Latest / a16z AI Policy Brief / AI and the First Amendment

Transcript

AI and the First Amendment · a16z AI Policy Brief

• 0:00
  SPEAKER_03: We've talked a lot about Dormant Commerce Clause analysis.
• 0:02
  I'd love to get your take briefly in the time we have left on
• 0:06
  another potential
• 0:07
  SPEAKER_03: concern that might be raised by some of these state laws.
• 0:10
  A lot of them have disclosure-related provisions as a part of
• 0:14
  them in some form,
• 0:15
  SPEAKER_03: and so that raises possible First Amendment considerations.
• 0:19
  SPEAKER_03: Paul, starting with you, can you walk us through that concept?
• 0:23
  Why would the First Amendment apply to a situation when the
• 0:26
  government's mandating
• 0:27
  SPEAKER_03: certain disclosures?
• 0:29
  SPEAKER_01: Sure.
• 0:29
  So the First Amendment obviously protects free speech and laws
• 0:33
  that it prevents
• 0:35
  SPEAKER_01: laws that restrict speech,
• 0:36
  SPEAKER_01: but it also prevents laws that compel speech.
• 0:40
  So laws that compel people to engage in speech they would not
• 0:44
  otherwise want to
• 0:45
  SPEAKER_01: engage in always trigger some level of First Amendment scrutiny.
• 0:49
  How much scrutiny depends on what kind of speech is being
• 0:53
  compelled and in what context.
• 0:55
  SPEAKER_01: There's a lot of nuance to this,
• 0:56
  SPEAKER_01: but at a high level,
• 0:57
  the way courts think about it is if the compelled speech is
• 1:02
  purely factual and
• 1:03
  SPEAKER_01: non-controversial,
• 1:05
  then it gets a much lower level of scrutiny and it's generally
• 1:07
  going to survive
• 1:08
  SPEAKER_01: under the First Amendment.
• 1:09
  SPEAKER_01: So you think about all kinds of factual labeling requirements,
• 1:14
  SPEAKER_01: nutritional labels on food,
• 1:15
  SPEAKER_01: things like that.
• 1:16
  SPEAKER_01: But beyond that,
• 1:17
  if the speech that's being compelled by the government goes
• 1:20
  beyond the purely
• 1:21
  SPEAKER_01: factual and non-controversial,
• 1:23
  SPEAKER_01: it gets a higher level of scrutiny and
• 1:25
  and those kinds of compelled requirements often end up being
• 1:29
  struck down.
• 1:30
  SPEAKER_01: So there have been a number of cases recently,
• 1:33
  SPEAKER_01: almost all of them out of California,
• 1:35
  SPEAKER_01: at least the ones I'm thinking of,
• 1:37
  where the state has required companies in different industries
• 1:40
  to prepare different
• 1:41
  SPEAKER_01: kinds of reports.
• 1:42
  There was a law,
• 1:43
  for example,
• 1:43
  that required social media companies to prepare reports
• 1:47
  discussing risks to children from their services and their data
• 1:51
  practices.
• 1:52
  There was a law requiring drug companies to prepare reports
• 1:56
  that explain their
• 1:56
  SPEAKER_01: pricing decisions.
• 1:58
  And there was also a law that actually was partially enjoined
• 2:02
  just today by the Ninth Circuit
• 2:04
  that required companies to prepare reports about
• 2:06
  climate-related risks from their operations.
• 2:09
  SPEAKER_01: So in all of these contexts,
• 2:11
  SPEAKER_01: parties have challenged the laws and said,
• 2:14
  SPEAKER_01: you,
• 2:14
  SPEAKER_01: the state,
• 2:15
  SPEAKER_01: are requiring us,
• 2:16
  SPEAKER_01: by requiring these reports,
• 2:17
  SPEAKER_01: you're forcing us to engage in speech that we don't want to engage in.
• 2:20
  SPEAKER_01: And it's not just purely factual and non-controversial.
• 2:23
  SPEAKER_01: It's actually requiring us to
• 2:25
  SPEAKER_01: express opinions on some pretty controversial subjects.
• 2:29
  SPEAKER_01: Some of those laws have been struck down.
• 2:30
  SPEAKER_01: Some have survived in the Ninth Circuit.
• 2:32
  Some are still being challenged or likely to be challenged in
• 2:35
  the Supreme Court.
