1
00:00:00,017 -->
00:00:04,117
The Innovation Research Interchange is a community of innovators and experts
2
00:00:04,117 -->
00:00:09,317
focused on today's top innovation growth drivers, from sustainability to ROI,
3
00:00:09,657 -->
00:00:12,317
the future-ready workforce, digitalization, and more.
4
00:00:13,317 -->
00:00:17,297
The IRI podcast features fresh insights from leading voices to help manufacturers
5
00:00:17,297 -->
00:00:19,757
solve innovation challenges and seize opportunities.
6
00:00:21,757 -->
00:00:25,337
Opportunities. Thank you everyone for your patience and for joining us today.
7
00:00:25,537 -->
00:00:29,817
And I want to give a big thank you to Iria Giuffrida who agreed to be interviewed.
8
00:00:30,017 -->
00:00:32,817
So, all right, you have to correct me if your title has changed,
9
00:00:32,937 -->
00:00:36,817
Iria, but I've got you down as Professor of Practice of Law and Assistant Dean
10
00:00:36,817 -->
00:00:43,737
for Academic and Faculty Affairs at the William and Mary Law School. Cool. Is that right?
11
00:00:44,857 -->
00:00:49,777
Perfect. Awesome. All right. And I'm Leigh Green, and I run IRI's Research and
12
00:00:49,777 -->
00:00:50,977
Strategic Initiatives.
13
00:00:51,037 -->
00:00:54,957
And just a few housekeeping items to get through before we can begin our conversation.
14
00:00:55,437 -->
00:00:59,297
So first, this session is being recorded, and you will get a link to the recording.
15
00:00:59,717 -->
00:01:04,357
We'll be distributing the recording as a podcast, so there won't be any visuals here today.
16
00:01:05,057 -->
00:01:09,077
Second, please hold all your questions until the end. And once we open the floor
17
00:01:09,077 -->
00:01:13,877
to questions, You can either raise your hand or you can unmute so that you can
18
00:01:13,877 -->
00:01:17,797
unmute and ask a question or you can type it into the chat and I'll read it out for you.
19
00:01:18,537 -->
00:01:22,497
Okay, boring stuff and technical difficulties out of the way.
20
00:01:22,837 -->
00:01:24,797
Iria, thanks so much for joining us.
21
00:01:25,537 -->
00:01:27,917
It's always a pleasure. So thank you very much for having me.
22
00:01:28,737 -->
00:01:32,617
So tell us about your background and your areas of expertise, please.
23
00:01:33,257 -->
00:01:36,837
So I'm not going to go through the titles because we want to get to the juicy stuff.
24
00:01:36,837 -->
00:01:41,037
As you mentioned, I do work at William & Mary Law School, and one of my areas
25
00:01:41,037 -->
00:01:45,517
of research and interest is sort of the intersection between artificial intelligence
26
00:01:45,517 -->
00:01:50,597
and the legal system, both terms to be interpreted sort of in the broadest possible sense.
27
00:01:50,977 -->
00:01:53,997
I want to give a couple of examples of things that I've been working on this
28
00:01:53,997 -->
00:01:57,897
summer. A colleague of mine, Daniel Shin, whom I believe you also have met.
29
00:01:58,607 -->
00:02:02,607
Daniel and I have just submitted what I think is the first draft of a chapter
30
00:02:02,607 -->
00:02:09,387
in a collection of contributions, thinking about how the executive order of October 2013,
31
00:02:09,967 -->
00:02:16,187
excuse me, 2023, in which the White House is trying to create a framework for
32
00:02:16,187 -->
00:02:21,327
the safe use of AI, how that may impact civil litigation in the United States. Right.
33
00:02:22,067 -->
00:02:27,687
So this isn't your first time speaking to us. We did this webinar last summer
34
00:02:27,687 -->
00:02:31,127
when everyone was blowing up and everyone's hair was on fire thinking,
35
00:02:31,227 -->
00:02:35,487
well, it's going to chat GPT is going to ruin my company or it's going to make it amazing.
36
00:02:35,967 -->
00:02:38,687
So you went through and you gave us a lot of updates then.
37
00:02:39,367 -->
00:02:43,987
And there were a lot of cases that were very much in the air around the use
38
00:02:43,987 -->
00:02:46,367
of LLMs and they were working their way through the courts.
39
00:02:46,787 -->
00:02:50,287
Can you give an update on the most significant cases?
