Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:01
Support for Decoder comes from SAP Business
0:03
AI. Sure, we've all had
0:05
fun messing around with AI image generators and
0:07
conversation bots, but AI is more than
0:10
a novelty. Businesses around
0:12
the world have found ways to harness
0:14
its potential, like spotting inventory shortages before
0:16
they happen, or supporting supply chain management.
0:19
And it's very possible that your business could
0:21
benefit from AI integration too. Unlock
0:23
the potential of AI and discover even
0:26
more possibilities with SAP Business AI. Your
0:29
technology, real world results. That's
0:32
SAP Business AI. Learn
0:34
more at sap.com/AI. Hello,
0:41
and welcome to Decoder. I'm Neil Aptel, editor-in-chief of
0:43
The Verge, and Decoder is my show about big
0:45
ideas and other problems. We're
0:47
doing two decoders a week now. On Mondays,
0:49
we're going to have our regular interviews, but
0:51
our new Thursday episodes, like this one, are
0:54
all about deep dives into big topics in
0:56
the news. And for the next few
0:58
weeks, we're going to stay focused on one of the
1:00
biggest topics of all, generative AI.
1:03
There's a lot going on in the world
1:05
of generative AI, and maybe the biggest thing
1:07
going is the increasing number of copyright lawsuits
1:09
being filed against AI companies like
1:12
OpenAI and StabilityAI. So,
1:14
for this episode, we brought on Verge features
1:16
editor Sarah Jong, who is a former lawyer
1:18
just like me, and we're going to talk
1:20
about those cases. And the main defense the
1:22
AI companies are relying on, an idea called
1:24
fair use. Let's back
1:26
up a sec. All the big generative AI
1:28
models from every company are trained on huge
1:30
swaths of data that are scraped from the
1:32
entire internet. And big media
1:34
companies like the New York Times and
1:36
Getty Images have filed copyright lawsuits against
1:38
those AI companies, saying that basically, they've
1:40
stolen their work and are profiting from
1:43
it. A claim that amounts
1:45
to straightforward copyright infringement. I made something,
1:47
you made a copy without my permission,
1:49
that's copyright infringement. If there's
1:51
one thing to know about copyright law, it's
1:54
that it's still very much rooted in
1:56
the idea of making copies, and regulating
1:58
which copies are legal. in which
2:00
aren't. And since computers can't do
2:03
anything at all without making copies, copyright
2:05
law shows up again and again in
2:07
the history of computing, and especially the
2:10
history of the internet, which allows anyone
2:12
to make and distribute perfect copies faster
2:14
than ever before. But there's a
2:16
check on all of the control that
2:18
copyright law provides. Fair use. Fair
2:21
use is written right into the Copyright Act,
2:23
and it says that certain kinds of copies
2:25
are okay. You can quote things. You can
2:28
quote books in commentary about books. You can
2:30
run clips of movies in video criticism of
2:32
movies. And you can make copies of articles
2:34
to share in a classroom. There's a long
2:36
list of these things in the Copyright Act,
2:39
but since the law can't list or predict
2:41
everything that people might want to do, it
2:44
also has a four-factor test written into it
2:46
that courts can use to determine if a
2:48
copy is fair use or not. But
2:51
here's the thing about the legal system
2:53
in general, and fair use specifically, it
2:56
is not deterministic or predictable. I
2:59
know we have a lot of engineers and product
3:01
managers in the decoder audience, and it's tempting to
3:03
think about the legal system like a computer that
3:05
you put in some inputs, and you can predictably
3:07
get some outputs. But that's not
3:09
how it works at all. And it is
3:11
especially not how fair use works. Every
3:13
court gets to run that four-factor fair
3:16
use test any way they want. And
3:18
one court's fair use determination isn't actually
3:20
precedent for the next court. That
3:23
means fair use is a very vibes-based
3:25
situation. It's anyone's guess how a lot
3:27
of copyright lawsuits are going to go.
3:30
Many of them feel like a coin flip. And
3:32
when you add in the amount of hype,
3:34
uncertainty, and money that comes with AI, it
3:37
gets even more complicated. So I
3:39
wanted Sarah to come on and help me explain what's
3:41
going on to everyone. Sarah is one
3:43
of my very favorite people to talk to about copyright law.
3:45
I promise you, we didn't get totally off the rails and
3:48
running out about it. But we went a little
3:50
off the rails. But we had to
3:52
start at the start. The first thing we had
3:54
to figure out was how big a deal are
3:56
these AI-conforming lawsuits? I feel
3:58
like there's sort of a... potential extinction level
4:00
event on the horizon. It's pretty weird because
4:03
all the lawyers seem to think so and
4:05
for whatever reason, like the CEOs don't seem
4:07
to think so. My read when
4:09
I talk to the CEOs is that they think
4:11
this is a money problem. That something's going to
4:13
happen and their general counsels and their
4:16
policy people are going to walk through
4:18
some court cases and maybe get some
4:20
policy changes passed in Congress and
4:22
they'll have to pay some money, but it will be fine.
4:24
And in the end, the money is always fine. You
4:27
are an extinction level event. I'm feeling
4:30
like the noise I'm hearing indicates
4:33
extinction level event. Why
4:35
do you think it's that bad? I
4:38
mean, like we lived through Napster,
4:40
right? Like we like it's it which
4:42
is weird because the CEOs also lived
4:44
through Napster, but maybe
4:46
they didn't like maybe they're from another universe.
