Latest / a16z AI Policy Brief / A Guide to Crafting a Liability Regime for AI

Transcript

A Guide to Crafting a Liability Regime for AI · a16z AI Policy Brief

• 0:00
  SPEAKER_01: Welcome back to the AI Policy Brief.
• 0:01
  SPEAKER_01: Today,
• 0:02
  SPEAKER_01: Jerry Ramaswamy,
• 0:03
  SPEAKER_01: Chief Legal and Policy Officer at A16z,
• 0:06
  SPEAKER_01: returns to discuss one of the most important questions in AI policy.
• 0:09
  SPEAKER_01: When AI is used to cause harm, who should be held responsible?
• 0:13
  SPEAKER_01: Answering that question is the job of a liability regime.
• 0:16
  SPEAKER_01: A well-crafted regime can hold bad actors accountable,
• 0:19
  SPEAKER_01: build trust among users and the public,
• 0:21
  SPEAKER_01: and create the conditions for a healthy,
• 0:23
  SPEAKER_01: durable AI ecosystem.
• 0:25
  but a poorly crafted one by contrast can leave people
• 0:28
  unprotected punish people
• 0:29
  building the technology in good faith and entrench incumbents
• 0:33
  Jay and I discuss how
• 0:34
  we might find a Goldilocks approach designing a liability
• 0:37
  framework that provides
• 0:37
  meaningful accountability without chilling innovation or making
• 0:40
  it harder for
• 0:41
  startups to compete we cover a wide range of current liability
• 0:45
  proposals like
• 0:46
  strict liability for developers duties of care and using AI as
• 0:49
  an aggregating
• 0:50
  factor in crimes and then we talk about how each one would
• 0:53
  impact little tech
• 0:55
  SPEAKER_01: This is a complex and nuanced issue, but getting it right matters.
• 1:10
  SPEAKER_01: The goal is a regime where people trust AI enough to use it,
• 1:13
  SPEAKER_01: while keeping the market open enough for little tech to compete.
• 1:16
  SPEAKER_01: Here's the conversation.
• 1:18
  SPEAKER_00: You're looking for a bit of a Goldilocks principle here, right?
• 1:21
  SPEAKER_00: What is the appropriate...
• 1:23
  SPEAKER_00: AI Policy Brief a16z
• 1:37
  SPEAKER_00: Those types of ecosystems can only be built on trust.
• 1:42
  SPEAKER_00: Trust is key.
• 1:43
  A system that you can have trust in that balances these
• 1:46
  equities well is the type
• 1:48
  SPEAKER_00: of system of liability that will produce a long-term healthy industry.
• 1:54
  SPEAKER_00: And conversely,
• 1:56
  a poorly crafted regime can absolutely destroy innovation and
• 2:00
  incentives to
• 2:01
  SPEAKER_00: innovate.
• 2:02
  SPEAKER_00: Those two things play off each other.
• 2:03
  SPEAKER_00: And that's why all of this is super important.
• 2:12
  SPEAKER_00: And we're back with the AI Policy Brief.
• 2:14
  SPEAKER_00: Jay, good morning.
• 2:15
  SPEAKER_00: Welcome.
• 2:16
  SPEAKER_00: Good morning.
• 2:17
  SPEAKER_00: Pleasure to be here.
• 2:18
  SPEAKER_00: And I can't wait for this discussion.
• 2:20
  This is a tough one that you and I have been having privately
• 2:23
  now for a long period
• 2:24
  SPEAKER_01: of time,
• 2:25
  SPEAKER_01: talking about AI liability.
• 2:28
  It's something that I think has been in your mind when you
• 2:30
  thought about AI policy
• 2:32
  SPEAKER_01: for a while now,
• 2:34
  SPEAKER_01: and it's just incredibly fraught and incredibly challenging.
• 2:36
  SPEAKER_01: How does this concept sit in your head?
• 2:38
  SPEAKER_01: Like, why has it been front and center for you for so long?
• 2:41
  SPEAKER_00: Well,
• 2:42
  SPEAKER_00: I think that I was anticipating that whether we like it or not,
• 2:47
  SPEAKER_00: whether others like it or not is irrelevant,
• 2:50
  SPEAKER_00: that the liability question,
• 2:52
  SPEAKER_00: because of the way our legal system works,
• 2:55
  SPEAKER_00: is going to work its way through.
• 2:56
  We're seeing a number of cases already kind of working their
• 2:59
  way through the system
• 3:00
  SPEAKER_00: that address this.
• 3:02
  SPEAKER_00: And so...
