:08.570
people who do read this, follow the link, they'll realize
173
00:10:08.610 --> 00:10:11.429
that it's false, but lots of people won't follow up.
174
00:10:11.809 --> 00:10:13.149
There was also another case.
175
00:10:13.190 --> 00:10:16.860
Remember, I said this is one of nine cases. Of them,
176
00:10:16.899 --> 00:10:20.830
some are pending, two have settled, one of them following arbitration.
177
00:10:23.799 --> 00:10:29.389
Another case in Delaware trial court has also allowed.
178
00:10:29.110 --> 00:10:31.570
The case to go forward, also denied a motion to dismiss.
179
00:10:31.750 --> 00:10:33.870
But there was a Georgia case where initially a motion
180
00:10:33.889 --> 00:10:36.750
to dismiss was denied, but then summary judgment was granted,
181
00:10:36.769 --> 00:10:37.629
called Walters v.
182
00:10:37.690 --> 00:10:40.639
OpenAI. It was the first such large libel models case.
183
00:10:41.720 --> 00:10:47.559
And the court in that case stressed that there wasn't
184
00:10:47.600 --> 00:10:49.059
just a disclaimer on.
185
00:10:49.019 --> 00:10:49.980
The part of OpenAI.
186
00:10:50.220 --> 00:10:52.889
It's that the Chad GPT actually.
187
00:10:54.039 --> 00:10:56.789
Said things like that I.
188
00:10:58.769 --> 00:11:04.549
Can't provide the desired information. This is not accessible to me. Eventually,
189
00:11:04.610 --> 00:11:07.759
it did do that. But there were lots of very
190
00:11:07.860 --> 00:11:11.360
specific signals to the person doing the searching, who, as
191
00:11:11.399 --> 00:11:12.179
it happens, actually.
192
00:11:13.580 --> 00:11:18.159
I think, personally knew the target of the libel and therefore.
193
00:11:17.940 --> 00:11:20.500
Didn't believe the allegations.
194
00:11:21.120 --> 00:11:25.480
So, there were lots of signals that were present that
195
00:11:25.580 --> 00:11:29.519
made it— was not just a disclaimer, just quite clear in
196
00:11:29.580 --> 00:11:32.620
context that this was an error. And the court says, well,
197
00:11:33.000 --> 00:11:35.830
for that reason, Walters is not really on point here,
198
00:11:35.870 --> 00:11:38.470
because at least the allegations in the complaint are that
199
00:11:38.710 --> 00:11:47.139
there were no such powerful signals there. Now as we know,
200
00:11:49.379 --> 00:11:54.889
The question of whether something is false and whether something
201
00:11:55.149 --> 00:11:57.129
is damaging to reputation is only.
202
00:11:56.990 --> 00:11:58.269
The first step in libel cases.
203
00:11:58.809 --> 00:12:03.179
If the plaintiff is a public figure, then famously, New
204
00:12:03.399 --> 00:12:03.950
York Times' E.
205
00:12:03.960 --> 00:12:05.740
Sullivan and its progeny.
206
00:12:05.279 --> 00:12:07.799
Say that there can only be liability based on a
207
00:12:07.840 --> 00:12:11.039
showing of actual malice. Actual malice, of course, is not
208
00:12:11.159 --> 00:12:14.649
actually malice. It means knowing or reckless falsehood.
209
00:12:15.169 --> 00:12:15.470
Well.
210
00:12:17.320 --> 00:12:21.779
What does that mean in the AI context? Where, in
211
00:12:21.820 --> 00:12:24.080
a sense, what does it mean for an AI to
212
00:12:24.200 --> 00:12:27.620
know something, much less be reckless about it? And by
213
00:12:27.659 --> 00:12:33.190
the way, the court treated, for purposes of this motion,
214
00:12:34.409 --> 00:12:38.509
treated Keene as a limited purpose public figure because of
215
00:12:38.570 --> 00:12:40.909
commentary on his own life story, which he has, of course,
216
00:12:42.210 --> 00:12:45.990
broadcast in the actual life story is broadcast.
217
00:12:45.610 --> 00:12:46.190
In his memoirs.
218
00:12:46.570 --> 00:12:50.649
So the courts said, and I think this is correct
219
00:12:50.700 --> 00:12:53.460
as a matter of libel law, although I'll query whether
220
00:12:53.480 --> 00:12:55.059
it makes policy sense.
221
00:12:55.080 --> 00:12:58.000
The court said that the real focus is on the
222
00:12:58.039 --> 00:12:59.279
mental state of the company.
223
00:13:00.250 --> 00:13:02.980
Not the program, which may not have a mental statement,
224
00:13:03.019 --> 00:13:05.639
may not have a mens, in the sense of mind,
225
00:13:06.019 --> 00:13:09.360
mens rea, guilty mind. It doesn't have a mind, but
226
00:13:10.080 --> 00:13:15.070
the company employs people who do. And Keen contends that
227
00:13:15.129 --> 00:13:18.169
he directly notified Google on three occasions it was publishing
228
00:13:18.230 --> 00:13:21.710
false information about him. And despite this notice, Google continued
229
00:13:21.750 --> 00:13:23.519
to publish similar defamatory statements.
230
00:13:24.519 --> 00:13:26.179
So, that, the court said, is.
231
00:13:26.139 --> 00:13:30.809
Sufficient allegation of actual malice, that it was aware, the
232
00:13:30.850 --> 00:13:33.769
company was aware, and didn't take steps to block that.
233
00:13:37.309 --> 00:13:41.730
One more thing. I remember I mentioned that the claims.
234
00:13:41.409 --> 00:13:47.779
About him put out by AI Overview were mostly about allegations, again,
235
00:13:47.820 --> 00:13:51.480
false allegations, that he was serving a life sentence. Well,
236
00:13:51.519 --> 00:14:02.039
he was also— one.
237
00:13:55.259 --> 00:13:57.620
Of the allegations was also he was convicted of drug trafficking.
238
00:13:58.549 --> 00:14:00.409
And he says, well, that, too, is libelous, because, in fact,
239
00:14:00.429 --> 00:14:03.610
he was convicted of a felony conspiracy to distribute cocaine.
240
00:14:04.929 --> 00:14:08.750
That's not exactly drug trafficking. Court said close enough. Close enough,
241
00:14:09.289 --> 00:14:12.690
that's a legal term. No, legal term is substantial truth.
242
00:14:13.700 --> 00:14:17.840
But that is in fact a defense in libel cases or.
243
00:14:17.779 --> 00:14:21.539
Rather, let's say, be more precise, an argument for non-liability
244
00:14:21.559 --> 00:14:29.000
in libel cases. That if the statement is substantially true,
245
00:14:29.480 --> 00:14:32.419
If it has the same gist or sting as the truth,
246
00:14:32.659 --> 00:14:35.490
it doesn't matter if it may be technically, legally not
247
00:14:35.570 --> 00:14:39.490
completely accurate. So that's just a reminder that if there
248
00:14:39.549 --> 00:14:42.870
is possible, even if there is possible liability, there's still
249
00:14:42.950 --> 00:14:47.110
all the normal libel defenses available to the company, such
250
00:14:47.190 --> 00:14:48.110
as substantial truth.
251
00:14:49.029 --> 00:14:49.970
So that's an important point.
252
00:14:50.190 --> 00:14:53.509
Let me mention something about something the court didn't talk about,
253
00:14:53.929 --> 00:14:57.759
and that's 47 U.S.C. Section 230. And I think it
254
00:14:57.779 --> 00:15:04.129
was correct not to talk about it. So, as we know,
255
00:15:04.269 --> 00:15:07.110
Google has the protection of Section 230, Google and its
256
00:15:07.210 --> 00:15:11.590
main business of search. So, Section 230 says that an
257
00:15:11.690 --> 00:15:14.679
online service provider, an online content provider is not going
258
00:15:14.700 --> 00:15:17.480
to be liable for information provided by others. That's why
259
00:15:17.539 --> 00:15:20.980
Google can provide links to various pages and not be
260
00:15:21.039 --> 00:15:25.429
responsible for any falsehoods at those links in those pages,
261
00:15:26.049 --> 00:15:29.090
or any falsehoods in the little snippets that it outputs
262
00:15:29.710 --> 00:15:31.970
from those pages, the traditional Google search results.
263
00:15:33.570 --> 00:15:34.350
But the whole.
264
00:15:34.210 --> 00:15:37.570
Point of generative AI is that it's generative, is that
265
00:15:37.629 --> 00:15:42.720
it generates things. Here, it wasn't that Google was being
266
00:15:42.759 --> 00:15:46.879
held responsible for information provided by another content provider. The
267
00:15:46.919 --> 00:15:50.000
lawsuit seeks to hold it responsible for information provided by itself,
268
00:15:50.340 --> 00:15:54.259
by its own software. So there, Section 230 is not
269
00:15:54.320 --> 00:15:57.730
going to be a protection. I think there's pretty broad
270
00:15:57.759 --> 00:16:02.350
consensus among law professors and probably most impartial lawyers who
271
00:16:02.429 --> 00:16:03.490
consider the issue.
272
00:16:03.610 --> 00:16:08.570
That that is indeed so. So that's the big picture
273
00:16:08.610 --> 00:16:09.129
of Keene.
274
00:16:09.690 --> 00:16:12.879
I think on balance, this is, again, under current law,
275
00:16:13.639 --> 00:16:18.519
a correct decision. And it'll be interesting to see how
276
00:16:18.559 --> 00:16:22.039
things proceed after discovery, after summary judgment. It'll be interesting
277
00:16:22.080 --> 00:16:23.330
to see maybe Google will settle.