4:49
But yeah, like it's the level of
4:51
copyright direness in these cases, the
4:54
effects on existing industries plus
4:57
how applicable law is lining up. It's
4:59
got Napster vibes to it. And
5:02
when Napster happened to the law, entire
5:04
companies went bust, entire industries went bust,
5:06
copyright changed forever in a way that
5:09
was not great. It was
5:11
an extinction level event and AI
5:13
has a similar thing going on there.
5:16
They got sued that went all the way to Supreme Court.
5:18
The Supreme Court made some changes
5:20
to copyright law in that case. When
5:23
most people think about Napster, I'm pretty sure they
5:25
think about Justin Timberlake playing Sean Parker in the
5:27
movie The Social Network. And they
5:29
might think about the idea that a
5:32
company that just quote unquote facilitates piracy
5:34
is a bad idea.
5:36
They do not think about the Supreme
5:38
Court eventually issuing changes to copyright
5:40
law wholesale that we now live inside. So
5:42
explain what you mean there. Quickly give people
5:44
the capsule summary of Napster and Groxner and
5:47
what happened to the law in those cases.
5:49
Yeah, we exit one era where we
5:51
had just sort of softened fair use
5:54
so that it was okay for people
5:56
to use their VCR setups to record
5:58
off the television. So that was the
6:00
arrow we were exiting, where it was like, okay, so
6:02
like, there's there are these new technologies, and people are
6:04
going to use them for themselves in
6:06
these, like, you know, pretty benign ways,
6:08
and that's okay. And copyright doesn't
6:10
have to restrict that. And
6:12
then we enter into sort of the Napster era where they go,
6:15
everyone can be a pirate now. And
6:17
that's not good. It could destroy this
6:20
industry. So now we have to change
6:23
copyright law in a way that we've never seen
6:25
before. That's
6:28
a really good example of something I want to
6:30
hammer on as we cover the AI companies. The
6:32
concept of fair use was enshrined into federal law
6:34
in the Copyright Act of 1976. It's
6:37
almost 20 years before the consumer
6:39
internet came along. So when
6:42
digital culture hit and companies like Napster arrived on the
6:44
scene, we had no idea what was going to happen.
6:47
No one in 1976 could predict Napster. And
6:49
the record labels and Napster had to go
6:51
to court to figure out if Napster was
6:53
legal at all. Turns out it wasn't. Napster
6:55
basically met its end in 2001 when a
6:57
federal appeals court upheld a ruling
7:00
that determined Napster by facilitating the
7:02
copyright infringement of its users was
7:04
also liable for copyright infringement. A
7:07
few years later, another peer-to-peer file sharing network
7:09
called Grokster went all the way to the
7:11
Supreme Court with a very similar lawsuit. Grokster
7:14
was a different company and the same
7:16
federal court that had shut down Napster
7:18
said Grokster was not infringing on copyright
7:21
law. But because Napster was the forerunner
7:23
of a whole bunch of file swapping
7:25
platforms ending in stir, Grokster ended up
7:27
being painted with the same brush. And
7:29
that left the Supreme Court to make
7:31
a big decision that had major ramifications
7:33
on everything that's happened online in the
7:35
last 20 years. Ultimately,
7:37
the court said if you market a
7:39
tool for people specifically to do copyright
7:42
infringement with, you are liable for the
7:44
copyright infringement that happens as a result.
7:47
That is a judicial construction. That idea had to be
7:49
invented. And there was a lot of disagreement about it
7:51
at the time. You can go into the history and
7:53
drama of that case, the cases that came before it
7:55
and the cases that came after it. But the point
7:58
I want to make is that no one knew
8:00
what was going to happen. The Supreme
8:02
Court had the power to effectively
8:04
create or destroy a company and
8:06
an entire industry based on its
8:08
understanding of copyright law at that
8:10
time. And at that moment,
8:12
they said, this is illegal and those
8:14
companies basically disappeared. And that is the
8:17
extinction level event that Sarah is describing.