• 3:03
  I just think it's one that we're going to have to grapple with
• 3:07
  and you know as
• 3:08
  we've discussed on numerous occasions there isn't a clear and
• 3:13
  easy answer to this
• 3:14
  one I think of all the issues that we've addressed I'm going to
• 3:18
  be riffing off of
• 3:19
  something that that you said in our conversations but you're
• 3:22
  looking for a bit of a
• 3:23
  SPEAKER_00: Goldilocks principle here right what is the appropriate
• 3:26
  SPEAKER_00: So Jay, why is all this important?
• 3:28
  Why does it matter to set up a liability regime that balances
• 3:30
  all these equities well?
• 3:54
  SPEAKER_00: Look,
• 3:54
  SPEAKER_00: at the end of the day,
• 3:55
  I think what people sometimes miss about firms like ours and
• 3:59
  the ecosystem we care
• 4:00
  SPEAKER_00: about is that we're buy and hold investors,
• 4:03
  SPEAKER_00: right?
• 4:04
  SPEAKER_00: And so from our perspective, the long term matters a lot.
• 4:07
  SPEAKER_00: It doesn't matter if you can make a quick buck.
• 4:10
  SPEAKER_00: in one or two years by some sort of fly-by-night thing that you do.
• 4:15
  SPEAKER_00: Because for us,
• 4:16
  we're interested in building technologies and ultimately
• 4:19
  industries and businesses.
• 4:20
  And that takes typically 10,
• 4:22
  12,
• 4:22
  sometimes increasingly 20 years to build.
• 4:26
  SPEAKER_00: And those types of ecosystems can only be built on trust.
• 4:31
  SPEAKER_00: Trust is key.
• 4:32
  And so a well-crafted liability regime gives consumers some
• 4:36
  trust in the kinds of
• 4:38
  SPEAKER_00: products that are being produced and what they're being exposed to.
• 4:42
  SPEAKER_00: And so
• 4:43
  A system that you can have trust in that balances these
• 4:46
  equities well is the type
• 4:48
  of system of liability that will produce a long term healthy
• 4:54
  industry where we can
• 4:56
  SPEAKER_00: fund companies that can then produce groundbreaking technologies.
• 5:00
  that's why all of this is super important and so those two
• 5:03
  things play off each
• 5:04
  other and conversely a poorly crafted regime can absolutely
• 5:10
  destroy innovation and
• 5:11
  incentives to innovate and so that's why this Goldilocks
• 5:14
  problem in the context of
• 5:16
  SPEAKER_00: liability regimes becomes really important and why
• 5:19
  Sometimes we're often seen as being on the one side of an issue
• 5:22
  and then the other
• 5:23
  SPEAKER_00: side people can get confused,
• 5:24
  but that's the reason we're trying to calibrate this in a way
• 5:27
  that establishes
• 5:28
  trust without destroying innovation so that we can build an
• 5:31
  industry over medium to
• 5:34
  SPEAKER_00: long term.
• 5:35
  so the right answer I think is really really hard finding the
• 5:38
  Goldilocks we have a
• 5:40
  sense of the wrong answers at the polls so maybe we should just
• 5:42
  like go through
• 5:43
  those so on one side you would have an ecosystem where there's
• 5:48
  just crushing my
• 5:49
  SPEAKER_01: ability
• 5:50
  SPEAKER_01: AI Policy Brief a16z
• 6:08
  Navigate Through or there are massive penalties so that you
• 6:11
  have to be like going
• 6:12
  back to a war chest to try to pull out money to pay for court
• 6:16
  settlements that's
• 6:17
  going to significantly advantage larger companies and I think
• 6:21
  that's a scenario
• 6:23
  that we have concerns about we also have some concerns about
• 6:25
  the other end of the
• 6:26
  SPEAKER_01: spectrum do you want to talk through that side
• 6:29
  I think at the end of the day,
• 6:31
  you need a regime that accomplishes a few things,
• 6:35
  right?
• 6:35
  SPEAKER_00: You need accountability placed on the right actors in the system.
• 6:41
  SPEAKER_00: You need a level of
• 6:43
  SPEAKER_00: A16z
• 7:12
  SPEAKER_01: A16z
• 7:30
  Even in a Section 230 world But it's perceived as a
• 7:32
  get-out-of-jail-free card It's
• 7:34
  perceived to be a zero-liability regime A full shield of
• 7:37
  liability Conferring
• 7:39
  SPEAKER_01: blanket immunity for tech platforms And I think we repeatedly hear from
• 7:44
  policymakers When they look at AI They conceive of 230 as a
• 7:47
  mistake And as
• 7:48
  something that they're trying to avoid And so I think when you
• 7:51
  talk about the
• 7:52
  SPEAKER_01: importance of
• 7:54
  Wrongful Actors Being Subject to Liability That is an important
• 7:59
  component for us
• 7:59
  because we're pushing back against the notion that prior
• 8:02
  technology waves have
• 8:05
  SPEAKER_01: provided too much of a liability shield.