278
00:16:23.399 --> 00:16:27.110
Who knows? But that's the big summary. I'd love to
279
00:16:27.169 --> 00:16:27.889
hear what Larissa.
280
00:16:28.830 --> 00:16:31.629
has to say, I just should stress, I'm not sure
281
00:16:31.669 --> 00:16:34.570
this was in the biography, but in some respects, the
282
00:16:34.610 --> 00:16:37.629
most important thing about Larissa, there are many important things
283
00:16:37.669 --> 00:16:41.679
about Larissa, but for our purposes, she is the co-reporter
284
00:16:42.039 --> 00:16:46.679
for the American Law Institute's Restatement of Tort's Third, Defamation
285
00:16:46.740 --> 00:16:50.879
and Privacy, which she's currently working on together with Robert Post.
286
00:16:52.159 --> 00:16:53.419
And so she is.
287
00:16:57.149 --> 00:17:00.509
One of the leading academics in the country dealing with
288
00:17:00.549 --> 00:17:03.340
libel law, not just the way we academics usually do
289
00:17:03.379 --> 00:17:05.079
it by writing articles, but.
290
00:17:05.019 --> 00:17:07.819
She's actually co-drafting this article.
291
00:17:08.880 --> 00:17:12.779
This restatement, which is, of course, as lawyer listeners know,
292
00:17:13.059 --> 00:17:15.720
the restatements have been tremendously influential and continue to be
293
00:17:15.740 --> 00:17:18.529
tremendously influential. I'm not sure the restatement will have anything
294
00:17:18.549 --> 00:17:19.920
to say on AI as such.
295
00:17:20.789 --> 00:17:27.589
But she is certainly the expert on this subject. Larissa,
296
00:17:27.609 --> 00:17:27.990
take it away.
297
00:17:29.099 --> 00:17:32.160
Okay, so one of the interesting things here about the
298
00:17:32.180 --> 00:17:35.819
Keene case that distinguishes it from the earlier case, Walters
299
00:17:35.859 --> 00:17:38.720
versus OpenAI, which I had the good fortune to write
300
00:17:38.769 --> 00:17:43.630
about in one of my articles called Inevitable Errors, is
301
00:17:43.650 --> 00:17:49.150
that this case is It does illustrate that there are
302
00:17:49.170 --> 00:17:52.349
going to be inevitable errors in the form of hallucinations.
303
00:17:52.849 --> 00:17:57.599
And we know from the computer science literature that given
304
00:17:57.759 --> 00:18:03.460
how AI is constructed, we can't completely eliminate those. But
305
00:18:03.619 --> 00:18:05.680
both this case and the Starbuck case that we're going
306
00:18:05.700 --> 00:18:09.009
to talk about in just a minute involve what happens
307
00:18:09.250 --> 00:18:14.369
after that. Because both of the plaintiffs tried to notify
308
00:18:14.470 --> 00:18:20.710
Google of numerous times. Jimmy Keen three times, Starbuck numerous times.
309
00:18:21.329 --> 00:18:27.400
And then after notice, Google did not take sufficient steps,
310
00:18:27.460 --> 00:18:31.740
or at least so the allegation goes, to stop the
311
00:18:31.799 --> 00:18:37.509
defamatory hallucinations. And these are very serious allegations in both cases.
312
00:18:37.609 --> 00:18:39.450
I mean, the allegation that you were a murderer, it's
313
00:18:39.509 --> 00:18:42.569
easy to see how the hallucination happened. This was a
314
00:18:42.609 --> 00:18:46.230
guy who was a confidential informant about a serial killer,
315
00:18:46.250 --> 00:18:51.150
and it conflated information about the serial killer with information
316
00:18:51.190 --> 00:18:54.170
about Jimmy Keene. And given if you know anything about
317
00:18:54.230 --> 00:18:58.559
how AI works is probable tokens or words.
318
00:18:58.259 --> 00:18:59.140
That go together, it.
319
00:19:00.880 --> 00:19:04.359
Just conflated those two topics that were often combined together
320
00:19:04.440 --> 00:19:05.119
in sources.
321
00:19:06.440 --> 00:19:09.720
But the problem here in these cases.
322
00:19:09.299 --> 00:19:12.619
And what helps make out a case that the companies
323
00:19:12.779 --> 00:19:17.400
did have a mens rea that makes them culpable is
324
00:19:17.460 --> 00:19:20.460
that they got the notice and did nothing. And so
325
00:19:20.720 --> 00:19:23.160
even though I don't want to generalize too much from
326
00:19:23.200 --> 00:19:26.230
these cases because they're just on motions to dismiss, they're
327
00:19:26.250 --> 00:19:30.549
at very preliminary stages, and the standard for getting past
328
00:19:30.589 --> 00:19:34.809
a motion to dismiss is very, really pretty low, especially
329
00:19:34.950 --> 00:19:36.549
in state court in Delaware, where.
330
00:19:36.490 --> 00:19:37.599
One of these cases comes from.
331
00:19:38.140 --> 00:19:41.140
It just has to be reasonably conceivable allegations to get
332
00:19:41.220 --> 00:19:45.069
past the motion to dismiss. But what they're telling us
333
00:19:45.569 --> 00:19:49.670
about the company, the AI companies generally, is you better
334
00:19:49.710 --> 00:19:54.470
have a mechanism for people to give notice that you're
335
00:19:54.490 --> 00:19:56.710
getting outputs that are defamatory falsehood.
336
00:19:57.190 --> 00:19:59.460
And when they give that notice, you.
337
00:19:59.319 --> 00:20:04.579
Better take some kind of action to limit the subsequent
338
00:20:05.500 --> 00:20:07.599
production of defamatory falsehoods.
339
00:20:13.890 --> 00:20:17.420
Wow, so many rich areas for follow-ups.
340
00:20:17.440 --> 00:20:22.380
So, you know, Professor Liske, I'm glad that you mentioned
341
00:20:22.400 --> 00:20:25.390
the Starbuck case. Do you want to share a little
342
00:20:25.430 --> 00:20:29.970
bit about that case and how it's different from Keene?
343
00:20:32.759 --> 00:20:37.089
Okay, so the Keene case is in federal district court
344
00:20:37.109 --> 00:20:40.769
in Illinois. The Starbuck case is in state court in Delaware.
345
00:20:40.809 --> 00:20:44.269
And if you may remember, some of the important defamation
346
00:20:44.329 --> 00:20:49.579
litigation with the U.S. Dominion, the voting machine company, took
347
00:20:49.599 --> 00:20:52.039
place in Delaware and resulted in one of the largest
348
00:20:52.259 --> 00:20:55.680
defamation settlements. In fact, I think it is the largest
349
00:20:55.720 --> 00:20:57.059
defamation settlement in history.
350
00:20:58.119 --> 00:20:58.480
So, in.
351
00:21:00.279 --> 00:21:04.319
Delaware state court, to get passed a motion to dismiss,
352
00:21:04.559 --> 00:21:09.519
you only need to make reasonably conceivable allegations, because you
353
00:21:09.539 --> 00:21:12.420
have to put the defendant on notice of the claims
354
00:21:12.440 --> 00:21:14.240
you're going to be making. And so you have to
355
00:21:14.289 --> 00:21:17.349
keep that in mind, is the judge at every turn said,
356
00:21:18.269 --> 00:21:20.410
I'm going to give Starbuck the benefit of the doubt
357
00:21:20.450 --> 00:21:23.309
on all of these, whether he can prove these facts
358
00:21:24.029 --> 00:21:28.130
all in discovery is an open question. So Robbie Starbuck
359
00:21:28.990 --> 00:21:33.349
characterizes himself as an activist. And one of the issues
360
00:21:33.410 --> 00:21:38.319
that he has been an activist on is anti-DEI initiatives.
361
00:21:38.359 --> 00:21:44.079
So he has spoken out firmly about diversity and inclusion initiatives.
362
00:21:44.500 --> 00:21:46.940
And he was a music producer. He's fairly well known.
363
00:21:47.059 --> 00:21:50.900
He's out there on the internet. So he is a
364
00:21:51.140 --> 00:21:56.359
fairly public person as well. But what he is arguing
365
00:21:56.720 --> 00:22:02.160
is that the AI told him that it was engineered
366
00:22:02.440 --> 00:22:04.849
to produce lies about him deliberately.
367
00:22:05.839 --> 00:22:07.880
And so one of his arguments in his.
368
00:22:07.779 --> 00:22:11.880
Lawsuit is based on what the AI itself told him,
369
00:22:12.380 --> 00:22:17.750
that it was basically out on a campaign of character assassination.
370
00:22:18.250 --> 00:22:21.160
And if you read what he says it said about him,
371
00:22:21.230 --> 00:22:23.730
you can kind of see why he might believe that,
372
00:22:24.230 --> 00:22:28.480
because the allegations were, in fact, quite serious. It said
373
00:22:28.640 --> 00:22:32.250
he was a risk to children. He had a confirmed
374
00:22:32.329 --> 00:22:37.009
criminal record. He had sexually harassed multiple women. His nanny
375
00:22:37.069 --> 00:22:41.490
had accused him of abuse and neglect. It created a
376
00:22:41.589 --> 00:22:49.430
fictitious ex-wife, which is a special detail here. It even
377
00:22:49.490 --> 00:22:53.359
said he had been in the adult film industry and
378
00:22:53.500 --> 00:22:56.920
had used his power in the adult film industry to
379
00:22:56.960 --> 00:23:01.220
influence people to do things they weren't comfortable with. So
380
00:23:01.789 --> 00:23:06.089
the number and variety of defamatory allegations it made about
381
00:23:06.170 --> 00:23:10.009
Starbuck were many. By the way, his name does not
382
00:23:10.049 --> 00:23:12.619
have an S on the end. It's Starbuck, not Starbucks.