• 2:36
  SPEAKER_01: So this is very much a developing area of the law.
• 2:39
  And can we look a little at how the doctrine might map on to
• 2:43
  some of the things
• 2:43
  SPEAKER_03: that we're seeing in it?
• 2:46
  SPEAKER_03: So we're seeing, I think, a few different types of things.
• 2:49
  One would be that company conducts impact assessments or does
• 2:52
  those kinds of
• 2:53
  SPEAKER_03: assessments of its own potential security risks.
• 2:56
  SPEAKER_03: Another would be making disclosures to the government.
• 2:58
  So that could be things like certain kinds of content that is
• 3:02
  certain kind of
• 3:02
  prompts that you receive or certain kinds of content that you
• 3:05
  produce could also be
• 3:07
  requiring a company to say something to the government about
• 3:10
  safety practices the
• 3:12
  third bucket would be warning labels things that say things
• 3:15
  like this content is
• 3:16
  generated using ai or or things that say warning you've been
• 3:20
  using the service for
• 3:21
  x amount of time you need to now disclose that um a provider
• 3:25
  needs to disclose to a
• 3:26
  user that it's been using a service for a specific period of
• 3:29
  time so willie how
• 3:30
  would a court evaluate those various different kinds of
• 3:33
  transparency requirements
• 3:35
  SPEAKER_03: under the first amendment
• 3:36
  SPEAKER_02: Yeah,
• 3:37
  SPEAKER_02: I'll add actually just one more,
• 3:39
  SPEAKER_02: which would be actually saying that anything that children can access,
• 3:47
  SPEAKER_02: the AI has to generate,
• 3:49
  SPEAKER_02: whether it's images or text or whatever,
• 3:51
  SPEAKER_02: with the possibility that minors are reading it by in mind.
• 3:55
  SPEAKER_02: So on some of the examples that you gave,
• 3:58
  SPEAKER_02: like the simplest examples are more like a government warning label,
• 4:03
  like the requirement that you tell people you've been using the
• 4:07
  AI for this amount
• 4:09
  SPEAKER_02: of time.
• 4:10
  SPEAKER_02: And just like there are,
• 4:11
  SPEAKER_02: you know,
• 4:11
  SPEAKER_02: there's a surgeon general's warning on cigarettes,
• 4:14
  which have been challenged over the years as compelled speech
• 4:17
  and various like
• 4:19
  SPEAKER_02: graphic pictures of diseased lungs have been struck down.
• 4:22
  But factual statements about smoking being hazardous to your
• 4:26
  health is still on the
• 4:28
  SPEAKER_02: cigarette pack.
• 4:28
  So I think that the type of government warning or disclosure is
• 4:35
  likely to be the
• 4:36
  SPEAKER_02: weakest type of First Amendment challenge.
• 4:39
  Things that basically don't require the AI company or the model
• 4:44
  itself to mouth
• 4:46
  SPEAKER_02: things that they might disagree with.
• 4:49
  So it's clearly a message from the government and it's clearly
• 4:52
  something that is not kind of
• 4:55
  SPEAKER_02: hijacking the message or content generated by the AI itself.
• 5:00
  SPEAKER_03: And Alon,
• 5:02
  SPEAKER_03: maybe for you,
• 5:02
  what about the kinds of things the government might require
• 5:05
  around disclosing
• 5:06
  SPEAKER_03: information about safety practices?
• 5:08
  So those were some of the laws out of California that Paul was
• 5:12
  talking about
• 5:12
  SPEAKER_00: earlier,
• 5:13
  where the intention was to force social media companies and
• 5:17
  others in that area to
• 5:20
  SPEAKER_00: articulate what their own policies were on harm and
• 5:24
  restricting speech and the type of threats to children or
• 5:27
  others,
• 5:28
  certain types of hate speech.
• 5:30
  And one of the reasons that those laws failed First Amendment
• 5:33
  scrutiny is that just
• 5:35
  trying to come up with a clear definition of what counts as
• 5:40
  harmful speech or hate
• 5:41
  SPEAKER_00: speech is usually not a neutral endeavor.
• 5:45
  And so actually the very fact of forcing a company to sort of
• 5:51
  articulate its terms in
• 5:53
  the language that the government identified was itself a change
• 5:58
  of their message.