40
00:02:50,987 -->
00:02:57,307
Yes. And going back to the discussion we had last summer, if I recall correctly,
41
00:02:57,587 -->
00:03:02,107
I think I grouped this kind of cases thinking about some were copyright cases
42
00:03:02,107 -->
00:03:03,347
and some were patent cases.
43
00:03:03,767 -->
00:03:09,527
So an update on the copyright front, summarizing what are very many complicated
44
00:03:09,527 -->
00:03:13,567
legal arguments, what plaintiffs in all these different cases have done,
45
00:03:13,787 -->
00:03:17,747
they've They've advanced the argument of an infringement of their copyright.
46
00:03:18,127 -->
00:03:21,047
Now, who are the plaintiffs? I think if you remember last year,
47
00:03:21,107 -->
00:03:25,407
we talked about a case amongst whose plaintiffs was Sarah Silverman.
48
00:03:25,507 -->
00:03:31,407
So medians have right speech, right? They write so that the copyright attaches to the written product.
49
00:03:31,707 -->
00:03:35,087
We talked about John Grisham, same thing, an author written product.
50
00:03:35,267 -->
00:03:40,787
We talked about visual artists, right? That produce whether it's physical art or digital art.
51
00:03:40,887 -->
00:03:45,327
We talked about musicians, right? So all these different category of cases were
52
00:03:45,327 -->
00:03:49,447
either had just been filed or had been filed a couple of months prior.
53
00:03:49,547 -->
00:03:50,647
So they were very, very thought.
54
00:03:50,947 -->
00:03:54,447
So as I said, the context is slightly different, but fundamentally,
55
00:03:54,627 -->
00:03:57,427
all the plaintiffs were saying, you can't.
56
00:03:57,896 -->
00:04:01,476
More often than not, it's OpenAI, but there are other large language models
57
00:04:01,476 -->
00:04:04,676
developers that are at the receiving end of these cases.
58
00:04:04,916 -->
00:04:10,596
So plaintively saying, you have used my work that is protected by copyright,
59
00:04:10,976 -->
00:04:17,976
it being written work, art, etc., without my consent or a license to train your large language model.
60
00:04:18,876 -->
00:04:24,376
So what I want now is a court to issue a permanent injunction forbidding you
61
00:04:24,376 -->
00:04:29,756
from using my data into my work into your large language model.
62
00:04:29,896 -->
00:04:36,556
I would like damages for the loss of opportunity of licensing my work to somebody else.
63
00:04:37,436 -->
00:04:42,096
And some of them also make sort of broader arguments about a market usurpation.
64
00:04:42,296 -->
00:04:45,936
So in fact, it's not every case. There is one case specifically that is still
65
00:04:45,936 -->
00:04:50,796
running in California where this unfair competition claim has been put forward.
66
00:04:51,136 -->
00:04:55,976
So without going into too many details, two cases have been dismissed.
67
00:04:56,856 -->
00:05:00,816
And again, generalizing what courts have said to the plaintiffs is,
68
00:05:00,896 -->
00:05:06,096
listen, if you're talking about a copyright infringement, I need one of two things from you.
69
00:05:06,096 -->
00:05:12,776
You either prove that what the defendant has done is a copy of your work,
70
00:05:12,816 -->
00:05:15,216
or that it's substantially similar.
71
00:05:16,256 -->
00:05:22,496
If you'd say anything else, this is not what copyright law, at present as it stands, allows me to do.
72
00:05:22,756 -->
00:05:28,516
And in both cases, one was Sibylman's case, the court dismissed the case,
73
00:05:28,516 -->
00:05:29,956
though without prejudice.
74
00:05:29,956 -->
00:05:34,496
This is very important because it's kind of hinting at the plaintiffs,
75
00:05:34,516 -->
00:05:39,496
maybe you can have another go at it, but I need a better statement of your claim.
76
00:05:40,333 -->
00:05:46,113
From the patent perspective, we haven't had a ton of sort of cases as such going on.
77
00:05:46,293 -->
00:05:50,453
However, we did discuss patent law last year.
78
00:05:50,613 -->
00:05:55,313
The question that had been asked in court and then had been denied was,
79
00:05:55,693 -->
00:06:02,933
can the author, excuse me, can the inventor in a patent application be an AI system?
80
00:06:03,373 -->
00:06:06,793
And the case rolled up and down the different courts. The answer is just no.