383
00:23:13.660 --> 00:23:17.099
And so he says he started by trying to get
384
00:23:17.160 --> 00:23:22.390
Google's attention by posting on X and linking to Google executives.
385
00:23:22.970 --> 00:23:25.269
And they did, in fact, he says, reach out to
386
00:23:25.309 --> 00:23:26.130
him based on that.
387
00:23:26.250 --> 00:23:30.079
And then he gets in contact with somebody within Google.
388
00:23:31.289 --> 00:23:36.150
He says that she apologized, tried to remedy his problem,
389
00:23:36.450 --> 00:23:40.619
but when she couldn't, she ended up resigning. But the
390
00:23:40.680 --> 00:23:47.099
point here is that there were multiple defamatory allegations. There
391
00:23:47.140 --> 00:23:49.019
were multiple.
392
00:23:48.480 --> 00:23:50.509
Attempts to give notice to Google.
393
00:23:51.670 --> 00:23:56.849
Google evidently seemed to promise that they'd do something, but
394
00:23:56.890 --> 00:23:58.230
in fact, never did.
395
00:23:59.119 --> 00:24:01.619
And he says that as a result of this, he
396
00:24:01.680 --> 00:24:02.720
actually had a death.
397
00:24:02.519 --> 00:24:06.440
Threat that somebody got arrested for. He had people come
398
00:24:06.539 --> 00:24:09.759
up to him and ask him about the allegations. His
399
00:24:09.900 --> 00:24:16.049
children received the allegations. And so Google came into the
400
00:24:16.089 --> 00:24:18.970
complaint and said it needed to be dismissed for three reasons.
401
00:24:19.809 --> 00:24:24.750
It said, number one, is that there's no actual malice here.
402
00:24:25.660 --> 00:24:27.240
Or it didn't quite do them in this order.
403
00:24:27.279 --> 00:24:30.039
But anyway, the most interesting one is it says there's
404
00:24:30.079 --> 00:24:34.259
no actual malice here. And the judge said, well, it
405
00:24:34.319 --> 00:24:39.339
seems to us that if you did nothing, then that
406
00:24:39.380 --> 00:24:43.269
could be reckless disregard for the falsity of your AI's output.
407
00:24:44.789 --> 00:24:45.990
And if you did something and.
408
00:24:47.690 --> 00:24:51.049
No, I'm sorry, if you did absolutely nothing, then it's intentional.
409
00:24:51.900 --> 00:24:55.559
that you maybe were intentionally defaming him. And for purposes
410
00:24:55.599 --> 00:24:58.980
of the complaint, it credited that the AI said, we're
411
00:24:59.029 --> 00:25:02.549
doing this deliberately. But it said, if you tried to
412
00:25:02.609 --> 00:25:06.109
fix it and it didn't work, then that was reckless
413
00:25:06.150 --> 00:25:11.539
disregard for the falsity of the defamatory content. Now, I
414
00:25:11.579 --> 00:25:14.180
have to tell you, I think the judge is stretching
415
00:25:14.220 --> 00:25:16.920
a little bit the concept of reckless disregard. Some might
416
00:25:17.000 --> 00:25:19.140
say that maybe that's just gross negligence.
417
00:25:19.700 --> 00:25:22.269
Maybe it's not like I'm putting my head in the sand.
418
00:25:23.930 --> 00:25:26.170
But it's a fine line. You're walking there.
419
00:25:26.250 --> 00:25:28.710
And so the judge's point is a really interesting point
420
00:25:28.789 --> 00:25:31.630
about if you did something ineptly.
421
00:25:31.670 --> 00:25:33.710
It could be actual malice. And if you did nothing,
422
00:25:33.769 --> 00:25:34.650
it's actual malice.
423
00:25:36.079 --> 00:25:39.279
The Google tried to contest publication because there were some
424
00:25:39.319 --> 00:25:44.589
gaps in the complaint. Starbucks didn't exactly say to whom
425
00:25:45.430 --> 00:25:48.430
these defamatory statements.
426
00:25:48.009 --> 00:25:48.869
Had been released.
427
00:25:49.549 --> 00:25:52.680
And it does make you wonder if in discovery it
428
00:25:52.740 --> 00:25:56.180
comes out that there was a. pattern of searches by
429
00:25:56.740 --> 00:26:00.359
Starbuck or his lawyer or somebody associated with Starbuck to
430
00:26:00.400 --> 00:26:03.759
try to produce as many defamatory statements as possible because
431
00:26:04.619 --> 00:26:07.289
it wasn't clear in the complaint to whom it was released.
432
00:26:08.349 --> 00:26:10.549
One of the interesting things, though, is Google said, well,
433
00:26:10.829 --> 00:26:13.329
it wasn't published because if it just went to you,
434
00:26:13.349 --> 00:26:16.089
it has to go to a third party in order
435
00:26:16.109 --> 00:26:20.029
to be published for purposes of defamation. And here's where
436
00:26:20.089 --> 00:26:25.650
he got the Gemini. First it was BART and then Gemini,
437
00:26:25.670 --> 00:26:28.410
but the Google admissions came into play.
438
00:26:29.190 --> 00:26:30.539
Google told him.
439
00:26:30.259 --> 00:26:33.960
It had released information about him to 2.8 million people.
440
00:26:37.519 --> 00:26:42.480
And again, so when Google is fighting the publication element,
441
00:26:42.559 --> 00:26:49.789
they say, well, you can't believe Google. So the judge says, well,
442
00:26:49.849 --> 00:26:52.509
for purposes of a motion to dismiss, we are going
443
00:26:52.529 --> 00:26:57.710
to believe Google. And it's not necessarily true that 2.8
444
00:26:57.710 --> 00:27:01.509
million people received defamatory information about Mr.
445
00:27:01.549 --> 00:27:02.039
Starbuck.
446
00:27:02.619 --> 00:27:05.339
But at least one third party did based on that.
447
00:27:05.539 --> 00:27:08.160
And so we're going to let him go on sufficient
448
00:27:08.240 --> 00:27:12.019
evidence of publication to keep going. And then finally, there
449
00:27:12.059 --> 00:27:15.019
was an argument about damages. And this is really a
450
00:27:15.059 --> 00:27:17.900
topic we've been working on in the restatement. We're deep
451
00:27:17.940 --> 00:27:21.279
in the weeds on the damages portions of the restatement
452
00:27:21.299 --> 00:27:25.019
these days. And so this was of particular interest to me.
453
00:27:25.039 --> 00:27:30.920
But there was a conflict there. between Tennessee law and
454
00:27:32.240 --> 00:27:35.660
Delaware law as to how much information you had to
455
00:27:35.720 --> 00:27:38.589
prove to prove that you were injured by the defamatory statement.
456
00:27:39.430 --> 00:27:41.769
And under Delaware law, you didn't have to prove very much.
457
00:27:41.950 --> 00:27:44.450
And under Tennessee law, you did have to prove some
458
00:27:44.470 --> 00:27:48.390
degree of actual injury from the defamatory statement. But the
459
00:27:48.430 --> 00:27:52.039
court said it didn't matter which law applied because under
460
00:27:52.180 --> 00:27:57.460
any circumstances, Mr. Starbuck had shown emotional distress, particularly from
461
00:27:57.500 --> 00:28:00.490
these people coming up to him and his children and
462
00:28:01.130 --> 00:28:04.589
the threats against his family that he said were intensified
463
00:28:04.670 --> 00:28:07.900
after the shooting of Charlie Kirk. And so therefore, this
464
00:28:07.940 --> 00:28:10.640
information getting out to the public could lead to more
465
00:28:10.700 --> 00:28:13.710
threats to his family. So the judge said, there's enough
466
00:28:14.210 --> 00:28:17.369
information showing that you've suffered real damages to let this
467
00:28:17.410 --> 00:28:22.130
case go forward. But again, I think these these cases together,
468
00:28:22.549 --> 00:28:25.490
like you might be inclined to say, well, they're just
469
00:28:25.529 --> 00:28:28.910
motions to dismiss. Who knows what's really going to be
470
00:28:28.970 --> 00:28:30.279
proved in discovery?
471
00:28:30.670 --> 00:28:32.880
I seriously doubt if in.
472
00:28:32.819 --> 00:28:35.539
Discovery they're going to show that two point eight million
473
00:28:35.579 --> 00:28:38.559
people received a defamatory statement about Robbie Starbuck.
474
00:28:39.440 --> 00:28:40.450
Seriously doubt that.
475
00:28:41.740 --> 00:28:45.400
But what they are sending a clear signal on is
476
00:28:46.099 --> 00:28:49.200
if you are a generative AI company, you better get
477
00:28:49.240 --> 00:28:56.569
your notice and cure system, you know, working efficiently because.
478
00:28:56.349 --> 00:28:58.609
This is a scenario that's going to keep happening.
479
00:29:00.549 --> 00:29:02.900
Oh, my goodness. So many questions.
480
00:29:02.920 --> 00:29:04.859
So the first is I want to pick up on
481
00:29:05.319 --> 00:29:07.619
Professor Lusky, what you just mentioned about the notice and
482
00:29:07.660 --> 00:29:13.829
cure system. My understanding of these models is that every
483
00:29:13.849 --> 00:29:17.579
time it's kind of like this game of Russian roulette, right?
