Latest / a16z AI Policy Brief / AI's Data Access Question

Transcript

AI's Data Access Question · a16z AI Policy Brief

0:00
SPEAKER_00: welcome back to the a16z ai policy brief i'm matt perrault
0:04
SPEAKER_00: the web success was driven in part by a simple but powerful idea
0:08
SPEAKER_00: if information is lawfully available people should be able
0:11
SPEAKER_00: to access it read it learn from it and build on it that openness helped create
0:15
SPEAKER_00: the internet as we know it but in the age of ai that bargain is being tested
0:20
SPEAKER_00: In this episode,
0:21
SPEAKER_00: I'm joined by Derek Slater,
0:23
SPEAKER_00: a co -founder of Proteus Strategies and an expert on information access,
0:27
SPEAKER_00: copyright,
0:28
SPEAKER_00: and free expression for discussion about what we call the freedom to learn.
0:32
SPEAKER_00: We talk about how the early web balanced openness with publisher control,
0:36
SPEAKER_00: how technical standards like robots .txt helped manage those preferences at scale,
0:41
SPEAKER_00: and how today's debates over copyright contracts and data access could reshape
0:46
SPEAKER_00: who gets to build with AI.
0:48
SPEAKER_00: a16z
1:18
SPEAKER_01: they are reinforcing monopoly right limits and exceptions
1:21
SPEAKER_01: to copyright have always been a core anti-monopoly lever and i think
1:26
SPEAKER_01: when people miss that and you know start
1:29
SPEAKER_01: to say well we should just propertize everything they don't get how google
1:33
SPEAKER_01: and facebook are going to be the biggest winner
1:42
SPEAKER_00: derrick welcome to
1:43
SPEAKER_00: the ai policy brief great to be here thanks for having me
1:46
SPEAKER_00: so you are an expert on all things information access and you
1:50
SPEAKER_00: and i have known each other for a long time i've come to you over a long time
1:54
SPEAKER_00: on those topics free expression copyright so you're the perfect person i think
1:59
SPEAKER_00: to talk to us today about freedom to learn principles and the thing i want to start
2:03
SPEAKER_00: with is how the conversation that you and i had been having evolved to this topic
2:10
SPEAKER_00: so we were originally i think discussing what is the approach that
2:14
SPEAKER_00: a firm like AC Team Z should take.
2:38
SPEAKER_01: the great thing in this space is there are lots of different lenses we can apply
2:40
SPEAKER_01: but i do think the freedom to learn is attractive in two ways first i think one way
2:46
SPEAKER_01: of thinking about this issue is where do my rights as a reader a learner
2:51
SPEAKER_01: a consumer somebody who's using a book music a website whatever
2:57
SPEAKER_01: where do my rights begin in the publisher's rights end right and what should i get
3:01
SPEAKER_01: to do with the thing that i have lawful access to that i bought and so on and
3:04
SPEAKER_01: so forth
3:05
SPEAKER_01: that's both sort of you know literally how can i access it can i take a website and turn on
3:11
SPEAKER_01: read aloud function and make it more accessible but i think also this other layer
3:17
SPEAKER_01: a16z
3:19
SPEAKER_01: a16z
3:41
SPEAKER_01: ii want there to be this commons of knowledge that belongs to all
3:46
SPEAKER_01: of us ideas facts lots of stuff that we rely on all day the way language works
3:51
SPEAKER_01: where we all rely on the fact that you can't get a property right and
3:54
SPEAKER_01: then say no you can't do this with
3:57
SPEAKER_01: that you can't use my fact you can't use my idea it belongs to everybody
4:01
SPEAKER_01: and that's i think something that's really important to be reckoning with i think
4:04
SPEAKER_01: a lot of these issues can so easily be taken into there's this company
4:09
SPEAKER_01: and this company who are fighting
4:11
SPEAKER_01: that's true there's a lot of that and there's
4:13
SPEAKER_01: a lot to tease out there
4:15
SPEAKER_01: and we need to step back and think about what are the things
4:18
SPEAKER_01: that should belong to each of us into all
4:20
SPEAKER_01: of us together
4:20
SPEAKER_01: and that's where public policy really has a strong role
4:23
SPEAKER_00: so can you walk through some other venues where we confront some
4:27
SPEAKER_00: of these conversations like as you're describing i think a lot of people come into
4:30
SPEAKER_00: the ai copyright debate with some preconceived notions based
4:35
SPEAKER_00: on sometimes like do they like one side of the debate or the other like it's
4:38
SPEAKER_00: so crazy that companies would use copyrighted work as training or it's so crazy
4:43
SPEAKER_00: that doing this really innovative thing of training
4:46
SPEAKER_00: on publicly available copyrighted work would not be fair use
4:49
SPEAKER_00: and i wonder if it's helpful if we could just like step back a little bit and look
4:53
SPEAKER_00: at how for something like the web
4:55
SPEAKER_00: for instance like how do these general principles operate because that's a that's
4:59
SPEAKER_00: a world where i think like some of these things have been more resolved and we have
5:02
SPEAKER_00: a sense of like what are the rights that should belong to publishers and what are
5:06
SPEAKER_00: the rights that should belong to people who are roaming around the web
5:09
SPEAKER_01: going back to the beginning of the web is good and there are other technologies too
5:12
SPEAKER_01: and i think the key thing here is that even though you know things can look like on the web
5:18
SPEAKER_01: things have been working in a certain way and they've always worked their way actually at
5:21