8:19
The justices might say, well, actually all
8:21
this is illegal. And then the entire
8:23
AI industry might disappear. So
8:30
now we come to the AI companies, which are also making a bunch
8:32
of copies. And the
8:34
argument that is getting made
8:36
everywhere is this is something called fair
8:38
use. Yep, we acknowledge that we've made
8:40
the copies, but we've done them
8:42
in a way that makes it OK because of something called
8:44
fair use. Can you quickly explain what fair use is? So
8:47
fair use is the escape valve for
8:49
copyright because it is wild for
8:52
the law to restrict other
8:54
people's speech based on whether or
8:56
not you've published it in a book or set
8:58
it on a tape or whatever. And
9:00
so you have these four factors in
9:03
the law. You can look them up, 17 USC 107. They're
9:06
like intertwined. There's not really like a clear
9:08
logic behind how the four are like lined
9:10
up. And in fact, if you go and
9:13
look at the cases where fair use is
9:15
implicated, you can see the factors being weighed
9:17
very differently per case. You don't even need
9:19
to meet all four depending on the
9:21
use. It's not super
9:23
clear. It is meant to
9:25
be very flexible because speech is important
9:28
and you want to have a really
9:30
flexible escape valve. But that
9:32
also means it's not super
9:35
predictive in cases of new
9:37
technologies. So AI companies are
9:39
getting sued. The New York Times, for example,
9:41
has sued OpenAI. And the New York Times has complaint is
9:44
very compelling because it has all of these
9:46
examples where OpenAI will just spit out word
9:49
for word recitations of New York Times
9:51
articles. They've obviously copied the information. OpenAI's
9:54
response is, yep, we've acknowledged that we've
9:56
made these copies. But that copying is
9:58
fair use. We're allowed to do it
10:01
and we're going to show it to you by
10:03
going through the four factors So the first factor
10:05
is purpose and character of the use. What
10:07
does that mean? And how do you think applies in this case? so
10:10
what's the difference between for instance a middle
10:12
school or opening up the New York Times
10:14
and like quoting the
10:16
New York Times in their book report about Soybeans
10:21
Like go right like it's just like you
10:23
you it's a source So why not
10:25
go to the New York Times to
10:27
like get the definitive information on some
10:29
news story? People are also going to
10:31
open AI against the advice of their
10:33
teachers to like copy paste
10:35
a paragraph about soybeans for their book
10:37
report Right. It's like the purpose and
10:39
character of the use including whether such
10:41
use is of commercial nature or
10:44
for nonprofit educational purposes Yes,
10:46
open AI is ultimately going to make money
10:48
off of this thing. They charge you
10:50
for GPT for right now, right? Right pay
10:53
them a subscription and you get access to
10:55
someone else's information That
10:57
seems tough, right? It is tough the
11:00
big missing word in this that's sort
11:02
of been added over time Through
11:05
the courts actually is that
11:07
they're looking for transformative use
11:10
That's just something that's evolved over
11:13
the years if a work
11:15
is transformed by the copying
11:19
There's like a stronger argument to be made
11:21
that it was a fair use So
11:24
you've got like, you know a parody you've
11:26
got mashups. That's like a classic one if
11:28
you're like doing a YouTube
11:30
like clap back and you like have
11:33
a little clip of the person you're
11:35
clapping back like that's a transformative
11:37
use but You
11:39
can kind of tell from all of those easy
11:41
examples I used you can definitely think of a
11:43
time Someone got in trouble with
11:46
copyright law for doing exactly that Which
11:48
goes again to like their use is kind
11:51
of a funky thing where like because
11:53
it's case-by-case Even if it seems
11:55
like it's easy, you can still get in
11:57
trouble might win in the end, but you'll still
12:00
get in trouble. If you get into like a
12:02
much more difficult scenario like open
12:04
AI, something that has never been to court
12:06
period, you're up leveling
12:09
the difficulty to another place.
12:11
And I feel like in the case of
12:13
open AI, whether or not
12:16
copying all the information on the internet so that
12:18
a robot can spit it back out at you
12:20
in slightly different formats, whether
12:22
or not that's transformative is wildly up in the
12:24
air. In some cases, it
12:26
clearly is transforming. But
12:29
the New York Times has all those examples where
12:31
the robot just spits it back verbatim.
12:33
It's not transforming. And so you can
12:35
sort of see like the New
12:37
York Times is trying to preempt that
12:39
transformativeness debate. They're like, yeah,
12:41
like you like if it's spitting it out
12:43
verbatim, how much transforming is actually going
12:45
on in here? You get kind of like
12:48
an almost circular argument there
12:50
where it's like if it's not doing
12:52
it verbatim in some cases, then
12:56
when it is doing it verbatim, it's still
12:58
transformative because that like whatever internal
13:00
guts are happening in there, like it's
13:02
like clearly changing things just because we
13:04
got like a one off
13:06
like verbatim quote, surely
13:08
that means it's like it's
13:11
still okay. It's a weird one. Yeah,
13:13
I will just point out to the audience.
13:16
We're already in the middle of like a
13:18
deeply existential debate on the very nature of
13:20
how AI systems work. And what
13:22
if they're transforming the source text? We're
13:24
at the first factor. We haven't
13:26
even made it out of the first one. There's three more to
13:28
go. And they're all they're all like this. And some of them
13:30
are even walkier. This is why when
13:33
I say it's a coin flip. This is what I
13:35
mean. Like I feel like Sarah and I could just
13:37
sit here debating whether or not AI is transformative for
13:39
the rest of the show and not
13:41
reach a conclusion. And it's not us who's
13:43
deciding it in the end. It's a bunch
13:45
of judges. And I don't know what they're gonna think. We
13:49
have to take a quick break. When we come back, Sarah
13:51
and I will start diving into the other three factors in
13:53
a fair use case. A
14:01
Corporate A Coder comes from S A P
14:03
Business Ai It's all over the internet. A
14:05
I dare say I that your friend is
14:08
turning his cat into a Monet painting. your
14:10
coworker used a chap ah to read solid
14:12
about Pancakes A eyes in the stuff of
14:14
science fiction anymore, but it's also more than
14:16
the gimmicks we see on a day to
14:18
day basis. If your business owner A I
14:20
can offer real solutions to help you scale
14:23
an enemy, it might be time to check
14:25
out S A P Business Ai Sep Business
14:27
A I can help you automate repetitive tasks,
14:29
optimize inventory management and supply chain analysis. And
14:32
identify opportunities for growth in your
14:34
operations. Sep Business A I can
14:36
help you with finance, sales, marketing,
14:38
human resources, procurement, supply chains, and
14:40
so much more. Like guarding against
14:43
fraud with A I assisted anomaly
14:45
detection or receive data driven prescriptive
14:47
guidance A critical decision points. They
14:49
even have a generative Ai road
14:51
map to help you discover upcoming
14:54
and cutting edge innovations for your
14:56
business. Who knows when innovations are
14:58
around the corner? Revolutionary technology, real
15:00
world results. That s a P
15:02
Business A I learn more it as
15:04
a P.com/a I. Was
15:12
back. We were just talking about factor one of
15:14
the ferries test which is purpose and character of
15:16
use and we talked about the idea that some
15:18
copies or transform it into the original work and
15:21
he turned him into something else. That
15:23
is major we up for debate in the
15:25
case of a generative Ai system. My church
15:27
between. The
15:29
second series factor is the nature of
15:32
the work is widely available. Is it
15:34
a secret is had stashed away the
15:36
copy, the new drugs? everyone else. How
15:38
does nature of the work place in
15:40
these cases? Some. Things are
15:42
considered like a little more. In.