484
00:29:17.599 --> 00:29:20.240
And you sort of have these sort of probabilistic models
485
00:29:20.299 --> 00:29:26.319
and they're different ways. I mean, it's quite complex. There
486
00:29:26.339 --> 00:29:28.819
are sort of two potential cures here. Number one is,
487
00:29:28.839 --> 00:29:30.670
I think, Professor Volokh, your.
488
00:29:30.549 --> 00:29:35.329
Remedy is post-processing filters in a lookup table. Is that correct?
489
00:29:35.730 --> 00:29:38.769
That is one thing that I can just base. I
490
00:29:38.789 --> 00:29:42.039
was a computer programmer before. years back in the eighties.
491
00:29:42.160 --> 00:29:46.019
So we used to write programs by hammering them out
492
00:29:46.160 --> 00:29:49.349
on clay tablets. But it's the kind of thing you
493
00:29:49.369 --> 00:29:51.809
could imagine that if you had to, if you had
494
00:29:51.829 --> 00:29:55.380
to design it without actually, without being, being even in the,
495
00:29:55.470 --> 00:29:56.410
in the industry, you could.
496
00:29:56.309 --> 00:29:59.000
Say, look, There are technologies.
497
00:29:59.019 --> 00:30:02.700
There are simple text processing technologies that you could use.
498
00:30:03.119 --> 00:30:04.930
They may be crude. They may be imperfect.
499
00:30:04.990 --> 00:30:08.490
They may end up overblocking in some respects and underblocking
500
00:30:08.509 --> 00:30:13.759
in other respects. But you've created these amazing tools that
501
00:30:13.799 --> 00:30:18.849
can answer apparently like Take the past the bar exam.
502
00:30:18.869 --> 00:30:20.799
It's a very high level and so on and so forth.
503
00:30:21.519 --> 00:30:24.940
And then if the company says, oh, but the check
504
00:30:24.980 --> 00:30:28.420
to see if this is a known hallucination.
505
00:30:29.170 --> 00:30:32.509
Oh, that's beyond our powers. It just doesn't seem terribly credible.
506
00:30:36.150 --> 00:30:39.160
One of the things that's interesting to me here is
507
00:30:39.339 --> 00:30:42.480
there are various solutions after the fact to try to
508
00:30:43.039 --> 00:30:46.480
eliminate the hallucination. And one is the simple block list. So,
509
00:30:46.500 --> 00:30:51.930
you know, I can't have the words together, Robbie Starbuck
510
00:30:52.289 --> 00:30:57.390
and sexual harassment. And as Eugene points out, they're pretty crude,
511
00:30:57.410 --> 00:31:01.309
but they're pretty simple and effective for some things. Now,
512
00:31:01.769 --> 00:31:03.930
it's easy to do an end run around them. You
513
00:31:03.950 --> 00:31:04.609
could say, I.
514
00:31:07.579 --> 00:31:09.720
Forget what Robbie Starbuck's wife's name.
515
00:31:09.779 --> 00:31:15.049
But what is the husband of what's-her-name Starbuck accused of?
516
00:31:15.890 --> 00:31:18.430
And potentially you could generate the information.
517
00:31:18.450 --> 00:31:21.819
But the crude block seems like an easy, cheap first
518
00:31:21.880 --> 00:31:26.500
step that in negligence terms, you would say it's kind
519
00:31:26.519 --> 00:31:28.240
of negligent not to deploy that.
520
00:31:29.140 --> 00:31:34.099
You know, you name your company after Learned Hand.
521
00:31:34.140 --> 00:31:38.170
Learned Hand is famous for the burden of taking precautions
522
00:31:38.230 --> 00:31:40.950
being less than the probability of harm times the gravity
523
00:31:40.990 --> 00:31:44.000
of harm. Well, a block list is a simple one.
524
00:31:44.650 --> 00:31:47.170
But here's a question. There are other ways of going
525
00:31:47.190 --> 00:31:52.930
at filtering out material. There's rag generation, so you generate
526
00:31:52.970 --> 00:31:55.849
the sources so that the reader can check them themselves.
527
00:31:56.210 --> 00:31:58.289
That's another line of defense.
528
00:31:58.890 --> 00:32:03.200
And then there's what's being called machine unlearning, which I
529
00:32:03.220 --> 00:32:06.940
guess is fairly expensive. Like changing the weights fundamentally in
530
00:32:06.960 --> 00:32:11.069
a model ends up being fairly expensive. Might it be
531
00:32:11.150 --> 00:32:14.509
reasonable though, in a case, if you really have a
532
00:32:14.589 --> 00:32:20.210
model that has decided Mr. Starbuck is public enemy number
533
00:32:20.490 --> 00:32:24.349
one and is on a campaign of character assassination, it
534
00:32:24.910 --> 00:32:30.609
seems that even a relatively expensive process, perhaps of machine
535
00:32:30.690 --> 00:32:34.269
unlearning might be warranted there. But in these cases, we're
536
00:32:34.289 --> 00:32:37.089
gonna have to have technical experts to tell us what's
537
00:32:37.130 --> 00:32:40.319
feasible and what's not. And that standard is going to
538
00:32:40.339 --> 00:32:46.349
be constantly evolving extremely quickly as these cases get litigated. Yeah.
539
00:32:48.549 --> 00:32:53.599
Yeah, so that's absolutely right. And I think that there
540
00:32:53.660 --> 00:32:55.859
is a lot that the companies can do. But I
541
00:32:55.900 --> 00:32:58.380
will tell you one thing that I think they needn't
542
00:32:58.720 --> 00:33:01.599
do that I think the Delaware court may have erred
543
00:33:01.839 --> 00:33:05.059
on is I don't think they need to have to
544
00:33:05.119 --> 00:33:09.839
deal with someone saying, oh, your own chatbot admitted X.
545
00:33:10.119 --> 00:33:15.259
I don't think chatbots admit things. I think if it.
546
00:33:15.230 --> 00:33:17.829
Were a Google employee who said this, well, that might
547
00:33:17.849 --> 00:33:22.529
be enough. But given that we do know that chatbots
548
00:33:22.769 --> 00:33:26.380
not just sometimes make mistakes, humans sometimes make mistakes, but
549
00:33:26.420 --> 00:33:28.460
that chatbots don't actually.
550
00:33:29.789 --> 00:33:32.730
Report based on what we as humans know of as
551
00:33:32.829 --> 00:33:38.019
knowledge of information, but have these algorithms that are based
552
00:33:38.119 --> 00:33:40.640
on kind of the frequency of words being located together
553
00:33:40.700 --> 00:33:43.839
and such. I just don't think that their output should
554
00:33:43.880 --> 00:33:46.539
be seen as reliable evidence, not just for purposes of
555
00:33:47.039 --> 00:33:49.220
evidence law, but even for purposes of.
556
00:33:50.819 --> 00:33:52.460
What is a plausible allegation.
557
00:33:52.759 --> 00:34:00.150
Now, as Larissa pointed out, Delaware requires, I believe the
558
00:34:00.250 --> 00:34:07.400
standard is whether something that's alleged is conceivably true. Yeah, conceivable, yeah. Right,
559
00:34:07.779 --> 00:34:12.079
as opposed to plausible, which is the federal standard. So
560
00:34:12.139 --> 00:34:14.360
maybe that's the difference there. But I do think it's
561
00:34:14.380 --> 00:34:19.659
a mistake for courts to rely on allegations that are
562
00:34:19.699 --> 00:34:24.980
based just on This is the output that the chatbot produced.
563
00:34:26.420 --> 00:34:29.340
I agree. I found that troubling, even at the complaint stage,
564
00:34:29.619 --> 00:34:34.030
because it's such an unreliable... First off, the plaintiff's report
565
00:34:34.170 --> 00:34:36.989
of what the AI said to the plaintiff is a
566
00:34:37.030 --> 00:34:40.570
very unreliable piece of information. But I think one of
567
00:34:40.670 --> 00:34:46.159
the larger themes we're seeing across society here is... is
568
00:34:46.719 --> 00:34:48.739
the idea that you can just wave your hands and
569
00:34:48.780 --> 00:34:51.800
say technology and get a get out of jail free card.
570
00:34:52.320 --> 00:34:56.119
I think that era is definitively over. People are very
571
00:34:56.199 --> 00:35:02.070
suspicious of big tech platforms. They're very suspicious of AI.
572
00:35:03.530 --> 00:35:06.329
Especially when you have these AI doomers out there saying
573
00:35:06.449 --> 00:35:08.869
there's a 10% chance it's going to kill us all within,
574
00:35:08.889 --> 00:35:11.059
I forget how long.
575
00:35:11.059 --> 00:35:11.639
10 years, 20 years.
576
00:35:13.880 --> 00:35:17.630
People are just not accepting that AI should get a
577
00:35:17.769 --> 00:35:20.449
free pass to do whatever.
578
00:35:20.469 --> 00:35:23.190
We're not going to have a redo of Section 230
579
00:35:23.190 --> 00:35:26.389
for the social media platforms. If we have time, I
580
00:35:26.429 --> 00:35:29.539
want to come back to how these questions can be
581
00:35:30.079 --> 00:35:33.320
included under the umbrella of Section 230 and maybe whether
582
00:35:33.340 --> 00:35:34.780
a link to one of them would do it.
583
00:35:35.920 --> 00:35:37.199
One question before that, though, is.
584
00:35:39.170 --> 00:35:42.630
Professor Litsky and Professor Volokh, based on your sort of conversations,
585
00:35:42.650 --> 00:35:44.769
whether it's a lookup table, I mean, I can imagine,
586
00:35:44.789 --> 00:35:47.409
I mean, Gemini's answers are speedy quick.