SPEAKER_01: the beginning things were very contested when it comes to what does it mean to have
5:25
SPEAKER_01: freedom to learn or to have an open web what does that mean so i think it really came down
5:29
SPEAKER_01: i found it two things that went on early days of the web so early days people created websites
5:36
SPEAKER_01: the web was a standard that was open what did that mean anybody could build a browser
5:41
SPEAKER_01: or piece of software to go and access your site and render it in whatever way was best for them
5:47
SPEAKER_01: you know if that meant blowing up the font if that met reading it aloud if that
5:51
SPEAKER_01: meant doing it in different colors you as the reader had that control because of
6:15
SPEAKER_01: a16z
6:30
SPEAKER_01: is the home for robots.txt the small text file anybody can put on their website
6:35
SPEAKER_01: and say hey you know you can crawl this part but not that part
6:38
SPEAKER_01: and there are other standards they're like well you can crawl my site but only at
6:42
SPEAKER_01: the speed that works for me
6:43
SPEAKER_01: because i only have this one server not 50 million servers i don't have
6:47
SPEAKER_01: a data center
6:48
SPEAKER_01: and those things had to evolve over time
6:50
SPEAKER_01: and i think there were mutual benefits that led them
6:54
SPEAKER_01: but it's not worth you know
6:55
SPEAKER_01: it's worth you no
6:56
SPEAKER_01: losing sight of
6:57
SPEAKER_01: these are contested
6:58
SPEAKER_01: these were fought over
6:59
SPEAKER_01: or rather
7:00
SPEAKER_01: there was both fights in the standards processes about that
7:03
SPEAKER_01: there was
7:04
SPEAKER_01: legal threats and litigation
7:06
SPEAKER_01: both about search engines
7:08
SPEAKER_01: and about things like deep linking
7:28
SPEAKER_01: and so there are always these tussles that go on when you have this sort
7:31
SPEAKER_01: of interface between publisher reader you know software and website and so one way
7:36
SPEAKER_01: that these things got reconciled was norms and standards like at the IETF and part
7:40
SPEAKER_01: of it was law and the balance that copyright has helped set which is say
7:45
SPEAKER_01: when you know search engines were sued in the sort of early days of
7:49
SPEAKER_01: the web late 90s early 2000s court said look you own this expression sure
7:56
SPEAKER_01: but what this thing is doing is transformative it's analyzing this expression
8:01
SPEAKER_01: to derive facts and relationships and things that aren't copyrightable
8:07
SPEAKER_01: and it shows this little snippet of text but that's so small that it
8:11
SPEAKER_01: a16z
8:35
SPEAKER_01: courts had to sort a lot of that out and and you know as we see today are continuing
8:38
SPEAKER_01: to wrestle with this not just an ai but with other sorts of cases
8:42
SPEAKER_01: where you know someone wants to access data on a site and
8:46
SPEAKER_01: the site you know doesn't like it
8:48
SPEAKER_01: to give another example you know facebook was faced with and linkedin too
8:53
SPEAKER_01: and other social media sites faced some service providers that wanted
8:56
SPEAKER_01: to help customers want to help users consumers
9:00
SPEAKER_01: a16z
9:18
SPEAKER_01: a16z
9:41
SPEAKER_00: so robust ad text seems like an interesting effort at a compromise i'm wondering
9:44
SPEAKER_00: if you can talk about
9:45
SPEAKER_00: a couple different components of it so one is
9:48
SPEAKER_00: the
10:17
SPEAKER_01: yeah i mean i think from the site owner's perspective you know i think the piece
10:21
SPEAKER_01: of this is control and then also i think what a lot of publishers would describe
10:24
SPEAKER_01: as maybe value exchange or who gets the value from this right so what's the bargain
10:29
SPEAKER_01: if you're coming with your crawler what do i get
10:31
SPEAKER_01: and what do you get okay you get data you get to create your search index
10:35
SPEAKER_01: iii
10:54
SPEAKER_01: a16z
11:09
SPEAKER_01: if at the beginning of things you know yahoo or other search engines had had to go
11:14
SPEAKER_01: and negotiate individually with each site and ask permission first and negotiate
11:19
SPEAKER_01: a separate contract first that never would have scaled to the you know billions
11:22
SPEAKER_01: of websites that come out
11:25
SPEAKER_01: so robots.txt both allowed people to just sort of continue operating as normal and
11:30
SPEAKER_01: then when somebody puts up a hand and says actually i have a different preference
11:33
SPEAKER_01: a simple way to respect it and i think the fact that it was simple machine readable
11:39
SPEAKER_01: and sort of fit within the broader construct of like okay we if you don't want
11:42
SPEAKER_01: to be called
11:43
SPEAKER_01: i don't like this traffic fine it's sort of you know landed as a reasonable norm not in
11:49
SPEAKER_01: and to your point to just underscore yes it's a norman that means sure there are people who
11:54
SPEAKER_01: will not respect it and some of them you know may be less
11:58
SPEAKER_01: sympathetic you know bad actors to some mind others it's for you know good reasons so
12:03
SPEAKER_01: look at the you know the internet archive there are some types
12:07
SPEAKER_01: of sites certain you know government
12:08
SPEAKER_01: sites i think in particular where they might not respect robots.txt because they're an archive
12:14
SPEAKER_01: they're a library their whole job is capturing the historical record and you know libraries
12:20
SPEAKER_01: have a job to do too that's very different from say a commercial search engine or ai or
12:25
SPEAKER_01: some other sort of data collector.