15:45
The purview of copyright than
15:47
others. So like you know,
15:49
arts plays, that anything you
15:51
know, creative works. that's. Much
15:54
more like. valuable than
15:56
say the the sequence
15:58
structure an organization of
16:00
Java, the language, right?
16:03
Which is, by the way, copyrightable. But
16:05
it's like kind of
16:07
a little bit less copyrightable, like on the
16:09
down low, than the other stuff. Because it's
16:11
weird. It's just a weird thing that you
16:13
don't really want to be like, oh yeah,
16:15
this is clearly what the founding fathers
16:18
wanted us to protect with copyright
16:20
to the law. I
16:22
always think of nature of the work as how
16:25
the judge feels about it. Like
16:28
it's just up for whatever it is. It's like,
16:30
how do I feel about the song Pretty Woman?
16:32
Is it important that I'm going to turn up
16:34
the dial on this one and say you've got
16:36
to overcome it a lot? Is
16:38
it a list of APIs in a spreadsheet in
16:40
order? I'll turn the dial down a little bit
16:42
because that seems pretty silly. And that, as
16:45
far as I can tell, is how this
16:47
one gets assessed. It feels
16:49
like the entire contents of the New
16:51
York Times archive might get the, I
16:53
feel like this is pretty important, waiting.
16:55
It's hard to know. I think the dial
16:57
is kind of in the middle on this one because on the one
16:59
hand, it's the kind of
17:02
creativity that you want to incentivize. But on the other
17:04
hand, it's full of facts. And
17:06
when it comes to facts and
17:08
history, facts themselves aren't copyrightable.
17:11
And you don't want to put a big fence
17:13
around the first draft of history, essentially. So
17:15
I think it does end up sort of
17:17
in the middle there. The part
17:19
of the nature of the work factor that I think is
17:21
also a bit of a coin flip here is
17:24
how widely available something is. So the
17:26
New York Times thinks of
17:28
itself as the news. It has
17:30
a very high self regard. And it thinks of itself
17:33
as pervasive. And it is the first draft of history
17:35
in all the ways the Times thinks about itself. It
17:37
is also famously behind a paywall. It's a thing you
17:39
have to pay for. You have to pay a lot
17:41
of money for it. They can mail you pieces of
17:43
paper, which is a very interesting way of receiving the
17:46
news. There's a tension there. The
17:48
tension is in the complaint. They're saying,
17:50
look, Chat CPT lets people get past the paywall. It's
17:53
the thing you have to pay for. And you can pay someone else to
17:55
get it for free. The argument is you're
17:57
reducing the market for the New York Times. How do you think those two
17:59
things? square. Yeah, like chat GPT,
18:02
letting you get around the
18:04
New York Times paywall, and the
18:06
New York Times being paywalled in the
18:08
first place. That does
18:11
seem to run in the
18:13
favor of the New York Times and against
18:15
open AI. Yeah. And again,
18:18
I will tell the audience, you can see
18:20
how complicated this is from the jump. This
18:22
is not, there's nothing deterministic about
18:24
this analysis. At every point, you can have
18:27
a pretty existential argument. But to
18:29
me, it's the nature of the work, it is
18:31
the least deterministic. Because if the judge decides that
18:33
they don't like the New York Times that day,
18:36
they can just turn the knob down and
18:38
say something like the news is the news. It's
18:41
not a poem. Everyone
18:44
has the news. It's open, I didn't steal it from
18:46
the New York Times, they would take it from the AP. And it's the same
18:48
news. And like, that is a
18:50
thing that they could logically say here to
18:52
devalue this factor. Yes, they
18:55
could say that. We're
18:58
about to get into the third and fourth theories factors.
19:00
And one thing we're going to talk a lot about
19:02
is Google, specifically Google books. 20 years
19:04
ago, in 2004, Google said it wanted to
19:06
scan and make searchable all of the books
19:09
in some research libraries. The
19:11
copyright holders, authors and publishers said, No,
19:13
you can't do that. And they filed
19:15
copyright lawsuits. It took nearly
19:17
a decade. The case was
19:20
finally resolved in Google favor in 2015.
19:22
The federal appeals court held that turning
19:24
books into search snippets was fundamentally transformative
19:26
of the original work. There's a factor
19:28
one, and said that even though Google
19:30
has scraped the entirety of the books
19:33
and made a profit from offering its
19:35
services, it is not hurting the
19:37
market for the original books. That's
19:39
where factors three and four come in. Factor
19:44
three in the analysis is how much of the
19:46
work was used. This one
19:49
might be the most obvious one on its face, which
19:51
is that's all of it. It's bad. If it's as
19:53
little as possible, it's good. Is there any nuance to
19:55
this one? I actually do think there's a little bit
19:57
of nuance to this one. This is where the Google
20:00
Books cases, I think, sort of cut
20:02
in favor of OpenAI where like, yes,
20:04
they're taking 100% of the
20:07
New York Times, like as much of the New York Times as
20:09
they can get. Yes. But like, what's
20:12
spitting out is small
20:15
in comparison to what they're taking. And
20:17
so like, the fact that they've taken everything
20:20
is sort of minimized in the Google
20:22
Books cases, for instance, where they're like,
20:25
yeah, they had to read
20:27
everything. But because a
20:29
robot was doing all of the reading, that
20:32
means a lot less. That
20:35
makes sense. Like the human literate
20:38
part of the copying
20:40
is what's important here is sort
20:42
of, I think, the upshot
20:44
of like those cases. And
20:46
so here, I think it's
20:48
a little bit more mitigated for OpenAI.