81
00:06:06,793 -->
00:06:12,533
However, the USPTO, so the Patent and Trademark Office, understands that this is an issue.
82
00:06:12,733 -->
00:06:16,493
And just over the course, I wanted to say a year, in fact, it's not even a year,
83
00:06:16,573 -->
00:06:20,553
it's over the course of 2024, has published three guidance.
84
00:06:21,053 -->
00:06:27,333
One guidance to users of AI systems as they innovate, right?
85
00:06:27,633 -->
00:06:32,973
One guidance on inventorship. The most recent came out two days ago,
86
00:06:33,033 -->
00:06:38,533
and I'm going to have to read this, apologies, but on guidance on subject matter eligibility.
87
00:06:38,853 -->
00:06:41,493
So what can be patented?
88
00:06:42,133 -->
00:06:47,333
And they have a request for comments that is open until the 29th of July,
89
00:06:47,513 -->
00:06:52,973
asking for comments from the sector, industry, academics, etc.,
90
00:06:52,973 -->
00:06:55,933
on certain criteria for patentability.
91
00:06:55,933 -->
00:07:01,473
So the patent office is listening to the anxiety that is now sort of moving
92
00:07:01,473 -->
00:07:06,213
into the innovation sphere, but the law hasn't changed.
93
00:07:06,453 -->
00:07:10,253
So we are, from a law perspective where we were last year, we better got this.
94
00:07:11,247 -->
00:07:14,907
Okay. All right. So what else should businesses be looking at?
95
00:07:14,947 -->
00:07:18,087
Because we're seeing, you know, an increased use of them.
96
00:07:18,647 -->
00:07:22,087
What are some other kind of watchouts that you have for us? Yes,
97
00:07:22,227 -->
00:07:28,727
I think expanding on the privacy point, I would say the protection of businesses
98
00:07:28,727 -->
00:07:30,007
on intellectual property.
99
00:07:30,247 -->
00:07:33,947
So the example that I had in mind when I mentioned an employee putting some
100
00:07:33,947 -->
00:07:37,467
information on ChatGPT, it was actually a Samsung engineer.
101
00:07:37,727 -->
00:07:41,427
So we're not talking about a sophisticated employee that put code,
102
00:07:41,687 -->
00:07:47,347
proprietary code, Samsung proprietary code into ChatGPT. it.
103
00:07:47,667 -->
00:07:52,867
And again, it wasn't done with any bad intention at all, right?
104
00:07:52,947 -->
00:07:57,387
This is just not understanding and not even thinking about the consequences
105
00:07:57,387 -->
00:07:59,107
of putting that kind of information.
106
00:07:59,487 -->
00:08:02,867
And it seems very easy for us, you know, something happened and now we can talk
107
00:08:02,867 -->
00:08:04,247
about it and we'll criticize the employee.
108
00:08:05,047 -->
00:08:09,007
But it's hard to train employees to think about, right?
109
00:08:09,367 -->
00:08:14,347
This is proprietary information. I think that it's easy to some extent,
110
00:08:14,587 -->
00:08:18,387
you think about, okay, this is a book, this is a painting, it's a piece of music, et cetera.
111
00:08:18,907 -->
00:08:23,207
Patents are different because patents are by their own nature out there, right?
112
00:08:23,267 -->
00:08:25,987
You have a patent, you have to describe everything in the process.
113
00:08:26,047 -->
00:08:28,527
So I'm not talking about that kind of intellectual property.
114
00:08:28,727 -->
00:08:33,247
I'm talking about trade secrets, the recipe of the Coca-Cola, right?
115
00:08:33,807 -->
00:08:38,267
What if an employee for another purpose puts a little bit of that information
116
00:08:38,267 -->
00:08:42,927
out there that then is triangulated with other data that other employees are
117
00:08:42,927 -->
00:08:45,667
putting in the system. How do you educate your workforce?
118
00:08:45,887 -->
00:08:49,087
This is, I think, one of the real challenges of businesses today,
119
00:08:49,187 -->
00:08:52,327
because the advantages of these tools are clear.
120
00:08:52,507 -->
00:08:55,907
They can accelerate the output. They can increase the quality.
121
00:08:56,227 -->
00:09:01,427
They can allow, for instance, greater reflection on, for instance, work product.
122
00:09:01,707 -->
00:09:05,447
This is sort of an argument I'm putting forward. What have I not considered?