• 8:07
  SPEAKER_00: Yeah, I think that's right.
• 8:08
  I think Section 230 is instructive because I think you have hit
• 8:11
  the nail on the head.
• 8:13
  There's a popular perception of it and then there's the kind of
• 8:15
  legal application.
• 8:17
  SPEAKER_00: On the one hand, there is a broad sort of protection against liability
• 8:22
  SPEAKER_00: A16z
• 8:37
  SPEAKER_00: A16z
• 8:56
  SPEAKER_00: David Gunn elizgunn a16z
• 9:11
  SPEAKER_00: AI Policy Brief a16z
• 9:26
  SPEAKER_00: AI Policy Brief
• 9:54
  SPEAKER_01: You've been working on these concepts for a while.
• 9:57
  One liability proposal that you had significant concerns about
• 10:02
  from a couple of
• 10:02
  SPEAKER_01: years ago was SB 1047 in California,
• 10:05
  SPEAKER_01: which would have imposed liability on developers for downstream misuse.
• 10:09
  SPEAKER_01: Why was that liability concept one that was so concerning?
• 10:13
  SPEAKER_00: Look,
• 10:13
  SPEAKER_00: I think for the reasons that I just mentioned,
• 10:16
  SPEAKER_00: which is if you are truly producing a product for general applicability,
• 10:21
  SPEAKER_00: The downstream uses and impacts,
• 10:24
  SPEAKER_00: some may be foreseeable,
• 10:26
  SPEAKER_00: some may be foreseeable if you squint and take a really careful look,
• 10:29
  SPEAKER_00: and some may not be foreseeable.
• 10:31
  And so what kind of regime are you actually placing on those
• 10:34
  developers and
• 10:36
  SPEAKER_00: A16z
• 10:51
  SPEAKER_00: A16z
• 11:14
  SPEAKER_00: A16z
• 11:38
  I'm just trying to take through what are all the different ways
• 11:40
  that this is
• 11:41
  SPEAKER_01: surfacing in current context that we have some concerns about.
• 11:46
  Strict liability proposals,
• 11:48
  imposing developer liability for downstream misuse.
• 11:51
  I think the concern there is just are you actually putting
• 11:54
  liability on the
• 11:54
  SPEAKER_01: responsible party if a developer is just putting a general purpose tool,
• 11:58
  SPEAKER_01: someone else uses it in a harmful way.
• 12:00
  SPEAKER_01: A16z
• 12:15
  SPEAKER_01: One that we've seen recently that is really interesting,
• 12:18
  SPEAKER_01: I think,
• 12:19
  SPEAKER_01: I don't really know where it comes from,
• 12:21
  but on the criminal side is the idea that AI should essentially
• 12:25
  be an aggravating
• 12:26
  SPEAKER_01: circumstance in a crime.
• 12:28
  So if you use AI to commit a crime,
• 12:30
  you'd be subject to heightened penalties.
• 12:32
  SPEAKER_01: We just saw this in Tennessee,
• 12:34
  SPEAKER_01: where it would have been a Class A felony to do certain things with AI,
• 12:38
  SPEAKER_01: some that would be really problematic,
• 12:40
  SPEAKER_01: like,
• 12:40
  SPEAKER_01: you know,
• 12:40
  SPEAKER_01: assistant commission of a crime or a homicide.
• 12:43
  but one of them was misrepresent your whether you had a license
• 12:49
  to practice medical
• 12:51
  SPEAKER_01: services to practice medicine and making that a class A felony would mean
• 12:56
  misrepresenting whether you had a license or not would be at
• 12:58
  the same criminal
• 12:59
  level in Tennessee as first-degree murder which is a bizarre
• 13:04
  way to think about AI
• 13:05
  that the penalties would be so significantly heightened right
• 13:09
  and one way you know
• 13:11
  we often talk about proportionality in liability regimes the
• 13:14
  question I would have
• 13:15
  is would somebody representing themselves as a licensed
• 13:19
  physician who is not in
• 13:21
  fact a licensed physician be subject to those penalties as
• 13:25
  somebody using AI to
• 13:27
  SPEAKER_00: make that same representation and it seems that there should be some
• 13:31
  proportionality because those two sets of conduct are equally
• 13:35
  egregious it seems to
• 13:36
  SPEAKER_00: me and if they're not being punished similarly then it's a kind of odd
• 13:41
  Duck Statute I would think yeah I think the concept of
• 13:45
  proportionality is really
• 13:46
  important to how we think about leaving room for innovation and
• 13:50
  the right way so
• 13:51
  SPEAKER_01: you don't want to allow innovators to
• 13:54
  SPEAKER_01: AI Policy Brief
• 14:13
  Misrepresentation Related to Licensed Services I think that
• 14:16
  results in developers
• 14:17
  really building away from offering tools anywhere close to that
• 14:21
  which some people
• 14:22
  might say is great like we don't want an AI tool to suggest
• 14:26
  that it has a medical
• 14:28
  license or can practice mental health or can operate as a
• 14:32
  lawyer if it doesn't have
• 14:33
  the requisite credentials but if the liability is so crushing
• 14:38
  that any mention of
• 14:39
  SPEAKER_01: any of those topics all of a sudden
• 14:42
  SPEAKER_01: AI Policy Brief a16z
• 15:10
  SPEAKER_00: A16z
• 15:25
  SPEAKER_00: A16z
• 15:49
  the kinds of trade-offs that one would want to see in a
• 15:52
  liability regime.