20:52
So even though OpenAI has taken everything the fact
20:54
that it's literally really everything, the
20:56
fact that it's not sort of immediately available,
21:00
kind of diminishes this factor. Yes.
21:02
That's fascinating. I honestly wonder
21:05
if that argument will be
21:07
made well, and if the judges will accept it
21:09
well, given that the main thing
21:11
that AI companies have to do using the state is
21:13
train on it, right? Google Books is like we made
21:15
an index of all the books, and
21:18
we can show you parts of the index and then kick
21:20
you out to buying a book. OpenAI
21:22
is like, we copied everything, we trained
21:24
a model on it. And
21:26
we might have even thrown away the database of
21:28
copies that we made. And now the model can
21:30
just go converse. And
21:32
it's like you had to take all
21:35
of the stuff to train the model.
21:37
And that just seems like
21:39
very complicated to me, because there's
21:41
a little bit of technical nuance
21:44
to how AI models work, that
21:46
requires all the stuff, even if all the stuff isn't
21:48
out in the world, or
21:50
exposed to the user. I mean, that's also
21:52
how search engines work, right? Like, it's like you have to
21:54
have all the stuff in order to be able to search
21:56
it. I mean, the technology is subtly different,
21:59
how the technology is used. is being used is subtly different. The
22:02
other part of it that no one wants
22:04
to like really even think about or hear
22:06
about is that the Google Books cases, we
22:09
don't have a Supreme Court decision out of
22:11
them. Like the important bits come out of
22:13
the appellate courts and it's been
22:15
10 years. So it's like almost 10 years.
22:17
So it's like you're, we've seen a shift
22:20
in how copyright, especially fair use is
22:23
being addressed by the Supreme Court. It's a completely different
22:25
court. Like we could get something
22:28
very strange out of this that
22:30
is unexpected. Let's talk
22:32
about the fourth factor here, which feels like
22:34
the most important one in this
22:37
case and often feels like the most important
22:39
one in any fair use case, the effect
22:41
of your copy on the market for the
22:43
original work. I make a song, it samples
22:46
your song. The sample
22:48
is de minimis in some way, it's just
22:50
like background noise. My song is really popular.
22:53
That doesn't mean people are going to listen to my song
22:55
instead of listening to the original. And in fact, it
22:58
might mean that people love the sample so much that
23:01
sales of the original song go up. So
23:03
you've had some impacts on the market. It could
23:05
be positive or negative or nothing. And we're going
23:07
to try to figure that out. And if it's
23:09
positive or nothing, maybe that use
23:11
is fair. If I have taken so much
23:13
of your stuff and replaced your work
23:15
with it and your market goes down and
23:17
your sales go down, that's negative and that's
23:20
going to cut against fair use. I
23:23
don't know how to evaluate that in the case of AI
23:25
at all. It feels like
23:27
they're racing the market for all human
23:29
generated content in the world. But
23:32
then I use the tools and I'm like, I
23:34
think I think I still have something to say here. I
23:38
mean, I actually think that the fourth
23:40
factor is really, really
23:42
against open AI. And
23:44
I think that it's because of the
23:46
Warhol case. So we have this case
23:48
where you get like an Andy Warhol
23:51
style portrait of Prince, or
23:54
you've got the famous Marilyn Monroe thing
23:57
where it's like the cutouts and
23:59
the choppy. print thing. So
24:01
a magazine thinks about
24:03
putting a picture like a photo of
24:05
Prince on their cover, and they're like, no,
24:07
everyone's going to put a photo of Prince
24:09
on their cover after he died. Let's
24:12
get an Andy Warhol style thing instead
24:15
from the Andy Warhol Foundation. So
24:17
they like make a portrait of
24:20
Prince in the
24:22
style of Andy Warhol. And
24:24
the base that they use is a
24:26
photograph that someone else took. They
24:29
do not license the photograph. And
24:32
the person who took the photograph
24:34
is the kind of person whose
24:36
photographs get licensed to
24:38
be put on the cover of magazines. And
24:40
the court basically just goes, look, you like snapped
24:43
up an opportunity that this person theoretically had.
24:45
And the Warhol Foundation is like, no, this
24:47
is like, it's not the thing. They didn't
24:49
want a photograph. They wanted Andy Warhol. They're
24:51
like, yeah, but it's like the same market.
24:53
It's about the same. Right. I
24:55
don't think that there's been a Supreme
24:57
Court case that emphasize Factor 4 that
25:00
heavily before. And it's,
25:03
I think, like sort of a
25:05
warning shot, actually, for these new
25:07
technologies. I don't know if they had the new
25:10
technologies in mind. But like, definitely, this is
25:12
not something we've seen out of courts
25:14
before. Is that heavy of an emphasis
25:16
on Factor 4? One thing that's
25:18
interesting about that note on Factor 4, which is
25:20
the economic factor, you could call it, is that
25:22
in a time since you and I graduated from
25:24
law school, and now there is
25:26
a movement called law and economics in
25:29
the law that really emphasizes these ideas.