123
00:09:05,747 -->
00:09:09,327
This is very, very useful. But that risk remains, and that risk,
124
00:09:09,327 -->
00:09:11,287
I think is quite significant.
125
00:09:12,207 -->
00:09:15,947
Problems with bias, again, they're not unique to large language models.
126
00:09:16,067 -->
00:09:19,567
They exist with all AI. The lack of transparency in the sense,
127
00:09:19,687 -->
00:09:22,127
okay, what data was used to train this large language model?
128
00:09:22,567 -->
00:09:31,427
Those businesses that essentially use ChatGPT and then create their own custom APIs don't know.
129
00:09:31,527 -->
00:09:37,087
We don't know what data OpenAI trained ChatGPT. We know it was the internet,
130
00:09:37,407 -->
00:09:41,867
minus some horrid side of the internet, but that's not sufficient, right?
131
00:09:42,467 -->
00:09:46,587
And sort of another aspect is privacy.
132
00:09:47,283 -->
00:09:52,083
Thinking about some of these systems are becoming ever sophisticated to auto-correct.
133
00:09:52,703 -->
00:09:55,983
So if an error is brought to the system attention, and again,
134
00:09:56,043 -->
00:09:59,643
I do lots of air quotes because I'm using language that kind of doesn't make
135
00:09:59,643 -->
00:10:00,623
sense to a machine, right?
136
00:10:01,303 -->
00:10:06,163
It can change, can tweak the way it deals with certain problems,
137
00:10:06,223 -->
00:10:07,263
questions, prompts, etc.
138
00:10:08,263 -->
00:10:13,063
All right. So regulations. Do you see any of those on the horizon?
139
00:10:13,283 -->
00:10:16,943
What form do you think they'll take? Do you think we'll ever get it together
140
00:10:16,943 -->
00:10:18,863
to do any sort of regulations?
141
00:10:19,443 -->
00:10:24,103
So I'm going to start, I'm going to create a kind of a conceptual pyramid,
142
00:10:24,383 -->
00:10:26,163
starting at state level.
143
00:10:26,283 -->
00:10:31,583
Actually, at state level, there has been an awful lot of activity in the space
144
00:10:31,583 -->
00:10:36,403
of AI more generally, which is what we have not seen at federal level.
145
00:10:36,623 -->
00:10:42,903
So different states have started to tackle and have been quite forceful in trying
146
00:10:42,903 -->
00:10:44,463
to tackle some of the challenges.
147
00:10:44,643 -->
00:10:47,663
For instance, we know that we are in an election year.
148
00:10:47,803 -->
00:10:52,643
Some states have made it a criminal offense to create a deep faith of either
149
00:10:52,643 -->
00:10:59,663
candidate making him say or her vice president something that is contrary to their view.
150
00:10:59,883 -->
00:11:05,603
Right. And that's step forward. You know, two years ago, we wouldn't even contemplated this.
151
00:11:06,485 -->
00:11:11,025
So I think the states are becoming, are aware of the states that are getting
152
00:11:11,025 -->
00:11:14,145
more significant and so are taking some action.
153
00:11:14,345 -->
00:11:20,665
Colorado, that very few people perhaps would have put on the map for being innovators,
154
00:11:20,765 -->
00:11:25,065
has come up with a piece of legislation that has expanded the notional sensitive
155
00:11:25,065 -->
00:11:30,945
data precisely to prevent the abuse in the context of large language models.
156
00:11:31,225 -->
00:11:33,945
Now, do all of these things work? We're going to have to see.
157
00:11:33,945 -->
00:11:40,605
But states are gradually engaging, even states that perhaps we wouldn't have
158
00:11:40,605 -->
00:11:44,725
anticipated as being at the forefront of regulation for AI systems.
159
00:11:45,345 -->
00:11:51,525
And the EU has, I think last year when we discussed the draft EU Act,
160
00:11:51,745 -->
00:11:55,245
I told you it's going to come out any minute. It didn't take a minute. It took more than that.
161
00:11:55,405 -->
00:11:58,825
It finally was published in the official journal last week.
162
00:11:58,905 -->
00:12:08,145
And so it officially becomes law, if you wish. comes into effect on August 2nd this year.
163
00:12:08,245 -->
00:12:13,425
However, it's 144 pages of European juicy regulations.
164
00:12:13,985 -->
00:12:18,865
So you can imagine businesses and maybe some of our colleagues are sort of starting to pull their hair.