• 15:53
  SPEAKER_00: Because I think that's important as well.
• 15:55
  If we want to maintain a competitive world in AI,
• 15:58
  you need open source and you need little tech.
• 16:00
  SPEAKER_00: And those two are obviously related.
• 16:03
  And so a liability regime can very much change the market
• 16:06
  structure and the market
• 16:07
  SPEAKER_00: composition depending on how it's imposed.
• 16:10
  But talk a little bit about that perspective because I think
• 16:12
  that's sometimes forgotten.
• 16:14
  SPEAKER_01: Yeah, I think the key thing there is
• 16:17
  SPEAKER_01: AI Policy Brief a16z
• 16:35
  SPEAKER_01: You are not, when you are developing that tool, engaged in wrongdoing.
• 16:39
  And therefore,
• 16:39
  for the most part,
• 16:40
  there shouldn't be significant liability attached to that.
• 16:43
  SPEAKER_01: And it shouldn't be hard for you to do that,
• 16:45
  SPEAKER_01: for you to just do the basic work of building software,
• 16:49
  SPEAKER_01: of building AI.
• 16:50
  AI Products that can go out in the world and have a lot of
• 16:52
  positive value.
• 16:54
  The proposals that target that phase of development and make it
• 16:57
  harder for
• 16:58
  developers to build are the kind of paperwork favors the
• 17:01
  powerful proposals that we
• 17:03
  SPEAKER_01: find to be so problematic.
• 17:05
  SPEAKER_01: If they're very paperwork heavy,
• 17:07
  SPEAKER_01: I love that phrase, but can you unpack that a little bit?
• 17:10
  SPEAKER_01: I wish it was our phrase.
• 17:13
  SPEAKER_01: I love the phrase.
• 17:15
  SPEAKER_01: I wish it was our phrase.
• 17:16
  SPEAKER_01: It's Jen Palka's phrase.
• 17:17
  SPEAKER_01: She was the deputy chief technology officer in the Obama administration.
• 17:21
  SPEAKER_01: She wrote a great book.
• 17:22
  SPEAKER_01: A16z
• 17:44
  just trying to build an AI tool is not wrongdoing there's
• 17:46
  nothing wrong with it and
• 17:47
  I think that is like that feels like we're saying something
• 17:50
  obvious but I do think
• 17:51
  there are parts of the policy debate that are tilted away from
• 17:53
  that concept if
• 17:54
  SPEAKER_01: you're thinking about aggravated penalties just because you use AI um
• 18:00
  That suggests that the building of an AI tool has something
• 18:04
  unsavory about it or
• 18:06
  SPEAKER_01: something second class or something nefarious or problematic about it.
• 18:11
  The fact that you used AI to commit harm is different than
• 18:15
  using any other tool to
• 18:16
  SPEAKER_01: commit the exact same amount of harm.
• 18:19
  SPEAKER_00: And in our view, you use internet for false advertising, right?
• 18:22
  SPEAKER_00: AI is meaningfully worse than that.
• 18:24
  SPEAKER_00: It's unclear that it is.
• 18:26
  SPEAKER_01: Yeah,
• 18:26
  SPEAKER_01: if you use,
• 18:27
  SPEAKER_01: you know,
• 18:27
  SPEAKER_01: video editing software to alter a video,
• 18:30
  SPEAKER_01: but the video editing software is not AI,
• 18:32
  SPEAKER_01: that's,
• 18:33
  SPEAKER_01: you create a deepfake,
• 18:34
  SPEAKER_01: that's different from using AI to create a deepfake,
• 18:36
  SPEAKER_01: even if the deepfake,
• 18:38
  SPEAKER_01: using a deepfake in a fraudulent way,
• 18:40
  SPEAKER_01: for instance,
• 18:40
  SPEAKER_01: that should,
• 18:41
  SPEAKER_01: I think,
• 18:41
  SPEAKER_01: probably be equally harmful,
• 18:43
  SPEAKER_01: regardless of the tool that's used.