25:31
Like the law should be measurable. We
25:33
can apply economic thinking to it. That
25:36
was not so much in vogue when
25:38
the Napster cases were getting decided, the Groschke cases
25:40
were getting decided. And so now you have this
25:42
other fair use thing where it's like, is a
25:45
painting replacing the market for a photo? And
25:47
the judges are like, we can do some economic
25:49
thinking here. And we're going to prove it by
25:51
saying, yes, there is a market for
25:54
depictions of prints. And
25:56
this painting can serve that market as well as
25:58
the photograph, which sounds... But
26:02
I think in a case of open AI,
26:04
the economics of it actually become pretty
26:07
direct. There is a
26:09
market for information or writing or what
26:12
books in this robot for
26:14
20 bucks a month can
26:16
just substitute for
26:18
users all of the other kinds of products that they
26:20
might otherwise buy. Again,
26:23
I think there's complexity
26:25
here because I actually don't think the
26:28
GPTs right now, the GPT-4,
26:31
can't actually do the work.
26:33
It's not quite good enough. So
26:35
you have to pull your mind ahead to where it
26:37
obviously will be good enough in the future. But
26:41
right now, I wonder if the difference between
26:43
what it can do right now and what
26:45
it might do will weigh into this analysis.
26:48
You were putting open AI in the position of going like, our shit's
26:50
not that good. So you
26:52
can't sue us because our... I mean, any big
26:54
company will argue that it is a piece of
26:56
crap if that is legally advantageous. I
26:59
mean, yeah, but that's
27:02
where they're going to have to go in order
27:04
to make Factor 4 work with them, right? It's just
27:06
like, oh, yeah, we're not that great. We actually suck
27:09
and we're always going to suck. Like,
27:11
therefore, we will never impact the
27:14
commercial value of this work. I actually don't
27:16
know if they're going to be willing to go that far because
27:18
it's a bit much. I straight up
27:20
think that they are going to have to really
27:22
minimize Factor 4 as much as they can and
27:24
just talk around it and try to really push
27:27
their case on the other factors. I think Factor
27:29
4 is like, that's a
27:31
rough one for them. I think it's especially
27:33
rough given that the
27:35
most recent various cases we have out of SCOTUS
27:37
is a Factor 4 case. I
27:40
think that might be the biggest sign to me that
27:42
we're headed towards an extinction level event. We
27:46
have to take another quick break. We'll be right back. Thank
27:58
you. AI.
28:01
It's all over the internet. AI this, AI
28:03
that, your friend is turning his cat into
28:05
a Monet painting, your co-worker used a chatbot
28:07
to write a song about pancakes. AI
28:10
isn't the stuff of science fiction anymore, but it's
28:12
also more than the gimmicks we see on a
28:14
day-to-day basis. If you're a business
28:16
owner, AI can offer real solutions to help you
28:18
scale and innovate. It might be time to check
28:21
out SAP Business AI. SAP
28:23
Business AI can help you automate
28:25
repetitive tasks, optimize inventory management, and
28:27
supply chain analysis, and
28:29
identify opportunities for growth in your operations.
28:32
SAP Business AI can help you
28:34
with finance, sales, marketing, human resources,
28:36
procurement, supply chain, and so much
28:39
more. Like guarding against fraud
28:41
with AI assisted anomaly detection, or receive
28:43
data-driven prescriptive guidance at critical decision
28:46
points. They even have a
28:48
generative AI roadmap to help you discover
28:50
upcoming and cutting-edge innovations for your business.
28:53
Who knows what innovations are around the corner? Revolutionary
28:56
technology, real-world results. That's SAP
28:58
Business AI. Learn more
29:01
at sap.com/AI. We're
29:09
back. I'm talking about fair use with Verge Features
29:11
Editor, Sarah John. We've talked about
29:13
what the four factors in a fair use case are,
29:15
but that leaves us with a really big question. How
29:18
is this all gonna play out for the AI companies?
29:22
I often use the blurred lines case. Oh
29:24
no. An example of how much of a
29:26
coin flip all of this is. The
29:29
state of Marvin Gaye and Pharrell
29:32
Williams get into a dispute. The estate of
29:34
Marvin Gaye wins, even though there's no copying
29:36
in that song at all, which is still
29:38
crazy to me. Later
29:40
on, the estate of Marvin Gaye goes
29:42
after Ed Sheeran. There is actual musical
29:47
similarity in the two songs there, but Ed
29:49
Sheeran takes the stand. He's very sympathetic. He
29:51
says, this is the death of all music. And the jury agrees with him,
29:53
and he wins. Total
29:55
coin flip. The facts actually,
29:58
in both cases, were the what they
30:00
needed to be for the outcomes in my opinion. But
30:02
they are straightforward fair use analyses. Why do
30:04
you think that's applicable or not applicable in
30:06
the case of opening eye? I don't think
30:08
it's applicable. The thing about the
30:10
blurred line skater, it's a case about vibes, right?
30:13
Does this sound vibe with Marvin Gaye? And
30:15
is that infringement? It's just
30:17
such a terrible case. Yeah, it's really bad. Yeah,
30:20
it's a really terrible, I mean, it's a good
30:22
example of the fact of like, this
30:24
is why you never want to go to trial for
30:26
literally anything. The court is random, right? The jury
30:29
is random and the decision is random. And then
30:31
eventually you end up in an appellate court with
30:33
a bunch of unelected weirdos and they're extraordinarily random
30:35
lately. This is why the CEOs think
30:37
it's just gonna be some money, right?