165
00:12:19,365 -->
00:12:22,085
This could be a while. It could be a while.
166
00:12:22,185 -->
00:12:29,205
However, so there are different implementation dates throughout the regulation.
167
00:12:29,305 -->
00:12:32,905
So not everything becomes binding next week.
168
00:12:33,545 -->
00:12:38,645
Nonetheless, for businesses, so the main deadlines are that something comes
169
00:12:38,645 -->
00:12:43,765
in February and then August 25, August 26 and August 27.
170
00:12:44,405 -->
00:12:48,065
You know, individually may seem a lot of time, for businesses it isn't.
171
00:12:48,525 -->
00:12:54,045
It's not a lot of time, right? And so for businesses that use AI systems and
172
00:12:54,045 -->
00:12:55,625
are planning to use them in Europe,
173
00:12:55,805 -->
00:12:59,725
I think the pressure is on because they're going to have to start planning how
174
00:12:59,725 -->
00:13:04,845
to comply, even though they're not currently mandated to comply with those norms.
175
00:13:06,243 -->
00:13:08,683
Yeah, that's hard to keep up
176
00:13:08,683 -->
00:13:12,343
with, for sure, especially when we see the patchwork on state regulations.
177
00:13:12,543 -->
00:13:17,783
We've got the same going around sustainability and everyone's privacy. Privacy. Yes.
178
00:13:18,443 -->
00:13:25,203
And I think it puts businesses in the most enviable position because complying
179
00:13:25,203 -->
00:13:27,463
with regulation is costly.
180
00:13:27,563 -->
00:13:30,723
And I'm a regulation girl. I'm really regulation, but that's why I don't have
181
00:13:30,723 -->
00:13:35,023
a business, right? It's easy for me to say. They're on the safe side. But I don't understand.
182
00:13:35,623 -->
00:13:39,663
Compliance is costly. What a comment that you just made, Lee,
183
00:13:39,683 -->
00:13:44,763
which is investing in trying to work out what the regulatory regime is.
184
00:13:45,123 -->
00:13:49,103
And if you do businesses in multiple states or you do businesses abroad,
185
00:13:49,423 -->
00:13:55,403
the cost of compliance becomes sometimes so disproportionate that businesses
186
00:13:55,403 -->
00:13:57,723
are put off from engaging, right?
187
00:13:57,783 -->
00:14:03,763
Entering in new areas or engaging in other activities. So it's very complicated.
188
00:14:03,983 -->
00:14:07,743
It is very complicated for businesses, for innovators, for startups, right?
189
00:14:07,883 -->
00:14:13,343
And they do bring new ways of doing things that are very beneficial.
190
00:14:13,523 -->
00:14:16,423
So it's not a happy time, I think.
191
00:14:16,423 -->
00:14:19,803
The common view is
192
00:14:19,803 -->
00:14:23,683
probably not as much because we
193
00:14:23,683 -->
00:14:29,603
are talking about AI systems that use data or they are deployed in the EU this
194
00:14:29,603 -->
00:14:35,043
is very different from us being outside of the EU and then surfing on a website
195
00:14:35,043 -->
00:14:41,963
that services both American audiences and European audiences so I think that the
196
00:14:42,043 -->
00:14:50,063
geographic restriction is much more likely to affect only businesses that actually
197
00:14:50,063 -->
00:14:52,363
deploy these tools in the EU.
198
00:14:52,883 -->
00:14:57,803
This is, though, one consideration. It is still possible for a business who
199
00:14:57,803 -->
00:15:02,003
says, if that's the highest standard, I'm going to meet that standard because I don't have time.
200
00:15:02,163 -->
00:15:09,163
So it could be some kind of de facto. But the difference is that in the EU,
201
00:15:09,383 -->
00:15:10,663
there are certain kinds.
202
00:15:10,743 -->
00:15:14,463
So the way the EU AI Act sort of thinks about AI is about risk.
203
00:15:14,843 -->
00:15:19,463
So it's not about the kind of AI you come up with. It's the kind of risk it
204
00:15:19,463 -->
00:15:20,803
poses depending on the use.
205
00:15:21,683 -->
00:15:26,463
And so if you're using AI systems to make decisions that affect the people's
206
00:15:26,463 -->
00:15:28,963
access to benefits, that's forbidden.
207
00:15:29,723 -->
00:15:35,043
So how could it possibly affect us outside of the EU? You see what I mean? It doesn't.