12:26
SPEAKER_00: so you used to work for
12:28
SPEAKER_00: a small little commercial search engine called google why would
12:32
SPEAKER_00: that search engine respect robots.txt it's a great question
12:37
SPEAKER_01: and i think we could sort of characterize it over time if i roll
12:40
SPEAKER_01: the clock back i would say that it was in recognition of this need to reach
12:46
SPEAKER_01: an equilibrium of sorts right that is
12:50
SPEAKER_01: ii take the bright line hardwall all or nothing
12:53
SPEAKER_01: that could end up worse people start locking up their sites
12:56
SPEAKER_01: behind more restrictive things and paywalls right that would sort of be
12:59
SPEAKER_01: the inevitable end and that's sort of what we see today of okay
13:02
SPEAKER_01: if you let's imagine they decided not to
13:05
SPEAKER_01: tx then people put stuff behind paywalls or behind logins
13:09
SPEAKER_01: and that's even worse right so it's about finding that equilibrium
13:12
SPEAKER_01: that mediates this relationship so
13:14
SPEAKER_01: that there are many winners it's not just you know the search engine wins
13:17
SPEAKER_01: the publisher loses now i think today
13:20
SPEAKER_01: yon of the questions and now this is you know why do they follow it
13:23
SPEAKER_01: now there are ways in which this is more of a regulatory issue for company like
13:27
SPEAKER_01: google. There are more obligations for them
13:29
SPEAKER_01: in part because that's different, they have a different amount of power than in
13:32
SPEAKER_01: those days but I think if you go to those early days
13:35
SPEAKER_01: a16z
14:04
SPEAKER_01: you know rights holders come forward and say hey i don't want my book even indexed
14:08
SPEAKER_01: and google would respect it that wasn't required by the law it wasn't required by
14:13
SPEAKER_01: the fair use decision that the court ended up rendering and because google is
14:18
SPEAKER_01: in repeat relationships with publishers
14:20
SPEAKER_01: kind of makes sense to do something that strikes a medium recognizes
14:23
SPEAKER_01: that both sides have an interest in making this work well
14:26
SPEAKER_01: so that's what things look like
14:27
SPEAKER_01: when they're working well ish now again there was litigation over
14:30
SPEAKER_01: the web there was litigation over google book search by the authors guild and so on
14:34
SPEAKER_01: and so forth so none of this is easy and i think that's still in the background
14:38
SPEAKER_01: of how do we have something that leads to many winners rather than just
14:41
SPEAKER_01: a winner take all situation
14:43
SPEAKER_00: so we've talked a lot about the freedom to learn in the context of tech universe in
14:49
SPEAKER_00: the context of development of the web but can we back up even further and talk
14:53
SPEAKER_00: about freedom to learn generally how do these principles govern how humans learn in
14:56
SPEAKER_00: the world
14:57
SPEAKER_01: you know going back to where we started of where you know remember buying a book
15:02
SPEAKER_01: a physical book i don't know if you still do that i still do it from time
15:05
SPEAKER_01: to time you actually get a set of rights when you have that book
15:09
SPEAKER_01: the person can't say to you you're not allowed to give this away
15:12
SPEAKER_01: or you're not allowed to sell it
15:13
SPEAKER_01: to somebody else right that's how we have libraries right they're not allowed to do
15:18
SPEAKER_01: that with the book you can quote from it you can generate summaries generally to
15:23
SPEAKER_01: a point you can learn from it in the sense of
15:25
SPEAKER_01: elizabeth gunn the author's style they write in this way or these are the themes they're using
15:30
SPEAKER_01: but this is how they structure the novel you can learn facts right you buy a book
15:34
SPEAKER_01: on world war ii
15:35
SPEAKER_01: get a bunch of facts from them you can reuse those you could also take that physical book and
15:56
SPEAKER_01: i think what we're seeing now is a you know debate over how do these translate to ai
16:00
SPEAKER_01: and i would say generally the parallel is there i think people get tripped up over
16:05
SPEAKER_01: the idea of well humans are allowed to learn but this is about machines using
16:09
SPEAKER_01: the stuff and machines don't have rights okay fine but i'm the one using the machine
16:15
SPEAKER_01: i should have the right to use a machine to help me learn in ways that are in line
16:21
SPEAKER_01: with law and norms and so on so forth right and that's what's happening
16:24
SPEAKER_01: with ai models it's you know i could watch and have watched you know 30 sci-fi films
16:29
SPEAKER_01: and try to figure out the different concepts and themes
16:32
SPEAKER_01: george lucas can draw on kurosawa and then people can draw
16:35
SPEAKER_01: on george lucas right all that's your game if i use you know an algorithm a machine
16:40
SPEAKER_01: to do that same sort of pattern matching
16:41
SPEAKER_01: and learning we call it data science it doesn't change that i'm
16:44
SPEAKER_01: the one executing it i'm the one learning i think that's something
16:47
SPEAKER_01: that we should be celebrating people's ability to do that and to use that freedom
16:52
SPEAKER_01: to learn in new ways with machines as their assistants
16:55
SPEAKER_00: can you talk a little bit about how the arguments that publishers are making on
16:58
SPEAKER_00: the other side like they understand and recognize freedom
17:01
SPEAKER_00: to learn principles generally as well what is the concern for them
17:05
SPEAKER_00: about ai training how that relates to these principles yeah i think there are sort
17:09
SPEAKER_01: of three buckets of concerns and the way people have come
17:14
SPEAKER_01: at this i think one is really around traffic traffic to websites and the cost
17:20
SPEAKER_01: of traffic this is less i think with large commercial publishers which
17:24
SPEAKER_01: you know they can scale their servers and it's probably a rounding error but you
17:28
SPEAKER_01: don't mean to say that generally i've seen
17:30
SPEAKER_01: as a bigger issue for other types of publishers like wikimedia right
17:34
SPEAKER_01: wiki media although content or
17:36
SPEAKER_01: they say our content is free our infrastructure
17:38
SPEAKER_01: not
17:39
SPEAKER_01: a16z
18:04
SPEAKER_01: then how are we going to get new users and new members and new contributors and then