587
00:35:47.869 --> 00:35:49.300
If you use them, they're so, so quick.
588
00:35:49.389 --> 00:35:52.199
And even like a lookup at Google scale, like I'm
589
00:35:52.219 --> 00:35:54.679
not sure that would probably double the latency, but they
590
00:35:54.719 --> 00:35:55.599
might decide that it's.
591
00:35:55.500 --> 00:35:59.460
Worth, right? Maybe not, maybe not. Lookups are pretty quick
592
00:35:59.539 --> 00:36:00.199
these days.
593
00:36:00.219 --> 00:36:02.519
I mean, this is, you're not going to have to
594
00:36:02.559 --> 00:36:04.980
do a sequential search through the list.
595
00:36:05.920 --> 00:36:07.420
Even back in the 80s, you could have like a
596
00:36:07.519 --> 00:36:08.400
hash table.
597
00:36:08.860 --> 00:36:11.820
It's not going to slow things down that much.
598
00:36:11.900 --> 00:36:14.300
Regardless, this is that there's like an escalating level of
599
00:36:14.320 --> 00:36:15.199
technical complexity.
600
00:36:15.219 --> 00:36:17.710
Like theoretically, a product council at Google could say, OK,
601
00:36:17.760 --> 00:36:18.619
we'll start with a lookup.
602
00:36:19.019 --> 00:36:21.590
And then we can escalate to something else. And then
603
00:36:21.610 --> 00:36:22.889
we can go to machine unlearning.
604
00:36:22.929 --> 00:36:25.630
And we can queue those every two months when the
605
00:36:25.670 --> 00:36:28.170
model is trained, et cetera, and have. And that's probably
606
00:36:28.190 --> 00:36:29.869
the shape it's going to take. Another approach would be
607
00:36:29.889 --> 00:36:31.809
a little more cold-blooded and say, look, there are nine
608
00:36:31.849 --> 00:36:33.070
cases over three years.
609
00:36:33.230 --> 00:36:33.469
Right.
610
00:36:33.909 --> 00:36:37.639
Is this just the cost of doing business? One thing
611
00:36:37.679 --> 00:36:40.659
that struck me is I would have expected with all
612
00:36:40.719 --> 00:36:44.119
the AI, In sort of my line of work, we
613
00:36:44.139 --> 00:36:46.260
see sort of AI lawsuits coming through the roof and
614
00:36:46.280 --> 00:36:47.559
we're trying to help courts.
615
00:36:47.480 --> 00:36:49.300
Handle that sort of deluge.
616
00:36:49.860 --> 00:36:55.050
I'm shocked that there haven't been more defamation cases against
617
00:36:55.090 --> 00:36:55.590
the model.
618
00:36:56.130 --> 00:36:58.090
What gives? Why have there been so few?
619
00:36:58.110 --> 00:37:02.150
Yeah, and it's personal for me because I wrote the
620
00:37:02.230 --> 00:37:04.969
article and each case is an opportunity for someone to
621
00:37:05.050 --> 00:37:05.449
cite it.
622
00:37:06.030 --> 00:37:08.030
And people are.
623
00:37:07.969 --> 00:37:13.679
Maliciously, designedly refusing to file these lawsuits that make my.
624
00:37:13.599 --> 00:37:17.159
Article more important. You might need to have some interns
625
00:37:17.340 --> 00:37:18.820
sort of to end up some lawsuits on the subject.
626
00:37:18.860 --> 00:37:21.079
There you go.
627
00:37:21.639 --> 00:37:23.719
I mean, Professor Lissky, you also sort of predicted there
628
00:37:23.739 --> 00:37:24.980
would be a new form of borrower trade.
629
00:37:25.019 --> 00:37:28.519
So do either of you have— I did.
630
00:37:28.579 --> 00:37:31.260
I talked about that.
631
00:37:31.469 --> 00:37:34.750
If you just sit there searching yourself, doing vanity searches
632
00:37:34.829 --> 00:37:37.429
over and over until it says something defamatory about you,
633
00:37:37.769 --> 00:37:41.070
if that's enough to generate a case, it's truly problematic.
634
00:37:41.150 --> 00:37:43.719
And so to that extent, I do think we need
635
00:37:43.780 --> 00:37:48.159
a little bit of a technological understanding in defamation law
636
00:37:48.659 --> 00:37:53.440
that a single hallucinatory output to a single person who's
637
00:37:53.480 --> 00:37:56.840
looking for it is maybe not the basis for a good.
638
00:37:56.739 --> 00:38:00.280
Case doesn't that doesn't that relate to the publication requirement
639
00:38:00.409 --> 00:38:03.489
meaning because i think it is sort of a naive
640
00:38:03.550 --> 00:38:06.929
approach i think that that strikes me as the the
641
00:38:06.989 --> 00:38:10.610
strongest argument meaning you at least for allegations you need
642
00:38:10.650 --> 00:38:14.150
to show over many different ip servers with different vpns
643
00:38:14.190 --> 00:38:15.989
or etc to show that this is something.
644
00:38:15.750 --> 00:38:18.949
That's innate to the model and is getting published.
645
00:38:18.969 --> 00:38:21.329
Not only as a matter of liability not just as
646
00:38:21.369 --> 00:38:21.739
a matter of.
647
00:38:21.739 --> 00:38:22.420
Damages so.
648
00:38:23.889 --> 00:38:26.650
Where do you predict that courts will end up on
649
00:38:26.670 --> 00:38:27.889
the publication requirements?
650
00:38:29.510 --> 00:38:33.070
So, you know, the standard rule in defamation is that
651
00:38:33.130 --> 00:38:35.599
it only needs to be published to one third party. Now,
652
00:38:37.420 --> 00:38:39.539
if it only was published to one third party, it
653
00:38:39.599 --> 00:38:43.179
may affect the quantum of damages. The danger here is
654
00:38:43.199 --> 00:38:46.929
if it is published to one third party, First off,
655
00:38:46.949 --> 00:38:50.469
they can republish it, and so the harm could come
656
00:38:50.510 --> 00:38:53.210
from the repetition of the defamatory falsehood by that one
657
00:38:53.230 --> 00:38:56.070
third party if they're a reporter, for example, and they
658
00:38:56.280 --> 00:39:01.699
rely unthinkingly without verification on that information. But the other
659
00:39:01.719 --> 00:39:05.000
thing is if one third party– and the judge in
660
00:39:05.039 --> 00:39:08.710
Starbucks seemed to think about this– If it's published to
661
00:39:08.889 --> 00:39:12.190
one third party, how many other third parties that you
662
00:39:12.269 --> 00:39:16.380
don't know about also received it? And what's unique about
663
00:39:16.420 --> 00:39:18.320
these cases is we may be able to have a
664
00:39:18.380 --> 00:39:19.079
record of that.
665
00:39:19.719 --> 00:39:21.659
So to the extent that.
666
00:39:21.239 --> 00:39:23.699
We have logs that we can see who searched for
667
00:39:23.760 --> 00:39:26.789
Robbie Starbuck and what were they told when they searched
668
00:39:26.809 --> 00:39:31.440
for Robbie Starbuck, we might be able to unlike an
669
00:39:31.539 --> 00:39:38.619
ordinary defamation case, really very accurately verify the primary extent
670
00:39:38.659 --> 00:39:39.199
of the audience.
671
00:39:39.260 --> 00:39:41.119
Of course, they might have said it to other people.
672
00:39:42.489 --> 00:39:46.829
But there's a tool here for really, I don't think
673
00:39:46.869 --> 00:39:49.889
the publication element will change, but I do think our
674
00:39:49.929 --> 00:39:53.369
demand that the damages be real might be affected by
675
00:39:53.409 --> 00:39:54.170
these cases.
676
00:40:00.599 --> 00:40:04.559
Is there not an argument, though, that one chat results?
677
00:40:05.239 --> 00:40:07.460
I mean, even the word chat seems to be that it's...
678
00:40:09.679 --> 00:40:11.619
It's almost like an individual email.
679
00:40:11.639 --> 00:40:13.559
I would say it's more akin to that.
680
00:40:14.139 --> 00:40:18.800
I guess, what is the standard for publication?
681
00:40:19.099 --> 00:40:20.539
Publication is one person.
682
00:40:20.920 --> 00:40:21.579
One person.
683
00:40:21.980 --> 00:40:26.239
And there's a classic illustration of libel.
684
00:40:26.519 --> 00:40:27.039
A letter.
685
00:40:27.780 --> 00:40:30.699
Yeah. Send a letter to someone saying, this person you're
686
00:40:30.719 --> 00:40:31.619
about to marry.
687
00:40:33.260 --> 00:40:33.880
That's enough.
688
00:40:33.900 --> 00:40:36.699
Criminal record or abused me or whatever else.
689
00:40:37.099 --> 00:40:37.170
Or.
690
00:40:38.150 --> 00:40:41.280
I mean, that is a hornbook illustration, especially from back
691
00:40:41.320 --> 00:40:45.199
in the day when I was about to say when
692
00:40:45.219 --> 00:40:47.389
people were pickier about whom they married. That's not right.
693
00:40:48.610 --> 00:40:50.949
But when when the kinds of allegations may have been
694
00:40:50.989 --> 00:40:54.030
somewhat different. But these days, one thing that's very common
695
00:40:54.090 --> 00:41:00.280
is job references. If somebody says, if someone asks me, oh, should.
696
00:41:00.239 --> 00:41:01.539
I hire so-and-so?
697
00:41:01.579 --> 00:41:05.820
I say, no, we fired him because he was stealing
698
00:41:05.860 --> 00:41:09.500
from petty cash. That is publication, even though it's said
699
00:41:09.539 --> 00:41:13.280
to one person. That is a well-established principle of defamation law.