• 18:44
  SPEAKER_01: And so our focus, I think, is focus on the harm itself.
• 18:48
  SPEAKER_01: And that really, I think,
• 18:49
  SPEAKER_01: AI Tool with the intent
• 18:59
  I want to pivot a little because you know one of the principles
• 19:02
  that's kind of
• 19:28
  SPEAKER_00: I think it tracks back to the Coase Theorem and Ronald Coase,
• 19:37
  a University of Chicago economist who became very influential
• 19:40
  in legal circles.
• 19:41
  SPEAKER_00: But the idea is that you should put liability on the least
• 19:48
  SPEAKER_00: AI Policy Brief a16z
• 20:08
  Compete effectively in these areas or can it I guess would be
• 20:11
  the question it's a
• 20:12
  difficult area but it is embedded within sort of liability
• 20:16
  generally so I think
• 20:17
  SPEAKER_00: it's probably worth examining a little bit
• 20:19
  SPEAKER_01: Yeah,
• 20:19
  SPEAKER_01: well,
• 20:19
  I'd love giving your expertise in economics to hear a little
• 20:22
  bit more about your
• 20:23
  SPEAKER_01: views on it.
• 20:24
  But I think there are cases where clearly that would result in
• 20:29
  focusing on the developer.
• 20:30
  So in the case that I described of a developer who is really
• 20:35
  building a tool that's
• 20:35
  SPEAKER_01: solely going to be used for criminal activity,
• 20:37
  SPEAKER_01: that's the intent of the product.
• 20:39
  SPEAKER_01: AI Policy Brief a16z
• 20:54
  SPEAKER_00: AI Policy Brief a16z
• 21:16
  SPEAKER_00: But what has to be proven is that you're complicit in some way,
• 21:20
  SPEAKER_00: that you're not just a developer producing something,
• 21:22
  but you're actually working with and embedded with sort of
• 21:26
  criminal conduct and
• 21:27
  SPEAKER_00: activity.
• 21:27
  SPEAKER_01: Right.
• 21:28
  SPEAKER_01: And there's no defense in existing law.
• 21:31
  SPEAKER_01: You can't say, I used AI, and so therefore I'm entitled to a defense.
• 21:35
  SPEAKER_01: There's no AI defense in existing law.
• 21:37
  SPEAKER_01: So if existing law applies and you used AI, you'll be held liable.
• 21:41
  I think there are other cases where the least cost avoider
• 21:43
  principle is really focused on users.
• 21:46
  SPEAKER_01: AI Policy Briefs
• 22:13
  SPEAKER_01: how do you see the least cost avoider principle playing out here
• 22:36
  SPEAKER_00: Yeah, very similarly.
• 22:39
  SPEAKER_00: I think I'm going to geek out a little bit here,
• 22:41
  SPEAKER_00: but one of the paradigmatic examples of these kinds of principles,
• 22:46
  do you remember in law school there was this concept of res
• 22:49
  ipsa loquitur,
• 22:50
  SPEAKER_00: that the thing speaks for itself?
• 22:52
  And so the idea was that if you had a barrel kind of that fell
• 22:56
  out of a window and
• 22:57
  SPEAKER_00: harmed someone,
• 22:59
  SPEAKER_00: you know,
• 22:59
  SPEAKER_00: in a manufacturing district,
• 23:01
  SPEAKER_00: that in a sense the thing spoke for itself,
• 23:03
  SPEAKER_00: that the responsibility,
• 23:04
  SPEAKER_00: and I'm mashing this up from my 1L understanding,
• 23:09
  In some senses the lowest cost avoider is the person who has
• 23:12
  control of that barrel
• 23:13
  and they should be putting appropriate guardrails on it so it
• 23:16
  doesn't fall out the
• 23:17
  window and harm someone rather than putting it on the person on
• 23:20
  the street who
• 23:21
  SPEAKER_00: really didn't have a way of protecting themselves from it.
• 23:26
  That is almost reversed in the case of users who are doing
• 23:31
  harmful things with the
• 23:32
  SPEAKER_00: technology versus the general purpose developer.
• 23:35
  SPEAKER_00: It seems to me that in that situation,
• 23:37
  it really is the user who knows what they're going to deploy
• 23:42
  this thing for,
• 23:43
  SPEAKER_00: whether that's an individual,
• 23:44
  SPEAKER_00: a company or whatever else.