• 5:59
  And then you've got sort of even more extreme cases like the
• 6:02
  Supreme Court case out
• 6:03
  SPEAKER_00: of Florida and Texas,
• 6:05
  SPEAKER_00: where those companies were purporting,
• 6:07
  SPEAKER_00: you know,
• 6:08
  SPEAKER_00: passing laws that were trying to essentially
• 6:11
  alter what speech the social media companies would make
• 6:15
  available and push to their users.
• 6:18
  SPEAKER_00: So,
• 6:19
  SPEAKER_00: you know,
• 6:19
  if you used a particular type of algorithm that screened out
• 6:23
  speech that seemed
• 6:24
  SPEAKER_00: like it was hate speech,
• 6:25
  SPEAKER_00: that might run afoul of the Texas or Florida law.
• 6:28
  And what the Supreme Court held is that those companies had a
• 6:31
  First Amendment
• 6:32
  SPEAKER_00: right,
• 6:33
  SPEAKER_00: or at least,
• 6:33
  SPEAKER_00: you know,
• 6:34
  SPEAKER_00: in that posture,
• 6:34
  they could make an argument that they had a First Amendment
• 6:37
  right to sort of choose
• 6:39
  SPEAKER_00: for themselves
• 6:40
  SPEAKER_00: what sort of censorship to engage in,
• 6:42
  SPEAKER_00: what speech to allow users to put on the platform,
• 6:45
  SPEAKER_00: and what speech not to.
• 6:47
  SPEAKER_00: There was a really interesting concurrence by Justice Barrett,
• 6:50
  SPEAKER_00: who said,
• 6:52
  SPEAKER_00: you know,
• 6:52
  SPEAKER_00: in this instance,
• 6:53
  it seems fairly clear that the companies have made a very
• 6:56
  deliberate choice,
• 6:57
  SPEAKER_00: which the First Amendment protects,
• 6:59
  SPEAKER_00: not to allow certain forms of speech,
• 7:01
  SPEAKER_00: which they've decided is harmful.
• 7:03
  SPEAKER_00: But you could have an instance where
• 7:05
  in which artificial intelligence is used essentially to give
• 7:08
  the user whatever it
• 7:10
  SPEAKER_00: is that it wants.
• 7:11
  SPEAKER_00: And it's not totally clear,
• 7:12
  SPEAKER_00: she said,
• 7:13
  SPEAKER_00: in that instance,
• 7:14
  SPEAKER_00: that there would be any sort of decision,
• 7:17
  SPEAKER_00: speech-based decision,
• 7:18
  SPEAKER_00: content-based decision that the First Amendment would protect,
• 7:21
  because it would essentially be just a sort of automatic
• 7:24
  consequence of an
• 7:25
  SPEAKER_00: algorithm that was generated not for First Amendment reasons,
• 7:29
  SPEAKER_00: but just for essentially consumer preference reasons.
• 7:33
  SPEAKER_03: Paul, how do you think about that last point that Elan made?
• 7:36
  SPEAKER_03: What does editorial discretion look like in an AI context?
• 7:40
  SPEAKER_01: Yeah, I think it's really interesting.
• 7:42
  SPEAKER_01: Just as Barrett's opinion raises some really interesting questions,
• 7:45
  I think she's right to say that what the First Amendment
• 7:49
  ultimately protects is
• 7:50
  human expression and expressive choices where people are trying
• 7:54
  to make decisions
• 7:55
  SPEAKER_01: about the ideas they want to communicate.
• 7:58
  But where you locate that expression and whether you can have
• 8:00
  expression in the
• 8:01
  SPEAKER_01: form of an algorithm is a really interesting question.
• 8:04
  SPEAKER_01: You know,
• 8:05
  SPEAKER_01: Justice Barrett sort of suggests,
• 8:06
  and this is an argument we've seen plaintiffs make in
• 8:09
  litigation over social media,
• 8:11
  that if a platform is using an algorithm to try to serve
• 8:18
  content that the users
• 8:19
  SPEAKER_01: want,
• 8:20
  SPEAKER_01: that that is somehow not expressive.
• 8:23
  And I wonder if they would apply that same thinking to sort of
• 8:27
  more rudimentary
• 8:28
  SPEAKER_01: algorithms that companies have used for a long time.
• 8:31
  SPEAKER_01: Think about something like.
• 8:33
  SPEAKER_01: a radio station that plays the top 40, right?