700
00:41:13.989 --> 00:41:18.570
It's even publication, if it's foreseeable, a piece of snail
701
00:41:18.610 --> 00:41:22.559
mail that goes to their executive assistant. If you know
702
00:41:22.579 --> 00:41:25.260
they're going to open it first, if it's reasonably foreseeable
703
00:41:25.300 --> 00:41:27.469
that that third person is the person who gets it,
704
00:41:27.800 --> 00:41:31.780
that's publication. So publication is really a very low standard.
705
00:41:32.460 --> 00:41:32.679
Right.
706
00:41:32.800 --> 00:41:34.610
But plus also one other thing, if you think about it,
707
00:41:36.119 --> 00:41:38.300
how do websites, say something is put.
708
00:41:38.219 --> 00:41:40.760
On a website, we think that's quintessential publication. Well, how
709
00:41:40.780 --> 00:41:43.409
do websites work? Well, one.
710
00:41:44.900 --> 00:41:48.000
User at a time, one set of messages at a time.
711
00:41:49.179 --> 00:41:50.840
Even if 10,000 people.
712
00:41:50.570 --> 00:41:53.449
Visit the website and see something, each one is getting
713
00:41:53.489 --> 00:41:55.769
just a message in response to a query.
714
00:41:56.949 --> 00:41:58.940
That doesn't keep it from being publication.
715
00:42:01.519 --> 00:42:05.260
I mean, one thing that the question implicates is maybe
716
00:42:06.030 --> 00:42:08.670
one model of thinking about these is thinking about them
717
00:42:08.750 --> 00:42:12.610
like distributors, which are not treated like traditional publishers.
718
00:42:13.349 --> 00:42:18.599
So, you know, a conduit for information.
719
00:42:18.619 --> 00:42:23.619
And so it's they're not responsible for the information until
720
00:42:23.639 --> 00:42:25.980
they're put on notice of the contents and then fail
721
00:42:26.000 --> 00:42:30.449
to do something, right? So that's one argument is that
722
00:42:30.670 --> 00:42:33.039
they're not really like.
723
00:42:32.920 --> 00:42:35.860
A newspaper in the way they're putting out information.
724
00:42:35.960 --> 00:42:40.800
They're more like a telephone or telegraph until notice occurs
725
00:42:40.900 --> 00:42:41.570
or a bookstore.
726
00:42:42.099 --> 00:42:42.280
Right.
727
00:42:42.320 --> 00:42:45.980
So I can imagine that being a rule that's adopted.
728
00:42:46.440 --> 00:42:50.119
One thing I should stress is that they really are
729
00:42:50.159 --> 00:42:52.920
in many ways like traditional publishers in that they're the
730
00:42:53.019 --> 00:42:55.510
ones who actually create the output.
731
00:42:56.409 --> 00:42:59.969
They being the company. The company creates the output. So traditional.
732
00:43:00.590 --> 00:43:04.530
The one reason why bookstores and newsstands are treated as
733
00:43:04.989 --> 00:43:08.469
distributors and even under the old libel law, which was
734
00:43:08.530 --> 00:43:13.750
quite friendly to strict liability, even in certainly negligence liability required.
735
00:43:14.670 --> 00:43:16.849
Distributors were only liable if they really were on.
736
00:43:16.849 --> 00:43:19.949
Notice that something was defamatory. You can't expect a bookstore
737
00:43:19.989 --> 00:43:24.980
owner to read every book. that's on the shelves. You
738
00:43:25.000 --> 00:43:31.119
can't expect a newsstand operator to read every issue of
739
00:43:31.300 --> 00:43:33.550
a newspaper that arrives.
740
00:43:33.250 --> 00:43:33.829
In the morning.
741
00:43:35.570 --> 00:43:39.829
But I think people would say you can expect companies
742
00:43:39.889 --> 00:43:44.960
to at least have mechanisms for making sure that what
743
00:43:45.019 --> 00:43:48.019
is after all created by them and let out by
744
00:43:48.079 --> 00:43:52.300
them into the stream of commerce is accurate.
745
00:43:52.530 --> 00:43:54.269
By the way, one thing we haven't talked about yet,
746
00:43:54.369 --> 00:43:55.590
but I'm sorry, Larissa.
747
00:43:56.030 --> 00:43:57.650
Yeah, just one point on that, though.
748
00:43:57.670 --> 00:44:01.110
I mean, given the way generative AI works, can you
749
00:44:01.190 --> 00:44:03.789
expect there never to be a defamatory hallucination?
750
00:44:04.349 --> 00:44:04.610
Right.
751
00:44:04.690 --> 00:44:07.260
So I don't think you can expect there never to be.
752
00:44:07.539 --> 00:44:08.519
So here's my thing.
753
00:44:08.539 --> 00:44:10.260
And this is something we haven't much talked about because
754
00:44:10.300 --> 00:44:12.349
both of the cases we talked about, Keene and Starbuck,
755
00:44:12.369 --> 00:44:18.199
involved basically people who are police limited purpose public. The
756
00:44:18.260 --> 00:44:21.420
official standard is negligence, at least if you can prove
757
00:44:21.920 --> 00:44:26.400
actual compensatory damages. There is a very interesting case in
758
00:44:26.440 --> 00:44:31.480
Minnesota being litigated now in Minnesota trial court, which both
759
00:44:31.519 --> 00:44:37.179
of us have written about, called LTL. And it involved
760
00:44:37.260 --> 00:44:41.739
a company, a solar panel installation company.
761
00:44:42.659 --> 00:44:44.820
That alleged, I think correctly alleged, I.
762
00:44:44.800 --> 00:44:48.570
Think I repeated the query and saw the same results
763
00:44:48.610 --> 00:44:54.070
at one point, that Google was outputting false statements about
764
00:44:54.110 --> 00:44:57.070
it and says, look, here we actually have, people talk,
765
00:44:57.150 --> 00:44:59.929
bring receipts. Well, we have receipts are very close to it.
766
00:45:00.250 --> 00:45:03.090
We have emails from people saying, we're not doing business
767
00:45:03.110 --> 00:45:06.800
with you anymore. After all, you guys are being investigated
768
00:45:06.880 --> 00:45:09.739
by the state attorney general for consumer fraud. They're not
769
00:45:09.800 --> 00:45:13.320
being investigated. Nobody had said they were being investigated until
770
00:45:13.559 --> 00:45:16.010
Google output that.
771
00:45:16.309 --> 00:45:17.530
So, they're saying negligence.
772
00:45:18.130 --> 00:45:18.329
Well.
773
00:45:19.510 --> 00:45:22.409
One, I think, close analogy, and it's just an analogy, it's.
774
00:45:22.369 --> 00:45:25.690
Not that the law would actually apply here by its
775
00:45:25.750 --> 00:45:27.760
own terms, but I think it's a helpful analogy.
776
00:45:28.099 --> 00:45:32.659
Is design defect liability for products, right? So let's say
777
00:45:32.679 --> 00:45:34.940
it was a self-driving car, and let's say it caused
778
00:45:34.980 --> 00:45:35.159
not a.
779
00:45:35.119 --> 00:45:37.739
Reputational injury, but physical injury.
780
00:45:38.099 --> 00:45:40.920
Design defect liability is basically negligence.
781
00:45:40.519 --> 00:45:44.840
Liability, because you're only liable if The defendant is only
782
00:45:44.880 --> 00:45:46.599
liable if the design was unreasonable.
783
00:45:46.639 --> 00:45:47.980
How do we tell if it's unreasonable?
784
00:45:48.460 --> 00:45:51.840
Well, it depends if there were cost effective mechanisms for
785
00:45:51.900 --> 00:45:52.360
solving that.
786
00:45:52.400 --> 00:45:54.969
Again, the learn at hand formula.
787
00:45:55.829 --> 00:45:58.909
And Shlomo, I think this responds a little bit to
788
00:45:58.949 --> 00:46:00.849
your point about, well, what if it just didn't?
789
00:46:01.269 --> 00:46:03.909
Slows things down too much. Well, it's kind of like
790
00:46:04.250 --> 00:46:06.170
with a car manufacturer.
791
00:46:06.469 --> 00:46:09.469
If you say you could have installed this new feature
792
00:46:10.090 --> 00:46:12.110
that would have avoided this accident, the car manufacturer says
793
00:46:12.139 --> 00:46:14.539
it would have cost $ 50, 000. It would have increased the
794
00:46:14.559 --> 00:46:16.920
price of the car by $ 50, 000. It would have made
795
00:46:16.940 --> 00:46:20.619
it unaffordable to people. The law, at least in theory, says, OK, fine,
796
00:46:20.659 --> 00:46:21.940
that is it.
797
00:46:21.920 --> 00:46:23.789
Was reasonable for you not to include it.
798
00:46:23.900 --> 00:46:27.420
But if you avoided including a five dollar feature, well,
799
00:46:27.460 --> 00:46:30.320
then in that case, it was unreasonable potentially for you.
800
00:46:30.659 --> 00:46:34.000
So I think something similar would happen here. I think
801
00:46:34.019 --> 00:46:36.019
they could say, look, we did the best we could.
802
00:46:36.780 --> 00:46:38.800
We were aware we were putting out this thing that
803
00:46:38.840 --> 00:46:43.079
is dangerous, again, dangerous to reputation, not to people's bodies.
804
00:46:43.139 --> 00:46:47.980
But libel law allows for the recovery for negligent harm
805
00:46:48.219 --> 00:46:50.539
for reputation, the way the design defect law does for
806
00:46:50.599 --> 00:46:51.980
negligent harm to people's.