208
00:15:35,583 -->
00:15:41,783
The parts where it may affect is if there are either relatively high or medium
209
00:15:41,783 -->
00:15:44,543
kind of risks, there are certain procedures that need to be in place.
210
00:15:44,683 -->
00:15:49,163
So the AI needs to be reported to the EU and then there's some mandatory governance.
211
00:15:49,823 -->
00:15:56,283
That's possible, right, that either it's brought over, it inspires action at EU level.
212
00:15:56,823 -->
00:16:04,003
But again, it doesn't have the same degree of permeability, if you wish, that the GDPR had.
213
00:16:05,154 -->
00:16:10,274
So kind of moving on to social and ethical implications.
214
00:16:11,294 -->
00:16:16,594
So different track now. So how are you seeing this play out already?
215
00:16:16,774 -->
00:16:18,154
Where do you see this going?
216
00:16:18,354 -->
00:16:23,534
Are we in doom and gloom phase? Are we optimistic? How do you feel?
217
00:16:25,694 -->
00:16:29,234
Probably moderately optimistic, I have to say. Of course, there are risks.
218
00:16:29,634 -->
00:16:35,034
And this is a powerful tool that has hit us all, right?
219
00:16:35,154 -->
00:16:38,714
So, for instance, I look at my profession, but I know that this is happening
220
00:16:38,714 -->
00:16:41,334
also in other disciplines, the risk of hallucination.
221
00:16:41,454 -->
00:16:47,834
You ask a question of chat GPT, right? These models are designed to produce
222
00:16:47,834 -->
00:16:54,194
a plausible output, not a correct output, right? That's a really good distinction.
223
00:16:54,614 -->
00:17:00,134
It's a massive distinction. So if, as some of my less sophisticated perhaps
224
00:17:00,134 -->
00:17:01,714
legal colleagues have done,
225
00:17:01,894 -->
00:17:10,354
if you ask ChatGPT to write a brief and if there are no cases to support certain arguments,
226
00:17:10,694 -->
00:17:15,214
ChatGPT will make them up because it is plausible, right? Right.
227
00:17:15,274 -->
00:17:21,314
To judge a GPT, again, air quotes, understands that to make an argument,
228
00:17:21,394 -->
00:17:22,494
you need to have a precedent.
229
00:17:22,974 -->
00:17:27,854
If there is no precedent, let me create one. Sometimes it's completely made
230
00:17:27,854 -->
00:17:30,854
up. Sometimes it's a case that exists about something completely different.
231
00:17:31,634 -->
00:17:39,854
So the risk of hallucination in the context of using professionally this kind of tools is high.
232
00:17:40,134 -->
00:17:43,674
And so, again, educational. How do I educate you? right it
233
00:17:43,674 -->
00:17:47,954
goes back to that to that kind of point it's not an impossible task
234
00:17:47,954 -->
00:17:50,854
to achieve but again it's a cost right you need
235
00:17:50,854 -->
00:17:53,794
to understand the risk and it's a cost in my that's sort
236
00:17:53,794 -->
00:17:56,834
of my academic industry cheating cheating it's
237
00:17:56,834 -->
00:17:59,614
it's it's very
238
00:17:59,614 -->
00:18:05,834
difficult for all educators because we now don't know how to teach students
239
00:18:05,834 -->
00:18:11,054
how to write right if i give my students as a hobby go home and you write your
240
00:18:11,054 -->
00:18:18,634
paper and I tell them you cannot use chat GPT 99% don't one does like the 1% does.
241
00:18:20,014 -->
00:18:22,754
And so far at least in my very
242
00:18:22,754 -->
00:18:26,474
very limited experience I have been able to spot the
243
00:18:26,474 -->
00:18:30,434
chat GPT generated out yeah you can kind of tell the language there's something
244
00:18:30,434 -->
00:18:36,874
I call it kind of metallic about it however this is obviously student students
245
00:18:36,874 -->
00:18:43,154
plural who have not who have We're not terribly sophisticated, right?
246
00:18:43,254 -->
00:18:44,694
Take it a step forward.
247
00:18:45,654 -->
00:18:49,494
Now, I say this, I have a colleague who actually teaches a course on how to
248
00:18:49,494 -->
00:18:50,954
use ChatGPT for legal research.
249
00:18:51,734 -->
00:18:58,094
And her point is, how is that different, not in the context of writing an exam,
250
00:18:58,274 -->
00:18:59,714
in the context of writing a motion?