18:07
SPEAKER_01: that hurts our sustainability and that you know you can think of that
18:11
SPEAKER_01: as another cost so that's one slice of things and there i think there are a lot
18:16
SPEAKER_01: of important things to be done of okay how do we like we did with robots.txt
18:21
SPEAKER_01: and other standards create the right ways of saying hey if you're going to run
18:25
SPEAKER_01: a bot it should behave kind of like this so it doesn't crash my server
18:29
SPEAKER_01: and you know if you would can you validate with this third party that you're
18:32
SPEAKER_01: a good bot not a bad bot and show me a credential well there are lots
18:36
SPEAKER_01: of interesting ideas like that we can play with but that's one which is really
18:39
SPEAKER_01: about like traffic costs and then that merges in this competitive one and this is
18:43
SPEAKER_01: where i think you're going with the publishers and this and the value exchange
18:47
SPEAKER_01: the bargain wait
18:49
SPEAKER_01: a second you're taking my content you're creating these trillion dollar companies
18:55
SPEAKER_01: and not only am i not necessarily getting some you know the benefit of
18:59
SPEAKER_01: that trillion dollars you're making a dollar i'm not even getting a dime
19:03
SPEAKER_01: and i should get a dime but also wait a second
19:07
SPEAKER_01: the fact you're displaying these facts and now if i'm a newspaper
19:11
SPEAKER_01: and i'm expecting people to come to me for the facts
19:13
SPEAKER_01: and you're taking these uncopyrightable facts and showing them they're not going
19:16
SPEAKER_01: to come to my site and then i lose traffic and i lose ad revenue and so on and
19:19
SPEAKER_01: so forth
19:20
SPEAKER_01: there's a real competitive angle there and you know it's to sort of round it out
19:25
SPEAKER_01: because it's easy to just think about news publishers it's a broader issue
19:28
SPEAKER_01: with web publishers right just last week amazon
19:32
SPEAKER_01: and perplexity i think there's another hearing in that litigation where amazon
19:36
SPEAKER_01: as sued perplexity for giving users about an agent that at
19:41
SPEAKER_01: the user's direction will do will navigate the amazon site
19:45
SPEAKER_01: and help them do their shopping so that instead of dealing with the amazon listings
19:51
SPEAKER_01: which prioritize ads and things that might maximize their profit
19:55
SPEAKER_01: it's as a duty to the user do what the user wants and amazon said i don't like
20:00
SPEAKER_01: that i want to control the interface of my site and use
20:03
SPEAKER_01: that same anti-hacking statute that i mentioned earlier and
20:06
SPEAKER_01: so that's another case like the news publishers where they're saying i want control
20:10
SPEAKER_01: over how my content is displayed you're competing with me you're taking some of
20:15
SPEAKER_01: that value
20:16
SPEAKER_01: i think some of these disputes it's about finding that equilibrium
20:20
SPEAKER_01: the win-win i mean like at some level backup
20:22
SPEAKER_01: from amazon perplexity those two companies like if you're a commerce site
20:26
SPEAKER_01: a shopping site and you're
20:27
SPEAKER_01: a commerce agent you have some shared interests commerce site wants
20:30
SPEAKER_01: to sell stuff you want to help consumers buy stuff there's reasons
20:34
SPEAKER_01: to come together make agreements make apis come up with standards
20:38
SPEAKER_01: i think we're starting to see some of that with news publishers too of how
20:42
SPEAKER_01: that value exchange happens as much as the ai developers think yes
20:46
SPEAKER_01: if your information is public you allow me to crawl it i can use it
20:50
SPEAKER_01: for training i can use at least purposes and if you have
20:54
SPEAKER_01: a paywall well you know you can say to an ai developer you have to pay me to acquire
20:59
SPEAKER_01: that content even if it's fair use to train on it or
21:02
SPEAKER_01: y
21:03
SPEAKER_01: developers have said we will respect certain sorts of opt outs you know like
21:06
SPEAKER_01: the search engines if
21:08
SPEAKER_01: want to opt out a foundation model training or you know use
21:11
SPEAKER_01: in these things like ai overviews as
21:13
SPEAKER_01: google announced last week we'll respect those two so we're seeing us see that
21:17
SPEAKER_01: a16z
21:40
SPEAKER_01: i get when people look at the equities of ai companies and newspapers say this is
21:45
SPEAKER_01: a fight over money and they should contribute to the publishers
21:48
SPEAKER_01: the equities feel very different to me when it's a publisher saying sure you paid
21:53
SPEAKER_01: for access to my
21:55
SPEAKER_01: but i want you to read it in this particular way and i'm actually gonna i'm out of
22:00
SPEAKER_01: a technical control that blocks you from using these from using an assistant
22:04
SPEAKER_01: to help you read it right so in my browser today
22:07
SPEAKER_01: and many other browsers there are ai tools you know that will
22:10
SPEAKER_01: let you summarize a page should a website publisher be able to tell you the user no
22:15
SPEAKER_01: you may not summarize you must read the full article the way i wrote it i mean we could have
22:20
SPEAKER_01: world that works that way it wouldn't be the freedom to learn that either serves the individual
22:25
SPEAKER_01: and i don't think really serve society ultimately now you know there are lines
22:29
SPEAKER_01: to be drawn here but
22:30
SPEAKER_01: i think that angle is really important to sort of put in the front can we run
22:34
SPEAKER_00: through how existing
22:36
SPEAKER_00: in law applies here
22:37
SPEAKER_00: in your overview you're talking about of several
22:40
SPEAKER_00: different
23:01
SPEAKER_01: yeah i mean it certainly varies around the world and governments around
23:04
SPEAKER_01: the world are looking at how to amend change reconcile ai
23:07
SPEAKER_01: with their copyright law you know just starting in
23:10
SPEAKER_01: the us i think there are it's very important to separate inputs from outputs that is
23:15
SPEAKER_01: to say training a model
23:18
SPEAKER_01: oi public information,
23:20
SPEAKER_01: you know,
23:20
SPEAKER_01: lawfully accessed either a website or information online elsewhere, or,
23:24
SPEAKER_01: you know,
23:25
SPEAKER_01: you go and you get a book and you scan it as as anthropic and others have done.