30:40
Because they're gonna say, look, I don't want to go to trial.
30:42
I doubt the New York Times wants to go to trial. Like
30:44
we'll just pay you a bunch of money and you'll go away
30:46
and we'll build our businesses under a
30:48
legal regime that exists now. Yeah, Getty,
30:50
The Times, the media, authors,
30:52
basically everyone who works in a
30:54
creative industry is very
30:57
mad and very concerned. And
31:00
they've seen sort of their bottom line eaten away
31:02
by big tech. And
31:05
this is like no longer an arrow where you
31:07
go, oh, well, we can form these partnerships with these companies.
31:09
And it'll work out for us in the long run. Like
31:12
people are tired of having, getting
31:14
their backstabs essentially. Like that's sort of
31:16
the perception, I think. I
31:18
think that like there is an appetite to take this
31:20
all the way to the end rather
31:22
than give in and
31:25
let things blow up the future.
31:28
And like whether or not that's like
31:30
a warranted feeling, maybe,
31:34
but it's like the feeling
31:36
is just different. So you don't think there are
31:38
settlements here? If you have nothing to
31:40
lose, then of course go to trial, right? If
31:43
you have literally nothing to lose. And I think like we've sort of hit
31:45
the wall where you go, oh, either
31:47
we go to court and we destroy copyright law
31:49
forever, maybe. Yeah. Or
31:52
we lose everything, right? Like I
31:54
think that we're like running up against the wall where
31:56
people are sort of like, oh, it's not just that
31:58
the robots are gonna take our jobs. the robots
32:00
are going to take like the
32:02
future literacy of humanity. Like it's
32:04
like people are starting to spin up these visions
32:06
of the future that are increasingly
32:09
apocalyptic. And I
32:11
think that there's like an appetite to take this to
32:13
the end. So I want to end there
32:15
on kind of a big thing. This
32:17
is an idea that I have ruthlessly stolen from you over
32:19
the years. It's the idea that
32:21
copyright law is the only real limiting regulation on the
32:24
internet. Because it's the only thing
32:26
that can consistently get things taken down. Like
32:28
everything else aside from child
32:31
pornography and sex trafficking to
32:33
some degree even. It's those
32:36
two things in copyright law. And
32:38
those are the regulating factors on the internet. Those
32:40
are things that you can send a letter
32:42
and something you can hand down if you claim it's copyright infringement.
32:45
And then you have this chaotic fair
32:47
use argument happening in the background. Like
32:49
the exit ramp is supposed to be
32:51
really flexible and you know
32:53
lend itself to these kinds of existential arguments
32:56
that might go either way in front of
32:58
a court. Entire industries might
33:00
live or die depending on how people are
33:02
feeling that day about the nature of the
33:04
New York Times. Is
33:06
this working? Like I, is
33:08
this the right way to go about it? Because it's
33:10
what we've been doing for a long time. And
33:13
I still don't know if
33:15
we've made any of the correct policy
33:17
choices using copyright laws are only real
33:19
tool. I think it's a
33:21
terrible tool to regulate
33:24
speech. It's clearly
33:26
not working out I think in the
33:29
context of like you know individual
33:31
creators. Like we've set up just
33:33
a very bizarre kangaroo court system
33:35
essentially through platforms. Everyone is
33:37
familiar with the idea of copyright
33:40
strikes and DMCA and everyone
33:42
sort of knows that it like it's a lot
33:44
of BS and doesn't work super well. Doesn't make
33:46
a ton of sense and it's weaponized.
33:50
That said it's like when you're looking
33:52
at sort of the changes that are
33:54
coming to the culture through generative AI
33:57
and what that poses for.
34:00
society and for the way we
34:02
live and you know all kinds
34:04
of things like how do we
34:06
learn in schools right the
34:08
nature of creativity itself the value
34:10
of literature and art I
34:13
like don't even know what how to quantify
34:15
the changes that are coming down the pipeline or
34:17
what to do to address them and historically
34:21
when you're looking at a technology
34:23
that's about to blow up culture itself bring
34:27
in the copyright actually like right like it's like
34:29
the printing press shows up you bring in sort
34:32
of proto-copyright like the stationers monopoly
34:34
right that you you bring in something
34:36
like copyright does it work super well
34:38
is it a good thing I don't
34:40
know kind of not
34:43
not totally I'm not 100% on
34:45
board but like yeah like traditionally what
34:47
we do when technology is about to
34:50
blow up culture is we bring on
34:52
something like copyright and so like I
34:55
don't know if that's the right tool for this because I
34:57
don't even know if we really understand what
35:00
generative AI is about to do to us but
35:02
I think it does make sense to me that
35:04
it's shown up at this time as
35:07
sort of the front line does it make sense
35:09
that it's shown up to you as sort of an extinction level
35:11
event for these companies it makes
35:13
sense to me in like oh
35:16
yes this is checkouts gun right moment
35:18
right like it makes sense and like
35:20
oh yes this was this was the
35:22
destiny of copyright law and the destiny
35:24
of generative AI but will
35:26
it be a good tool about
35:29
as good as anything else I think it's
35:31
like not great it's not super good but
35:33
like if I run my head
35:35
through anything else that we've got on the
35:37
books I don't think that there's like
35:40
something where I'm like oh yeah this isn't a copyright
35:42
thing this is a something else thing there's one I
35:44
can there's one I can think of actually yeah this
35:46
is a hint towards our next episode deepfakes
35:50
there's no copies there's some there's
35:52
a copy somewhere in the model but
35:55
then you you're looking at the deepfake of Taylor Swift
35:57
or Joe Biden or Donald Trump or whoever and it's
35:59
not a copy of anything. So
36:01
if those characters want to show up and say
36:04
take this down, they have to
36:06
use some other tool because they
36:08
can't just go to copyright law and say that you're not
36:10
authorized to use that photo of me. The
36:12
way that I don't know, even celebrity revenge
36:14
porn gets taken down because they own the
36:16
copyrights, the underlying images that gets stolen. Like
36:19
there's something else that needs to happen
36:21
for in particular deep fakes that
36:24
I don't think that we have an answer to yet either. It's like,
36:26
I don't know, we've had
36:28
senators on the show proposing new causes of
36:30
action around likenesses, which
36:32
just gets to it's just the other weird places.