• 23:47
  and the lowest cost avoider is probably that individual as
• 23:51
  opposed to the developer
• 23:52
  of the general purpose tool who has to go through a ton of
• 23:55
  ruminations to figure
• 23:56
  SPEAKER_00: out what is every conceivable thing that somebody could do using this.
• 24:02
  SPEAKER_00: And that's just a very difficult thing to foresee in circumstances.
• 24:06
  SPEAKER_00: So again, I'm not sure I got all of the economics and the law there.
• 24:11
  SPEAKER_00: I think it's been a few years since I took
• 24:14
  Tort Law but that's always been in my head one of the
• 24:16
  constructs one of the mental
• 24:17
  models that one can think about how one applies this and there
• 24:21
  are gradations in
• 24:21
  the middle there where I think those are clear sort of examples
• 24:25
  but there are
• 24:26
  probably gradations in the middle where you have to make some
• 24:28
  judgment calls and
• 24:30
  SPEAKER_00: that's where it becomes difficult
• 24:31
  SPEAKER_01: It seems like the analogy that we're concerned about is there's a barrel
• 24:35
  SPEAKER_01: manufacturer who makes a barrel pretty well.
• 24:38
  Someone buys the barrel,
• 24:39
  takes it home,
• 24:40
  pushes it out their window and someone gets hurt.
• 24:42
  SPEAKER_01: Should the barrel manufacturer have liability in that situation?
• 24:45
  And no,
• 24:46
  we would put it on the person who's decided to misuse the
• 24:48
  technology.
• 24:49
  SPEAKER_01: The barrel could be used for a lot of different purposes.
• 24:51
  SPEAKER_01: It's not supposed to be used to be pushed out a window to hurt someone.
• 24:54
  And so therefore putting liability on the person who did it is
• 24:57
  that that's the
• 24:58
  SPEAKER_01: right way to do it.
• 24:59
  SPEAKER_00: Yeah, I like that.
• 25:00
  SPEAKER_00: I like that analogy.
• 25:01
  It kind of brings full circle my archaic example that's in
• 25:04
  every single I think 1L
• 25:06
  SPEAKER_00: law textbook out there.
• 25:07
  SPEAKER_00: You have a better memory of 1L than I do.
• 25:09
  SPEAKER_01: I don't even remember the barrel.
• 25:11
  So maybe that's the right segue then to talking a little bit
• 25:13
  about the kinds of
• 25:14
  SPEAKER_01: things that we might actually want to see as part of a liability regime.
• 25:18
  SPEAKER_01: Yeah, I think that'd be great.
• 25:20
  One concept that we've seen in the ecosystem that we drew on
• 25:23
  for this piece that I
• 25:23
  think was really helpful was initially published by Alliance
• 25:27
  for the Future and
• 25:28
  that was the idea that there should be a presumption of user
• 25:31
  liability in cases
• 25:33
  SPEAKER_01: related to AI outputs and so that concept means that
• 25:38
  SPEAKER_01: A16z
• 26:03
  and principally to generate unlawful content to engage in
• 26:06
  unlawful conduct.
• 26:08
  SPEAKER_01: And in those kinds of cases,
• 26:09
  then obviously that presumption could be overcome and the
• 26:13
  developer could be held
• 26:14
  SPEAKER_01: liable for outputs that are unlawful.
• 26:16
  One I think that is important that actually gets at a feature
• 26:19
  of Section 230 that's
• 26:20
  often overlooked is figuring out how to use procedural tools to
• 26:24
  reduce the cost of
• 26:25
  SPEAKER_01: frivolous litigation.
• 26:26
  I think this is something that's really vital to little tech
• 26:29
  and just really often
• 26:31
  SPEAKER_01: gets overlooked.
• 26:32
  SPEAKER_01: It is not enough for small companies to eventually win lawsuits.
• 26:36
  The threat of a lawsuit and the likelihood that they'd have to
• 26:39
  pursue a lawsuit
• 26:41
  Long through litigation is something that could really shift
• 26:46
  the AI markets toward
• 26:47
  more concentration where larger companies really have dominant
• 26:50
  positions because
• 26:52
  they're the ones who are likely to have illegal teams that can
• 26:54
  litigate a case
• 26:55
  SPEAKER_01: through the end.
• 26:56
  SPEAKER_01: And I think from our perspective,
• 26:58
  burying those significant litigation costs all through the
• 27:01
  process is the kind of
• 27:02
  SPEAKER_01: thing that's going to be difficult for little tech.
• 27:04
  SPEAKER_00: Yeah.
• 27:05
  SPEAKER_00: Look, I think that's a really important principle.