23:29
SPEAKER_01: That's the input side.
23:31
SPEAKER_01: And if you're using that to train a model,
23:33
SPEAKER_01: which means to derive those uncopyrightable ideas and facts and so on,
23:37
SPEAKER_01: and not output the expression.
23:39
SPEAKER_01: then that's consistent with that's highly transformative it's not substitutional for
23:44
SPEAKER_01: the original work it helps people create new works but it's not replicating
23:48
SPEAKER_01: the old work there are corner cases here like what if the model memorizes stuff what
23:52
SPEAKER_01: if it's a copyrightable character
23:54
SPEAKER_01: a16z
24:17
SPEAKER_01: as countries around the world think about both sovereignty ai sovereignty and
24:22
SPEAKER_01: about what does it mean to have competition and choice across all the layers of
24:26
SPEAKER_01: the ai stack they're going to have to come to grips with the idea
24:30
SPEAKER_01: that they will need some sort of way of doing you know allowing this sort of use
24:36
SPEAKER_01: of public lawfully access content for training on the input side
24:40
SPEAKER_00: so the input output distinction seems really important so
24:43
SPEAKER_00: the input side is you're using copy publicly available copyrighted work to train
24:47
SPEAKER_00: a model and then the outputs of the model might look nothing like
24:51
SPEAKER_00: or certainly not exactly like the underlying work can you say a little bit more
24:55
SPEAKER_00: about the specific use case you have in mind in the output side but what would
24:59
SPEAKER_00: a potential infringing output look like
25:01
SPEAKER_01: yeah i mean start with the non-infringing because i think there's so much that is
25:05
SPEAKER_01: about particularly the the large language models that has been bound up in copyright
25:09
SPEAKER_01: and certain types of creators and that's fair enough and those objections
25:15
SPEAKER_01: and those rights holders are vanishingly small portion of all rights holders
25:19
SPEAKER_01: and creators talking about a vanishingly small percentage
25:23
SPEAKER_01: of uses most people are not using their llms to go and try and recreate a
25:29
SPEAKER_01: an article to spit out a new york times article they're creating new stuff
25:31
SPEAKER_01: as you said they're writing their own stuff they're you know they're getting
25:34
SPEAKER_01: an educational tutor they're doing scientific research right all sorts of sectors
25:39
SPEAKER_01: across the economy are using these large language models for things
25:43
SPEAKER_01: that have nothing to do with creative works and even when they do have to do
25:46
SPEAKER_01: with creative works it's generally about creating a new a novel thing okay
25:50
SPEAKER_01: so how would it be infringing
25:52
SPEAKER_01: let's say the model did memorize a full work the expression of a work and
25:57
SPEAKER_01: then outputted it right which is alleged in well some of the newspaper cases
26:01
SPEAKER_01: but also in the case brought by music publishers against anthropic
26:05
SPEAKER_01: who have said your model memorizes full song lyrics and outputs them readily
26:11
SPEAKER_01: a16z
26:26
SPEAKER_01: likely to memorize did they use any output filters how did that look
26:29
SPEAKER_01: that might be part of it well let's remember but that's all still that's still on
26:32
SPEAKER_00: the on the input inquiry right infringing output would be cognizable
26:37
SPEAKER_01: under copyright law if it outputted that article the question though is is
26:41
SPEAKER_01: the ai model developer responsible for that and i just i bring that up
26:44
SPEAKER_01: because look remember google books
26:47
SPEAKER_01: is a database of books it's like it is a it has memorized those books
26:52
SPEAKER_01: but there are output filters that limit what it outputs so in
26:55
SPEAKER_01: the google books case there was both us you know scanning them but also creating
26:59
SPEAKER_01: the index and the output and they're all you know different but they have
27:02
SPEAKER_01: another one that may be clearer is that it's sort of again the iron man example let's say i
27:06
SPEAKER_01: say to my image generator create a movie that is you know avengers doomsday
27:12
SPEAKER_01: my own version of it and then you go in you sell it I think because characters can
27:18
SPEAKER_01: be copyrightable there will be questions about those specific outputs and then what
27:22
SPEAKER_01: are the developers responsibility they're going to have to work through that
27:25
SPEAKER_01: my own perspective on it is
27:27
SPEAKER_01: it's going to come down to questions
27:29
SPEAKER_01: not of you know what's called direct
27:31
SPEAKER_01: liability but secondary
27:33
SPEAKER_01: liability that is the user is the one
27:35
SPEAKER_01: making the thing they're
27:37
SPEAKER_01: the one who type the prompt hit and
27:56
SPEAKER_01: i think professor matthew sag is one of the leading writers on on this topic wrote
28:02
SPEAKER_01: a really interesting article about you know fair use
28:04
SPEAKER_01: and ai's jagged frontier saying this uncertainty in these issues means
28:09
SPEAKER_01: that there's still an incentive to cut certain types of deals not necessarily for
28:13
SPEAKER_01: the input side but for this output side and that reminds me
28:17
SPEAKER_01: a lot reminded professor sag a lot of youtube right where youtube in the early days
28:24
SPEAKER_01: got blown stuff there's a notice and takedown process but they went further
28:27
SPEAKER_01: and said hey instead of just notice and takedown what if we add this option
28:31
SPEAKER_01: where we leave this stuff up and we run ads against it and you the rights holder get
28:35