251
00:19:00,234 -->
00:19:02,974
How is that different from the work that we do in a law firm?
252
00:19:03,921 -->
00:19:09,321
I, you know, I was a practitioner for 10 years, litigator, and litigators write
253
00:19:09,321 -->
00:19:12,121
very aggressive letters to each other. There's lots of brr.
254
00:19:12,581 -->
00:19:17,881
And the letter was always signed by the partner I worked for,
255
00:19:18,121 -->
00:19:22,141
who probably put two commas and maybe a full stop.
256
00:19:22,641 -->
00:19:26,481
Okay. And I'm exaggerating because I had my colleague was Marcus Rutherford.
257
00:19:26,661 -->
00:19:30,421
He's a fantastic lawyer, and we work very well together. But the point is,
258
00:19:30,521 -->
00:19:34,161
it's a collaborative effort,
259
00:19:34,281 -->
00:19:39,421
and it has to be so, because he didn't have the time to review however many
260
00:19:39,421 -->
00:19:42,501
cases we're working on to do the legal analysis, legal resources, et cetera.
261
00:19:43,001 -->
00:19:47,761
So her point is, how is that different from working, collaborating with somebody
262
00:19:47,761 -->
00:19:52,541
else if you're careful about hallucination, if you go and double check the sources,
263
00:19:52,901 -->
00:19:56,721
right, if you make sure that the arguments are good? So that's one point.
264
00:19:57,041 -->
00:20:02,561
How do you, how do as instructors we teach students to write in the presence
265
00:20:02,561 -->
00:20:07,901
and in existence of these tools is I think we need to help them use them.
266
00:20:09,043 -->
00:20:13,803
In a way that is first ethical, in a way that forces them to understand what
267
00:20:13,803 -->
00:20:14,923
good writing looks like.
268
00:20:15,423 -->
00:20:20,003
For instance, I made the argument, you know, you write something and you feed
269
00:20:20,003 -->
00:20:25,063
your chapter, your whatever, through ChatGPT and say, what's wrong with it?
270
00:20:25,583 -->
00:20:30,043
It's so useful. I tell you, it actually doesn't flow. You say this is an introduction,
271
00:20:30,223 -->
00:20:31,723
but you don't mention it anywhere.
272
00:20:33,003 -->
00:20:39,223
So it can be used as a useful tool. And I normally would do it with a colleague. Can you read this?
273
00:20:39,703 -->
00:20:43,963
Can you critique it? So it's very useful. But if you only use it for writing,
274
00:20:44,103 -->
00:20:46,203
then it's a problem. So hallucinations is one.
275
00:20:46,783 -->
00:20:50,403
The malicious uses, I feel, are underappreciated.
276
00:20:51,143 -->
00:20:56,083
And so think about teenagers on whatever social media platform,
277
00:20:56,303 -->
00:21:02,263
how this kind of tool is used to either defraud people.
278
00:21:02,263 -->
00:21:05,963
I think when we talked about it last year, it wasn't quite as common,
279
00:21:06,043 -->
00:21:11,983
but there are deepfakes tools that are linked to large language models that
280
00:21:11,983 -->
00:21:16,303
are essentially deployed to get people, mostly people who are unaware,
281
00:21:16,583 -->
00:21:20,103
to give vast amounts of money because their kid is in jail.
282
00:21:20,243 -->
00:21:26,303
I mean, it takes three seconds of my voice to a relatively sophisticated deepfake
283
00:21:26,303 -->
00:21:30,143
system to reproduce it and make me say anything I want.
284
00:21:30,623 -->
00:21:34,163
And this is with free tools. It's not you need to create a...
285
00:21:34,163 -->
00:21:36,823
This is with free tools. You just need to know how to use them.
286
00:21:37,463 -->
00:21:45,163
So it's the malicious uses that I think cause problems. And along the lines of misinformation.
287
00:21:45,503 -->
00:21:48,543
Again, it's not an LLM-only issue. This is an AI issue.
288
00:21:49,463 -->
00:21:52,603
The fact that the scale of the effects is variable.
289
00:21:53,063 -->
00:21:56,163
This is sort of some of the issues. The bias, the lack of transparency,
290
00:21:56,343 -->
00:22:00,403
they remain there. They just get a little bit more amplified when you think
291
00:22:00,403 -->
00:22:06,443
about the social media and how a piece of misinformation can go wild within a second.