• 27:07
  SPEAKER_00: And we see analogs of this, at least in federal law.
• 27:11
  SPEAKER_00: You know,
• 27:11
  SPEAKER_00: in my days when I practiced very briefly corporate law,
• 27:14
  there was a debate and eventually law passed the Private
• 27:18
  Securities Litigation
• 27:18
  SPEAKER_00: Reform Act.
• 27:19
  SPEAKER_00: And it did.
• 27:20
  SPEAKER_00: A16z
• 27:44
  You had to actually have substantial evidence and allege that
• 27:48
  substantial evidence
• 27:49
  SPEAKER_00: and a court could dismiss it at the so-called motion to dismiss stage,
• 27:53
  SPEAKER_00: so fairly early in the litigation,
• 27:55
  SPEAKER_00: which would reduce cost significantly.
• 27:57
  SPEAKER_00: Not only that...
• 27:58
  SPEAKER_00: A16z
• 28:13
  There would be a stay of discovery until the motion to dismiss
• 28:17
  had been decided and
• 28:18
  that allowed judges to sort of come in amongst other reforms
• 28:22
  that were introduced
• 28:22
  there allowed judges to sort of police this and at least reduce
• 28:27
  the cost of
• 28:28
  SPEAKER_00: litigation.
• 28:28
  SPEAKER_00: A16z
• 28:43
  SPEAKER_00: A16z
• 29:07
  put aside the legal fees that you've got to pay to the outside
• 29:10
  counsel just the
• 29:11
  time of involvement of that founding team can be enormous and
• 29:17
  if they've done
• 29:17
  something wrong then obviously liability needs to attach but it
• 29:21
  shouldn't be a
• 29:21
  penalty just to be small and to have kind of that level of it
• 29:24
  so I think those
• 29:25
  types of reforms from a frivolous litigation point of view are
• 29:28
  tremendously
• 29:29
  SPEAKER_00: important.
• 29:30
  So one thing that we've been concerned about that's relevant
• 29:32
  here is really
• 29:33
  SPEAKER_01: expansive theories of duty of care.
• 29:36
  SPEAKER_01: So a duty of care is part of a negligence lawsuit.
• 29:38
  SPEAKER_01: So that will happen in AI just as it happens in many other fields.
• 29:41
  But there are also congressional proposals that would create
• 29:44
  statutory proposals.
• 29:45
  SPEAKER_01: A16z
• 30:05
  the meritorious lawsuits that should proceed and those that are
• 30:08
  a little bit more
• 30:08
  frivolous and one concept that we've been batting around is the
• 30:11
  idea of having a
• 30:12
  safe harbor having a duty of care having a statutory federal
• 30:16
  duty of care but
• 30:17
  having some safe harbor that little tech companies would be
• 30:21
  able to take advantage
• 30:22
  of companies actually of all sizes but hopefully would be
• 30:24
  designed in a way that
• 30:25
  SPEAKER_01: just they wouldn't
• 30:26
  Claims that are brought for real violations of a clear duty
• 30:29
  would be able to have
• 30:30
  SPEAKER_01: recourse in court,
• 30:31
  SPEAKER_01: and we think that that's a positive thing.
• 30:46
  but having some sort of well-designed safe harbor I think would
• 30:49
  also help to get
• 30:50
  cases out earlier in litigation so might enable a defendant to
• 30:54
  get out at the
• 30:55
  motion dismiss phase which as you were describing before saves
• 30:58
  you the time and
• 30:59
  effort and energy and financial resources of the discovery
• 31:03
  process so that might be
• 31:04
  SPEAKER_01: a way to balance the equities here to give recourse to plaintiffs in some
• 31:07
  SPEAKER_01: circumstances but also protect little tech from a mountain of litigation.
• 31:12
  SPEAKER_00: Yeah, look, and I think safe harbors are an effective tool.
• 31:16
  SPEAKER_00: And if you pull back from the litigation point of view,
• 31:19
  SPEAKER_00: they're also an opportunity to shape behavior,
• 31:21
  SPEAKER_00: right?
• 31:22
  Because depending on the criteria that you establish for that
• 31:26
  safe harbor,
• 31:27
  SPEAKER_00: now that's what builders build to.
• 31:29
  And I'm a fan of bright line rules because I think bright line
• 31:32
  rules are effective
• 31:33
  in shaping behavior that then reduces the overall liability in
• 31:38
  the system and makes
• 31:39
  SPEAKER_00: it clear that if you don't abide by that,
• 31:42
  The devil then of course falls in the details which is what are
• 31:45
  those criterias?
• 31:57
  SPEAKER_00: that you want to use to provide that safe harbor.