SPEAKER_01: the majority of the money and that created a new revenue stream forever
28:39
SPEAKER_01: it was that win win win the user gets to leave this stuff up
28:42
SPEAKER_01: the rights holder gets money youtube makes money too and that keeps going
28:47
SPEAKER_01: and i think those sorts of things are i think on the output side part
28:50
SPEAKER_01: of what we'll see develop love some protection but there will also be areas
28:54
SPEAKER_01: that are clearly off limits if you're really outputting something that is
28:58
SPEAKER_01: an infringing work and commercializing it and so on and so forth
29:01
SPEAKER_00: let's talk about contract law which i have found to be really complicated
29:04
SPEAKER_00: so there's all the copyright related stuff which i think is probably like
29:09
SPEAKER_00: the headline issue here but then increasingly rights holders have said
29:14
SPEAKER_00: and i see this almost every time i pick up an ebook the book will say like
29:18
SPEAKER_00: the content of this book cannot be used in ai training any form which is a
29:22
SPEAKER_00: attempt to impose a unilateral contract and
29:24
SPEAKER_00: to hold any potential ai developer liable if they violate that provision
29:30
SPEAKER_01: in ebook how should we think of that that's a great question and a really tricky one
29:35
SPEAKER_01: because i think one of the key questions that gets out is when should you get to do
29:39
SPEAKER_01: if ever through contract what you can't do through copyright right
29:42
SPEAKER_01: so you know go back if i buy a book in the bookstore
29:46
SPEAKER_01: they can't simply put something in the book that says you may not resell and make
29:50
SPEAKER_01: that a binding contract and you know copyright would preempt that you know similar
29:54
SPEAKER_01: if they said no quoting you know no way and you know think
29:57
SPEAKER_01: about it newspapers could never live in that world newspapers are
30:01
SPEAKER_01: the greatest beneficiaries in many ways of copyrights limitations and exceptions
30:06
SPEAKER_01: that they can reuse facts they get from third parties they can quote
30:08
SPEAKER_01: from third parties as fair use right so
30:11
SPEAKER_01: a16z
30:11
SPEAKER_01: that's something we've always relied on and
30:13
SPEAKER_01: we're seeing more and more attempts where these
30:15
SPEAKER_01: things are coming to conflict where
30:17
SPEAKER_01: you know sites are saying in their terms of service
30:20
SPEAKER_01: you may not crawl my site
30:21
SPEAKER_01: to do x y and z there have been
30:23
SPEAKER_01: a set of cases that have started to reckon with
30:25
SPEAKER_01: a16z
30:41
SPEAKER_01: i'm sort of riffing on xv bright data which is a case of you know about x and
30:46
SPEAKER_01: then bright data is a data collector that helps with a lot
30:49
SPEAKER_01: of different data analysis you know there was a set of cases like that
30:53
SPEAKER_01: and this is still like a moving target and
30:57
SPEAKER_01: i think it's still very much evolving and there are these other tools that pop up
31:01
SPEAKER_01: that are contract but then are used to sort of accomplish the same thing so
31:05
SPEAKER_01: with your ebook they don't just put like a contract in front of you
31:09
SPEAKER_01: that says you may not give this away they use digital rights management digital
31:14
SPEAKER_01: locks that prevents you and if you were to circumvent those you could be liable and
31:18
SPEAKER_01: if somebody creates a tool that helps you circumvent they could be liable
31:22
SPEAKER_01: similarly going back to amazon
31:23
SPEAKER_01: and perplexity right i mean there are contract elements there but the core claim
31:27
SPEAKER_01: from amazon is you're hacking my site you're violating this anti-hacking law
31:32
SPEAKER_01: and both of these sorts of claims are being used you know
31:35
SPEAKER_01: in some cases news publishers have said robots.txt is the same as
31:41
SPEAKER_01: that technological protection measure that's on your ebook and so
31:46
SPEAKER_01: if we say no ai training in our robots.txt then you shouldn't be able
31:50
SPEAKER_01: to work all these things we put in our terms of service that's
31:53
SPEAKER_01: a technological protection all these sorts of things so a lot
31:56
SPEAKER_01: of claims are being thrown at the law but i think it comes back to what rights stay
32:00
SPEAKER_01: with the public
32:01
SPEAKER_01: a16z
32:20
SPEAKER_01: yeah and i think you know the main i think that is sort of tried and true
32:24
SPEAKER_01: and holds here rather there are allegations in various you know cases
32:28
SPEAKER_01: about circumvention of a paywall and so on and i think all of
32:33
SPEAKER_01: the major ai developers
32:35
SPEAKER_01: i would say that they respect paywalls login walls they do not try
32:40
SPEAKER_01: and circumvent those go around
32:41
SPEAKER_01: those now if the user has credentials to the site right i have a new york times login and
32:47
SPEAKER_01: subscription and then i use my ai well that's it
32:50
SPEAKER_01: i don't think that is the real challenge here
32:56
SPEAKER_01: and that's why we do see deals evolving for access to data i think the
33:02
SPEAKER_01: the you know people have ways of putting their data behind walls
33:06
SPEAKER_01: and locking it up already and then you know arranging with ai developers you know
33:13
SPEAKER_01: if you want to acquire this here's what it costs
33:15
SPEAKER_01: what's missing in that is it the vast majority of all works
33:18
SPEAKER_01: that have ever been created are not managed
33:20
SPEAKER_01: in any way right there are like over a billion websites most of them are not
33:23