• 8:37
  The producers at that radio station are not going out and
• 8:40
  choosing their favorite
• 8:42
  SPEAKER_01: music to play to listeners.
• 8:44
  SPEAKER_01: They're using top 40 is an algorithm.
• 8:46
  SPEAKER_01: It's not a very complicated algorithm,
• 8:48
  but it's an algorithm that looks at what music is popular and
• 8:51
  tries to serve that
• 8:52
  SPEAKER_01: music to listeners.
• 8:53
  SPEAKER_01: Now,
• 8:53
  SPEAKER_01: if the government came into a top 40 station and said,
• 8:57
  we're going to require you to play a certain amount of
• 8:59
  classical music every hour,
• 9:01
  and the reason we're allowed to do that is because we don't
• 9:04
  think you're engaged in
• 9:05
  SPEAKER_01: expression at all,
• 9:07
  SPEAKER_01: you're just using your top 40 algorithm,
• 9:09
  SPEAKER_01: I think courts would probably have a problem with that.
• 9:11
  I think they would very well might say that the decision to
• 9:16
  play top 40 music is
• 9:17
  SPEAKER_01: itself an expressive choice.
• 9:20
  SPEAKER_01: In terms of how that maps onto AI,
• 9:22
  I think courts,
• 9:24
  when they confront new technology,
• 9:26
  like to reason by analogy.
• 9:28
  SPEAKER_01: They like to try to figure out what things that I'm familiar with,
• 9:31
  SPEAKER_01: that courts have dealt with before,
• 9:34
  SPEAKER_01: does this look and feel like?
• 9:36
  SPEAKER_01: What is it similar to in relevant ways?
• 9:39
  SPEAKER_01: I think developers
• 9:41
  who want to claim protection for certain aspects of model
• 9:46
  development are going to
• 9:48
  SPEAKER_01: want to draw these analogies where they say,
• 9:50
  SPEAKER_01: this is a conduit for my speech.
• 9:53
  Decisions that I am making in model development are expressive
• 9:57
  choices where I am trying to
• 9:59
  SPEAKER_01: to influence or control the speech that is output by this model.
• 10:03
  SPEAKER_01: And on the other hand,
• 10:04
  SPEAKER_01: the argument on the other side is going to be,
• 10:06
  SPEAKER_01: there's no real expressive choice happening.
• 10:08
  SPEAKER_01: This is all basically the construction of a machine.
• 10:11
  And I think that's going to be the challenge for developers is
• 10:14
  to articulate how
• 10:16
  there are expressive choices being made as part of the
• 10:18
  development process.
• 10:19
  SPEAKER_00: One of the sort of interesting features of
• 10:22
  certain large language models is that when you ask the people
• 10:25
  who develop these
• 10:26
  models how it is that the model came to think that writing a
• 10:30
  particular response in
• 10:32
  SPEAKER_00: a particular way made sense,
• 10:34
  SPEAKER_00: they will often say,
• 10:34
  SPEAKER_00: look,
• 10:35
  SPEAKER_00: it's a black box.
• 10:36
  We don't actually know what it is that the model used in order
• 10:40
  to decide that this
• 10:41
  SPEAKER_00: word comes most appropriately after that word.
• 10:44
  SPEAKER_00: And so in some sense,
• 10:46
  SPEAKER_00: you might say,
• 10:46
  SPEAKER_00: well,
• 10:47
  you're not really making a choice that the First Amendment
• 10:49
  should respect.
• 10:50
  SPEAKER_00: On the other hand,
• 10:51
  they also do usually impose constraints like they won't give
• 10:55
  you an answer that
• 10:56
  SPEAKER_00: violates copyright law.
• 10:57
  They're not going to tell you the lyrics of a song or they're
• 11:01
  not going to tell you
• 11:01
  SPEAKER_00: the first or last page of a book just because you ask it to.
• 11:05
  SPEAKER_00: And that's a deliberate choice that's made in the product.
• 11:09
  SPEAKER_03: This is a helpful set of tools that we can use,
• 11:11
  SPEAKER_03: I think,
• 11:11
  in evaluating all the legislation we're likely to see ahead at
• 11:14
  the federal and
• 11:14
  SPEAKER_03: state level.
• 11:15
  SPEAKER_03: Alon, Paul, Willie, thank you so much.
• 11:19
  SPEAKER_01: It's been a pleasure.
• 11:19
  SPEAKER_01: Thank you.