• 31:59
  And that's oftentimes the sort of arguments around legislative
• 32:02
  sausage making,
• 32:03
  SPEAKER_00: I think,
• 32:03
  is what behaviors do you want to shape and what are those
• 32:07
  criteria that you want to
• 32:08
  SPEAKER_00: have implemented in order to get the benefit of safe harbor.
• 32:12
  SPEAKER_00: And I think they can be tremendously effective tools as well.
• 32:15
  That raises another good point about one of our principles,
• 32:18
  which is protecting good behavior.
• 32:19
  We think it's really important that a liability regime not
• 32:23
  punish companies when
• 32:24
  SPEAKER_01: they're trying to do something that's really helpful for
• 32:27
  SPEAKER_01: A16z
• 32:42
  SPEAKER_01: A16z
• 33:04
  SPEAKER_00: So Matt,
• 33:05
  SPEAKER_00: one thing we haven't gotten into,
• 33:06
  and I know it's one issue that you've been spending a lot of
• 33:10
  time thinking about,
• 33:12
  SPEAKER_00: is how federal law and state law interact in a world like this.
• 33:18
  We've talked a lot about the appropriate realm and sphere of
• 33:22
  state law and the
• 33:22
  exercise of state police powers and the appropriate exercise of
• 33:26
  federal powers.
• 33:28
  SPEAKER_00: What's your view of how that debate plays in the liability context?
• 33:33
  So our general view here is that the role of the federal
• 33:36
  government and the role of
• 33:37
  SPEAKER_01: state government should track what the Constitution outlines.
• 33:40
  Constitution gives a role to each,
• 33:42
  but Congress and state legislatures are not exactly the same.
• 33:45
  Congress largely regulates interstate commerce while states
• 33:49
  largely police harmful
• 33:50
  SPEAKER_01: conduct within their borders.
• 33:52
  SPEAKER_01: We think that's the right balance generally.
• 33:55
  And the way that it maps on to AI is that Congress typically we
• 33:59
  think should take
• 33:59
  SPEAKER_01: the lead in matters related to AI development.
• 34:02
  SPEAKER_01: So if you're a
• 34:03
  AI Developer in Texas or Florida or New York or California and
• 34:07
  you're thinking
• 34:08
  about trying to build a product that you can sell across
• 34:11
  borders not just within
• 34:12
  your state but throughout the country and you're thinking about
• 34:15
  a product that is
• 34:16
  SPEAKER_01: going to position the United States
• 34:18
  SPEAKER_01: A16z
• 34:33
  SPEAKER_01: If you're talking about using deepfakes for fraud,
• 34:35
  SPEAKER_01: if you're talking about criminal activity,
• 34:37
  if you're talking about local issues related to consumer
• 34:41
  protection and unfair and
• 34:42
  SPEAKER_01: deceptive trade practices,
• 34:43
  SPEAKER_01: then that we think is really fully states' domains.
• 34:46
  That's the kind of legislation and enforcement that states have
• 34:49
  historically done.
• 34:50
  SPEAKER_01: And so we think they should continue to do that.
• 34:53
  SPEAKER_01: It does also map onto how to think about liability because if states are
• 34:58
  legislating and enforcing in ways that interfere with
• 35:03
  development activity,
• 35:05
  SPEAKER_01: such as the California bill that we talked about earlier,
• 35:08
  SPEAKER_01: 1047,
• 35:08
  SPEAKER_01: which would have
• 35:10
  It's a liability regime that puts states in a lead position
• 35:22
  with AI development and
• 35:24
  SPEAKER_01: we think that burdens interstate commerce potentially in ways that are
• 35:26
  unconstitutional but regardless of the legality of it in ways
• 35:30
  that we think are
• 35:30
  SPEAKER_01: really counterproductive for having an AI ecosystem that we want to have.
• 35:34
  SPEAKER_01: At the same time,
• 35:35
  states need to be able to enforce their law and bring cases and
• 35:39
  impose liability
• 35:40
  SPEAKER_01: for things related to harmful conduct within their borders.
• 35:43
  SPEAKER_01: That's an important part of the ecosystem.
• 35:44
  And if developers could somehow be immune just because they're
• 35:48
  using AI and be
• 35:49
  SPEAKER_01: immune for local harmful use cases,
• 35:52
  SPEAKER_01: then that would be problematic as well.
• 35:54
  SPEAKER_01: Jay, this is great.
• 35:55
  SPEAKER_01: Thanks for this discussion on such an important issue.
• 35:58
  SPEAKER_00: Yeah,
• 35:58
  I know these conversations are always great and I hope the
• 36:00
  audience found it
• 36:01
  SPEAKER_00: interesting and useful.