SPEAKER_01: managed but still up there you know books from the 20th century
33:28
SPEAKER_01: most are not manage or licensed in any
33:30
SPEAKER_01: a16z
33:47
SPEAKER_01: right this is what came up in the anthropic case and i think that case
33:52
SPEAKER_01: in my reading took a narrower line and looked at okay it wasn't just the acquisition
33:59
SPEAKER_01: of the books but you kept the books and maybe you didn't secure them well enough
34:03
SPEAKER_01: and that's different from google books
34:04
SPEAKER_01: which put really good security controls around its books other people read
34:08
SPEAKER_01: that case differently and say no you are not allowed to use piloted material and
34:13
SPEAKER_01: if you had you know
34:15
SPEAKER_01: you know knowledge of some sort that this was from a pirate site don't use it and
34:21
SPEAKER_01: so there are all
34:21
SPEAKER_01: these now tons of lawsuits about the books 3 database libgen these other databases that
34:26
SPEAKER_01: you know were up there on the web just like any other website
34:30
SPEAKER_01: but people are saying oh those should have been off limits i think that's going
34:33
SPEAKER_01: to be a trickier issue because i i don't i don't see the easy distinction
34:38
SPEAKER_01: between these books that are publicly available in a you know a bit tort file and
34:43
SPEAKER_01: an html file that's sort of an artificial distinction
34:45
SPEAKER_00: so you kind of mapped out like the issues the underlying bodies of law
34:50
SPEAKER_00: that apply looking forward in in the piece we wrote
34:54
SPEAKER_00: about market evolution technical collaboration and public policy guidance
34:59
SPEAKER_00: as like three ways to think
35:00
SPEAKER_00: about addressing various different components here you've kind of teased each
35:03
SPEAKER_00: of these i think in your comments let's go into them a little bit in more detail
35:07
SPEAKER_00: so market evolution what do you think are the approaches there that would have
35:13
SPEAKER_00: delizabeth gunn
35:41
SPEAKER_01: a16z
35:42
SPEAKER_01: sincron acquired in that way and then the sort of the sum of human knowledge
35:45
SPEAKER_01: and history you know
35:46
SPEAKER_01: everything is copyrighted as soon as it's you know fixed in
35:49
SPEAKER_01: a tangible medium you write a letter
35:51
SPEAKER_01: you can write some letters on a napkin copyright and nobody's licensing
35:55
SPEAKER_01: and managing all of those
36:11
SPEAKER_01: create something
36:12
SPEAKER_01: that wouldn't be possible otherwise you know just under a fair use and whatnot you know i think
36:17
SPEAKER_01: while sora itself the open ai product is gone that was the nature of the deal
36:22
SPEAKER_01: that they cut with
36:23
SPEAKER_01: disney let's say okay we can use disney characters and outputs in this way that
36:27
SPEAKER_01: otherwise we might not be able to i imagine a lot of things like that will evolve over time so
36:32
SPEAKER_01: that's i think part of the business model evolution i think there will be other things
36:36
SPEAKER_01: that involve with publishers that could involve more of the sort
36:40
SPEAKER_01: of real -time information for sort
37:06
SPEAKER_01: Those markets would not be existing, but for fair use.
37:10
SPEAKER_01: That is the ability to use stuff for free under fair use enabled this new market
37:16
SPEAKER_01: to be created around AI, which is leading to new revenue sources.
37:20
SPEAKER_01: That is a tale as old as time in media and information.
37:23
SPEAKER_01: It's what we saw with the early internet platforms
37:26
SPEAKER_01: and search engines is what we sell with the VCR.
37:30
SPEAKER_01: a16z
37:47
SPEAKER_01: a16z
38:05
SPEAKER_00: what about technical collaboration i think the the main concept there is thinking
38:10
SPEAKER_00: about ways to implement some version of robot.txt for an ai world
38:16
SPEAKER_00: and i think there are considerations there that are most important to the
38:19
SPEAKER_00: a part of the tech ecosystem that we're focused on little tech which is you talked about cases
38:24
SPEAKER_00: like a provider burying a do not train provision in terms of service that obviously seems
38:30
SPEAKER_00: like it's the kind of thing that for small companies with small legal teams it's going
38:34
SPEAKER_00: to be really hard for them to implement and practice so what are the ways that
38:37
SPEAKER_00: you would think about making robots.txt workable for an ai ecosystem but also worked
38:44
SPEAKER_01: tomorrow there was a meeting of the internet engineering task force ai preferences working
38:49
SPEAKER_01: a16z
38:50
SPEAKER_01: which for the last year and a half has been working to wrestle with this question
38:56
SPEAKER_01: of how might we create a standardized vocabulary
39:02
SPEAKER_01: about different sorts of ai use cases foundation model training decisions about
39:07
SPEAKER_01: whether or not to be included in search versus you know ai overviews that are
39:12
SPEAKER_01: in you know google search what to do about other sorts of inference
39:16
SPEAKER_01: or use you know should i be able to tell a browser you may not summarize this for
39:20
SPEAKER_01: a user even if it's the user clicking
39:22
SPEAKER_01: the summarize button those have all been live issues and on
39:26
SPEAKER_01: the one hand standards work is really hard and frustrating and if you go through
39:29
SPEAKER_01: the mail list it's a mess
39:33
SPEAKER_01: i'm optimistic we're getting somewhere on the core issues of training
39:47
SPEAKER_01: ai providers and civil society it's not perfect but i i do think we can get there
39:52
SPEAKER_01: and i think i've seen a lot of willingness to arrive at that
39:56