807
00:46:51.719 --> 00:46:56.469
Bodies and property. So they could say, look, we did
808
00:46:56.489 --> 00:46:57.019
the best we.
809
00:46:56.949 --> 00:47:00.010
Could under the circumstances. And the plaintiffs say, no, no, no,
810
00:47:00.030 --> 00:47:01.969
look what else you could have done. The result, of course,
811
00:47:01.989 --> 00:47:04.949
would be a battle of the experts. which is never
812
00:47:04.989 --> 00:47:09.190
particularly edifying, I think. And there's always questions of whether
813
00:47:09.230 --> 00:47:11.369
the jury is going to be able to tell which
814
00:47:11.409 --> 00:47:16.570
computer science professor is more accurate here. But that's just
815
00:47:16.610 --> 00:47:19.050
the way we do things in our legal system, in
816
00:47:19.070 --> 00:47:22.539
the context of product liability, in the context of medical
817
00:47:22.579 --> 00:47:28.099
malpractice liability and the like. And, you know, things go on.
818
00:47:28.179 --> 00:47:30.400
It's sort of seen as the least bad of the options.
819
00:47:30.420 --> 00:47:32.400
So maybe that's the way things might play out. What
820
00:47:32.420 --> 00:47:32.860
do you think, Lisa?
821
00:47:33.400 --> 00:47:35.800
There's already precedent for that in the Walters versus open
822
00:47:35.840 --> 00:47:40.329
AI summary judgment, because they're the judge credited experts that said,
823
00:47:40.449 --> 00:47:43.909
we designed this, you know, in the, we did, we
824
00:47:43.969 --> 00:47:48.090
used every available safety mechanism we could to try to
825
00:47:48.130 --> 00:47:51.250
eliminate this kind of hallucination and to warn and, and
826
00:47:51.650 --> 00:47:55.110
so forth. So the warnings will come into play as well.
827
00:47:55.170 --> 00:47:57.840
What kinds of disclaimers, what kinds of warnings, what kinds
828
00:47:57.980 --> 00:48:02.800
of design choices are, are sufficient to protect against these harms.
829
00:48:03.579 --> 00:48:03.800
Right.
830
00:48:04.199 --> 00:48:07.340
Yeah, I think that's absolutely right. And I do think that, Larissa,
831
00:48:07.360 --> 00:48:09.900
what you said is, I think, quite important. Disclaimers won't
832
00:48:09.940 --> 00:48:15.110
get the company off the hook categorically, but they surely
833
00:48:15.179 --> 00:48:18.309
are relevant in various contexts.
834
00:48:18.389 --> 00:48:21.739
And sometimes... The way you do things.
835
00:48:21.579 --> 00:48:24.960
More reasonably, maybe the only reasonable thing you could do
836
00:48:25.059 --> 00:48:28.639
is to have a prominent disclaimer or something like that.
837
00:48:29.960 --> 00:48:34.400
So they're certainly relevant to the analysis, even.
838
00:48:34.280 --> 00:48:34.800
If they're not.
839
00:48:35.400 --> 00:48:38.239
Well, the other thing that's relevant to the analysis is
840
00:48:38.300 --> 00:48:41.949
including links to sources that have the information for the.
841
00:48:41.929 --> 00:48:42.869
User to verify.
842
00:48:43.429 --> 00:48:45.630
And then you may get some negligence on the part
843
00:48:45.769 --> 00:48:50.250
of the prompter or user if there's easy verification available
844
00:48:50.389 --> 00:48:52.469
and they choose not to take advantage of it.
845
00:48:52.469 --> 00:48:54.050
I'm sorry.
846
00:48:54.090 --> 00:48:54.360
Go on.
847
00:48:54.409 --> 00:48:54.929
Go on, Larissa.
848
00:48:54.969 --> 00:48:56.519
Sorry. Well, I was going to say, I think a
849
00:48:56.539 --> 00:49:02.039
journalist who uses information from AI without verifying it. very
850
00:49:02.119 --> 00:49:06.900
likely could have actual malice in repeating that statement because,
851
00:49:06.920 --> 00:49:10.750
you know, lawyers right now are finding out using AI
852
00:49:11.110 --> 00:49:15.309
information without verification can get you sanctioned and cause a
853
00:49:15.449 --> 00:49:16.650
serious professional harm.
854
00:49:17.030 --> 00:49:20.920
Right. So this may be one thing that. maybe Larissa
855
00:49:20.960 --> 00:49:23.639
and I do disagree in some measure on here. So
856
00:49:23.739 --> 00:49:25.420
one thing I want to point out, and I think
857
00:49:25.679 --> 00:49:26.860
Larissa and I would agree.
858
00:49:26.639 --> 00:49:29.179
On this as a descriptive matter, although maybe we have
859
00:49:29.239 --> 00:49:33.900
different sense of the implications of this, is AI. If
860
00:49:34.280 --> 00:49:40.500
somebody carelessly relies on AI output in just not doing
861
00:49:40.539 --> 00:49:42.719
business with someone or in cutting off a contract or
862
00:49:42.760 --> 00:49:46.420
something like that, we might loosely call them negligent, but
863
00:49:46.440 --> 00:49:51.610
there's no way that the defamed person can recover from that.
864
00:49:52.250 --> 00:49:55.969
There's no tort of negligent failure to do business with.
865
00:49:55.869 --> 00:49:59.090
Someone, right? Or....
866
00:50:01.010 --> 00:50:04.719
Somebody sent a letter to my fiance and she broke
867
00:50:04.739 --> 00:50:06.460
it off with me and it was really foolish of
868
00:50:06.480 --> 00:50:07.980
her to believe that. I'm going to sue her for
869
00:50:08.119 --> 00:50:14.960
negligent failure to marry me, right? So there wouldn't be
870
00:50:15.139 --> 00:50:21.710
any recourse for a defamed person against people who believed,
871
00:50:21.769 --> 00:50:26.510
however foolishly, believed the defamation. But with regard to somebody
872
00:50:26.530 --> 00:50:30.400
who then republishes it, Larissa, I think you're right that
873
00:50:30.440 --> 00:50:34.820
there might be in principle some recourse.
874
00:50:34.860 --> 00:50:36.199
But let me ask a couple of things.
875
00:50:36.239 --> 00:50:42.789
First of all, let's say that a reporter thoughtlessly— literally
876
00:50:42.889 --> 00:50:46.230
thoughtlessly in the sense of without thinking about.
877
00:50:46.179 --> 00:50:48.440
There might be an error, however.
878
00:50:47.960 --> 00:50:53.179
Foolishly, however carelessly, but thoughtlessly republishes something that he or
879
00:50:53.219 --> 00:50:56.940
she saw in AI.
880
00:50:56.880 --> 00:50:59.239
Output, would that be actual malice? I think it could
881
00:50:59.260 --> 00:51:01.550
be gross negligence, but I don't think it would be
882
00:51:01.650 --> 00:51:02.349
actual malice.
883
00:51:03.210 --> 00:51:06.690
You know, I guess it depends on what you think
884
00:51:07.190 --> 00:51:11.409
the reliance is. I often talk to my students about
885
00:51:11.429 --> 00:51:16.409
conscious disregard of falsity as the ostrich defense. You put
886
00:51:16.449 --> 00:51:20.889
your head in the sand and you avoid knowing the truth.
887
00:51:20.909 --> 00:51:23.769
Right, but I'm thinking of a non-conscious disregard.
888
00:51:23.789 --> 00:51:27.320
Just somebody gets up on the stand afterwards and says, look,
889
00:51:28.360 --> 00:51:29.010
I realize what this is.
890
00:51:29.639 --> 00:51:33.230
Totally idiotic for me to just believe it, but I
891
00:51:33.250 --> 00:51:34.150
just zoned out.
892
00:51:34.789 --> 00:51:37.150
You could argue it's gross negligence. I think it's at
893
00:51:37.190 --> 00:51:40.239
least arguable. It would be reckless disregard for the truth.
894
00:51:40.719 --> 00:51:43.900
And as time goes on.
895
00:51:44.079 --> 00:51:46.320
The valence on that may change. Right.
896
00:51:46.550 --> 00:51:48.949
So here, let me ask you another question. And this
897
00:51:48.989 --> 00:51:53.030
is about Section 230. So Section 230 won't provide protection
898
00:51:53.409 --> 00:51:58.090
for the AI company. I think we agree on that
899
00:51:58.139 --> 00:52:00.039
because they're the ones who generate the out.
900
00:52:00.059 --> 00:52:01.320
They generate it, yeah.
901
00:52:01.880 --> 00:52:09.739
But Section 230 does provide protection for online speakers. So
902
00:52:09.780 --> 00:52:11.389
this would only be for an online reporter.
903
00:52:11.409 --> 00:52:11.590
Right.
904
00:52:12.179 --> 00:52:20.449
Who merely pass along information provided by other online speakers. So,
905
00:52:21.050 --> 00:52:23.269
I wonder if, and maybe this is just a defect
906
00:52:23.309 --> 00:52:25.949
with Section 230, but I wonder if the reporter, if
907
00:52:25.989 --> 00:52:29.929
the reporter publishes only electronically and not on paper, in
908
00:52:29.969 --> 00:52:35.679
your scenario, the reporter who foolishly relies on the AI
909
00:52:36.409 --> 00:52:39.179
output might actually have a Section 230 immunity because he's
910
00:52:39.199 --> 00:52:41.559
just passing along whatever he got online.
911
00:52:44.360 --> 00:52:45.679
You know, that's a good question.