36:35
But it's like everyone will have the same rights as any
36:37
celebrity to endorse or not endorse
36:39
Twitter. And it's like that is really weird. But
36:42
it feels like we're going to need that for the deep fake problem.
36:45
Yeah, that problem is just another
36:47
rat's nest because it makes copyright
36:49
look easy. Because
36:51
like once you get into sort of the deep
36:53
fake problem and likenesses. Oh,
36:56
man, like you didn't copyright socks. Wait until
36:58
you get to the right of publicity slash
37:00
the right of privacy, which is the same
37:02
thing depending on which state you're in. Amazing.
37:05
Well, that's a big hint towards the next episode. I got to
37:07
ask you though, just to wrap this one up. How
37:10
do you think New York Times versus open eyes gonna play out? Oh,
37:13
why would you ask me? Like it's
37:15
you're basically just setting me up for
37:17
like, whatever I answer is the wrong
37:19
thing. I think that this is one
37:21
of those things where I think the
37:23
most you can really hope for
37:25
is that whatever comes out doesn't
37:28
damage copyright law in a way
37:30
that makes it unworkable. Like
37:32
that is like the worst case scenario. The worst
37:34
case scenario isn't actually that generative
37:36
AI gets banned forever, or
37:39
that it gets a green
37:41
light forever. The worst case is that
37:43
copyright law changes in a way
37:45
that's unworkable. I really do
37:47
feel like the tension you've identified with
37:49
the CEOs of these companies continue to
37:51
make huge investments is that
37:53
they feel they can solve the problem with money. And I
37:56
think the tension is you actually have to solve
37:58
the problem. My prediction is
38:02
that if the time's not just
38:04
a big early victory, OpenEye will just throw money
38:06
at the times and make it
38:08
go away, right? They will just throw money until the
38:10
time says, fine, we'll do a 10
38:12
year deal. But that doesn't stop the author's guild
38:15
and that doesn't stop Getty. And
38:17
then suddenly it becomes too costly to run. I
38:20
think that first case, the trial case
38:22
actually determines a lot of what happens next because
38:26
the authors win there, the creators win in
38:28
any of those early cases, the
38:30
entire AI industry is going to light up with
38:32
settlement offers. And then the prices
38:34
are just going to rise and maybe that will determine what
38:36
happens next. But if the AI
38:38
companies win first, I do
38:40
think it's existential for all these creative companies and they
38:42
are going to fight tooth and nail until they do
38:44
get in front of a Supreme Court and
38:47
then all bets are off. And there's the also
38:49
the sort of thing where some
38:51
of these companies have a worst case than the others. So
38:54
whichever ends up in front of a court first is also
38:56
going to be interesting, I think some
38:59
of these companies have played fast and loose with
39:01
copyright a little more than others. And
39:05
you're right that the first test bullet, like
39:07
the first trial balloon is going to determine
39:09
a lot of what happens next. Thanks
39:14
again to Verge features editor Sarah Zhuang for joining us on
39:16
the show. I hope you can tell Sarah and I love
39:18
talking about this stuff. That was a lot of fun. From
39:21
now on, we're going to keep bringing
39:23
you second episodes of Decoder every Thursday
39:25
to deliver more analysis and storytelling like
39:28
this. In addition to our classic regular
39:30
weekly interviews with CEOs, lawmakers and automakers,
39:32
stay tuned for parts two and three of our
39:35
series over the coming week. Your
39:37
thoughts about this episode or what you'd like to hear
39:39
us talk about more. You can email us at decoder
39:41
at the verge.com. We really do read every email. You
39:43
also hit me up directly on threads. I'm at retches
39:45
to lady. Also, I'm a tick off team. Check it
39:47
out. It's at decoder. You
39:49
like to go to please share with your friends.
39:51
Subscribe wherever your podcast. If you really like the
39:53
show, hit us with that five star. I wouldn't
39:56
want to be second. I'm
39:58
just saying that. Here is a production of
40:00
The Verge and part of the Vox If
40:28
you're looking for ways to innovate your business, it might be
40:30
time to consider SAP Business AI.
40:33
With dozens of potential integrations to
40:35
optimize sales, procurement, finance, human resources,
40:37
and more, SAP Business AI may
40:39
be able to improve your business
40:41
operations inside and out. Revolutionary
40:44
technology? Real world results.
40:46
That's SAP Business AI. Learn
40:49
more at sap.com/AI.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More