SPEAKER_01: as you were saying machine readable standardized in ways everybody can work with
40:02
SPEAKER_01: a16z
40:22
SPEAKER_01: so there's lots of other stuff going there it is tough and
40:25
SPEAKER_01: with all these standards process no one is ever like perfectly happy
40:28
SPEAKER_01: and it's left me with optimism that at least on the training and
40:32
SPEAKER_01: the search side we can start to square that circle we've been talking
40:36
SPEAKER_00: about this issue generally but our focus is really on smallest startups
40:41
SPEAKER_00: and entrepreneurs who are trying to build ai tools what are the equities here from
40:47
SPEAKER_00: a startup perspective
40:49
SPEAKER_01: let's start with access to data is so crucial for the development and use
40:53
SPEAKER_01: and refinement of ai systems right so that means that getting access to data
40:58
SPEAKER_01: and being able to use data that is an asset and a barrier to entry right
41:02
SPEAKER_01: and let's remember like the companies we're seeing today
41:05
SPEAKER_01: that are new open ai anthropic all these ones right access to public data
41:10
SPEAKER_01: to train their stuff that was startup capital
41:13
SPEAKER_01: that put them on a more level playing field with giants whether that's google or facebook
41:17
SPEAKER_01: and if you flip it around you think about a world where you have to get a license
41:22
SPEAKER_01: to do ai training and so on and so forth that means a huge new barrier to entry
41:28
SPEAKER_01: a16z
41:54
SPEAKER_01: a16z
42:06
SPEAKER_01: i thought every five hours right there like 500 million hours every you know added to
42:40
SPEAKER_01: so i think that is so huge for every sort of new entrance and startup i think
42:47
SPEAKER_01: to me this is one of the most challenging things in the in this debate
42:51
SPEAKER_01: where everything does have sort of okay what do we do about big tech and
42:57
SPEAKER_01: when people reach for copyright they are entrenching big tech
43:00
SPEAKER_01: they are reinforcing monopoly right limits and exceptions
43:04
SPEAKER_01: to copyright have always been a core anti-monopoly lever and when people miss that
43:10
SPEAKER_01: and you know start
43:11
SPEAKER_01: to say well we should just propertize everything they don't get how google
43:15
SPEAKER_01: and facebook are going to be the biggest winner out of that
43:18
SPEAKER_00: a16z
43:18
SPEAKER_00: and last but not least the law as often i think where people start as opposed to thinking about
43:24
SPEAKER_00: market evolution and technical collaboration but on the public policy side what's
43:28
SPEAKER_00: the roadmap that
43:29
SPEAKER_01: you think works best here this is said i think copyright law is consistent with
43:35
SPEAKER_01: the way ai is trained both copyrights function and purpose right copyright is there
43:39
SPEAKER_01: to incentivize
43:40
SPEAKER_01: creativity these are tools used to create new stuff and the training is on the un you know
43:46
SPEAKER_01: driving these uncopyrightable elements which is exactly the sort of in
44:05
SPEAKER_01: makes a lot of sense and to do text
44:07
SPEAKER_01: and data mining more generally
44:08
SPEAKER_01: or computational analysis more generally.
44:10
SPEAKER_01: And I think that's consistent
44:11
SPEAKER_01: with the flourishing market
44:13
SPEAKER_01: by the same token, as you said,
44:16
SPEAKER_01: right?
44:16
SPEAKER_01: Breaking into a site
44:17
SPEAKER_01: to get the data not okay, right?
44:20
SPEAKER_01: Outputting and selling,
44:22
SPEAKER_01: you know, full copies
44:23
SPEAKER_01: of somebody else's stuff
44:24
SPEAKER_01: and just saying, well, AI
44:25
SPEAKER_01: there's no AI get out card
44:27
SPEAKER_01: for that either.
44:29
SPEAKER_00: So meaning that if you're an AI developer
44:30
SPEAKER_00: you can engage in those practices
44:31
SPEAKER_00: as you would expect.
44:33
SPEAKER_01: yeah if your model is spitting out and selling you know copies of
44:39
SPEAKER_01: the latest stephen king novel or the latest marvel movie
44:42
SPEAKER_01: or whatever yeah that's not learning that's just copying it and forwarding along
44:48
SPEAKER_01: the copy so i think you know there are some lines to be drawn there and i think
44:53
SPEAKER_01: the basic line of you should be i a person should be able to use a machine
45:01
SPEAKER_01: i
45:08
SPEAKER_01: should be able to do that i think the the harder challenge here is going to come
45:12
SPEAKER_01: with you know and where a lot of the betas move these other elements of the law
45:16
SPEAKER_01: that we talked about contract anti-hacking law anti-circumvention law
45:21
SPEAKER_01: a16z
45:41
SPEAKER_01: i think consumer rights are a fundamental part of backstops about other types
45:47
SPEAKER_01: of bad market failures where consumers lose out where competitors lose out
45:51
SPEAKER_01: where innovation loses out and so there's a lot of room in between
45:55
SPEAKER_01: where there'll be litigation there'll be negotiation standard setting i'm not saying
45:59
SPEAKER_01: that you know we should force everybody to be interoperable and open no but
46:03
SPEAKER_01: but still i think we should live in a world where when i buy an ebook
46:07
SPEAKER_01: i should be able to use a read aloud function and they shouldn't be able
46:10
SPEAKER_01: to lock me out of that unfortunately that's not the world we live in today because
46:14
SPEAKER_01: of anti-circumvention law but i hope that you know this is actually a moment to try
46:19
SPEAKER_01: and reckon with some of that rather
46:20
SPEAKER_00: than expand those protections derek this is super fun thanks for coming on
46:24
SPEAKER_00: the podcast happy to