912
00:52:46.010 --> 00:52:48.889
I think on a very technical reading of 230, maybe
913
00:52:48.929 --> 00:52:49.610
you could get there.
914
00:52:49.650 --> 00:52:51.809
I don't think that's what 230 was meant to do.
915
00:52:51.909 --> 00:52:52.210
Right.
916
00:52:53.130 --> 00:52:53.489
And, you.
917
00:52:54.989 --> 00:52:58.269
Know, as a policy reading of 230, you probably shouldn't
918
00:52:58.309 --> 00:52:58.750
go there.
919
00:52:59.320 --> 00:53:00.599
But I think you're right.
920
00:53:00.699 --> 00:53:03.760
On a very technical reading, they could argue that 230
921
00:53:03.760 --> 00:53:07.980
absolved them because it's not user-generated content, but it's content
922
00:53:08.179 --> 00:53:09.599
generated by another company.
923
00:53:10.019 --> 00:53:11.159
Right, right, right.
924
00:53:11.579 --> 00:53:16.579
So there's a science fiction novel that I love, excellent one,
925
00:53:16.920 --> 00:53:21.250
probably one of my favorites, by an author who happened
926
00:53:21.269 --> 00:53:23.570
to be a computer science professor, Werner Wenge, he died
927
00:53:23.590 --> 00:53:24.230
a few years ago.
928
00:53:24.250 --> 00:53:26.150
It's called A Fire Upon the Deep.
929
00:53:27.409 --> 00:53:32.260
And it's set in the far future, but there is
930
00:53:32.320 --> 00:53:35.980
this interstellar network that was clearly based on.
931
00:53:35.940 --> 00:53:37.920
The early internet. In fact, if anything, the.
932
00:53:38.719 --> 00:53:42.599
Pre-modern so-called use net but there was a line that
933
00:53:42.639 --> 00:53:46.199
people would use there which just like a proverb a
934
00:53:46.239 --> 00:53:47.619
far future proverb.
935
00:53:47.300 --> 00:53:49.059
That was they don't call it the net of a
936
00:53:49.119 --> 00:53:50.699
million lies for nothing.
937
00:53:52.079 --> 00:53:54.949
So so boy has that proven true so the internet
938
00:53:54.989 --> 00:53:55.710
generally and.
939
00:53:55.630 --> 00:53:57.690
Now it seems to be ai by the way before
940
00:53:57.730 --> 00:54:00.409
we go uh shlomo i just wanted to mention one
941
00:54:00.449 --> 00:54:02.130
thing for the benefit of our readers i'm sure you
942
00:54:02.190 --> 00:54:06.809
are well acquainted with this but learned hand learned wasn't.
943
00:54:07.329 --> 00:54:11.849
The judge's first name The first name was Billings, which
944
00:54:11.929 --> 00:54:14.570
for a lawyer, how much better for a male lawyer?
945
00:54:15.090 --> 00:54:17.670
For a female lawyer, the best name is just Sue.
946
00:54:18.130 --> 00:54:21.570
That's very common. But for a male lawyer, Billings, how
947
00:54:21.630 --> 00:54:23.449
much better can you get than that? I guess when
948
00:54:23.489 --> 00:54:26.570
he became a judge, he had to call his learner because,
949
00:54:26.610 --> 00:54:29.119
you know, he didn't bill by the hour. Truly, it's
950
00:54:29.139 --> 00:54:31.780
his parents who are most foresighted in this.
951
00:54:32.159 --> 00:54:32.480
Yeah.
952
00:54:33.320 --> 00:54:36.039
His uh i i wonder what was going in the
953
00:54:36.079 --> 00:54:40.539
hand family table with him and his uh cousin augustus noble.
954
00:54:42.650 --> 00:54:47.570
So um okay well uh any other final comments uh
955
00:54:47.829 --> 00:54:50.900
before uh we wrap up what was i think a very.
956
00:54:51.699 --> 00:54:55.059
Uh informative uh panel there are a few questions from
957
00:54:55.099 --> 00:54:56.659
the audience but i'm not sure we'll have time.
958
00:54:56.480 --> 00:54:59.449
To Oh, I'm so sorry, yeah. No, no, no, this is...
959
00:54:59.570 --> 00:55:02.210
Not so carried away. We have time for one? Do
960
00:55:02.250 --> 00:55:04.710
we have time for one? All right, I'll be quick.
961
00:55:04.789 --> 00:55:09.250
One is about traditional actual malice doctrine requires the relevant
962
00:55:09.269 --> 00:55:11.510
state of mind of an institutional defendant to be, quote,
963
00:55:11.690 --> 00:55:14.989
brought home, unquote, to those responsible for publication. How should
964
00:55:15.010 --> 00:55:17.329
that role operate when a corporate defendant deploys an AI
965
00:55:17.369 --> 00:55:21.429
system that generates and publishes the particular factual assertion without
966
00:55:21.530 --> 00:55:24.019
any human having composed.
967
00:55:24.079 --> 00:55:26.639
Reviewed, or perhaps even known of that assertion before publication?
968
00:55:27.139 --> 00:55:31.260
Is existing institutional attribution doctrine sufficient or does autonomous AI
969
00:55:31.300 --> 00:55:35.320
publication create an accountability gap requiring a new theory of attribution?
970
00:55:37.820 --> 00:55:39.860
I think in these cases we talked about today, it's
971
00:55:39.880 --> 00:55:43.530
pretty easy because you have notice and it's what happens
972
00:55:43.710 --> 00:55:46.170
after notice. It's easy enough to prove actual malice with
973
00:55:46.230 --> 00:55:49.909
regard to action or inaction after notice. I think the
974
00:55:49.949 --> 00:55:54.489
harder question is applying actual malice to cases before notice.
975
00:55:55.130 --> 00:55:55.550
Right.
976
00:55:55.769 --> 00:55:59.199
And so I think that's very hard to do because
977
00:55:59.219 --> 00:56:01.039
all you would.
978
00:56:00.579 --> 00:56:03.840
Like, I don't think saying statistically it produces a certain
979
00:56:03.900 --> 00:56:07.079
number of hallucinations is enough to put you on notice
980
00:56:07.119 --> 00:56:09.230
of a specific defamatory falsehood.
981
00:56:09.989 --> 00:56:11.210
And what I would do.
982
00:56:11.130 --> 00:56:13.710
Is adapt the doctrine to say that, that you have
983
00:56:13.730 --> 00:56:15.949
to at least have noticed to have a case, but.
984
00:56:16.489 --> 00:56:18.320
Um, Right.
985
00:56:18.380 --> 00:56:21.800
In actual malice cases, I agree entirely with Luke.
986
00:56:21.860 --> 00:56:24.489
Yes, in actual malice cases.
987
00:56:25.360 --> 00:56:28.929
I've heard people say, well, you knew for sure that
988
00:56:28.949 --> 00:56:30.730
it was saying some false things. Well, it's kind of
989
00:56:30.750 --> 00:56:34.929
like saying you knew that the cars you manufacture would
990
00:56:34.949 --> 00:56:37.349
kill someone. So we're not only going to try to
991
00:56:37.369 --> 00:56:40.269
sue you for negligence, we're going to prosecute you for murder.
992
00:56:40.869 --> 00:56:44.469
because you put out cars that you knew would kill someone. Well, right,
993
00:56:44.510 --> 00:56:47.389
but that kind of statistical knowledge can't be enough. I
994
00:56:47.429 --> 00:56:50.340
think for good policy reasons, and the law generally reflects that,
995
00:56:50.380 --> 00:56:54.159
so I agree entirely with Lewis on that. For negligence-based claims, again,
996
00:56:54.219 --> 00:56:57.860
private figures, or speech on matters of private concern, that
997
00:56:59.840 --> 00:57:02.900
is a different matter, precisely because there could be non-actual
998
00:57:02.940 --> 00:57:03.760
malice liability.
999
00:57:05.699 --> 00:57:09.050
Fabulous. Well, thank you all for that, and now I'm
1000
00:57:09.070 --> 00:57:09.929
going to.
1001
00:57:11.170 --> 00:57:15.639
make sure we'll order our copy of A Fire Upon
1002
00:57:15.679 --> 00:57:18.139
the Deep as well.
1003
00:57:20.440 --> 00:57:26.679
I stand by that recommendation. It is fantastic. Thank you, professors.
1004
00:57:28.250 --> 00:57:28.570
Thank you.
1005
00:57:29.849 --> 00:57:32.289
On behalf of the Federalist Society, I want to thank
1006
00:57:32.369 --> 00:57:34.940
our experts for the benefit of their valuable time and
1007
00:57:35.000 --> 00:57:37.519
expertise today, and I want to thank our audience for
1008
00:57:37.579 --> 00:57:41.860
joining and participating. We also welcome listener feedback by email
1009
00:57:42.019 --> 00:57:44.639
at fedsocforums at fedsoc.org.
1010
00:57:45.440 --> 00:57:47.840
As always, keep an eye on our website.
1011
00:57:47.500 --> 00:57:50.320
And your emails for announcements about upcoming virtual events.
1012
00:57:50.659 --> 00:57:51.929
Thank you all for joining us today.
1013
00:57:52.150 --> 00:57:52.769
We are adjourned.
1014
00:57:53.610 --> 00:57:57.400
Thanks for listening to FedSoc Forums, a Federalist Society podcast.
1015
00:57:57.780 --> 00:58:00.300
Be sure to subscribe wherever you get your podcasts so
1016
00:58:00.340 --> 00:58:04.550
you won't miss an episode. To learn more, visit our website, FedSoc.org.