Latest / The Road to Accountable AI / Dean Ball: The World is Going to Be Totally Different in 10 Years

Transcript

Dean Ball: The World is Going to Be Totally Different in 10 Years · The Road to Accountable AI

• 0:00
  Dean, thanks so much for joining me.
• 0:02
  Thanks so much for having me.
• 0:05
  You're a great example of someone who became known as an expert on AI policy, largely through your writings and largely through your blog, Hyperdimensional.
• 0:16
  We're in an environment where anyone can create content.
• 0:20
  What do you think it takes to become influential in an area like this?
• 0:25
  ah It's a good question.
• 0:27
  um I think part of it was uh just sort of taking the time to develop a differentiated point of view on a topic, which sounds really simple or sounds kind of obvious, but in
• 0:45
  practice, most people don't do it.
• 0:47
  So it's obviously not easy.
• 0:48
  um And I actually wrote a...
• 0:51
  a post shortly before I went into the White House called How I Work, where I try to talk about like what my process is for developing a differentiated point of view.
• 1:00
  Because I do feel as though that's the single highest leverage thing.
• 1:03
  There's also some extent to which consistency I think matters.
• 1:07
  I think it's maybe somewhat overrated.
• 1:10
  I don't think you need to be writing every, I mean, I think writing every day as a practice is good, but I don't think you need to be publishing every day.
• 1:18
  And then,
• 1:20
  You know, there's also um luck is a part of it, right?
• 1:26
  Like I just happened to write about SB 1047, the California AI bill from 2024, which ended up being a very prominent debate.
• 1:36
  I was very early to that bill.
• 1:38
  um And, you know, I was talking about it probably four or five months before, um you know, it was a popular trending topic on Twitter or anything like that.
• 1:49
  And so I kind of had like a natural authority there.
• 1:52
  And that was mostly just luck.
• 1:53
  I just picked the right topic.
• 1:56
  And uh finally, I think it's, uh you I try to bring a literary sensibility as much as I can to the writing that I do.
• 2:07
  So I think it's probably the case that AI policy is like a very dry topic, even within AI, right?
• 2:14
  Like AI is a very exciting topic, but...
• 2:17
  policy is kind of this relatively dry cousin of it.
• 2:21
  But my view is that more think tank scholars in general should bring literary sensibilities to, and academics should bring literary sensibilities to what they do.
• 2:32
  And I don't think that makes your work any less serious because if anything, I think it makes it more serious.
• 2:40
  How would you describe what your differentiated viewpoint is on AI policy?
• 2:45
  Yeah, so I mean, I come from a classical liberal perspective, which is in and of itself, you know, actually committed classical liberal is kind of rare in the policy world.
• 2:59
  And I would say in general, particularly when I started writing around late 2023, you know, the attitude back then was very much like, it's not a matter of
• 3:15
  uh, you know, when, if we will regulate AI, it's a matter of when, and we're certain that there are like many problems that we have to solve through regulation.
• 3:24
  Um, and you know, my perspective was always like, uh, you know, I, I think that, I think that the people who are thinking that there's going to be like transformative AGI type
• 3:37
  capabilities soon, I'm very bullish on the capabilities of AI, but
• 3:42
  I think those people are underestimating the complexity of the world in a way that a classical liberal who is trained in thinkers like Friedrich Hayek would not underestimate
• 3:51
  the complexity.
• 3:52
  um And at the same time, there was this regulatory impulse that was coming both from, you know, one of the interesting things is that um there are a lot of people who say that AI
• 4:05
  is completely, like there's people like Gary Marcus who say AI is totally fake and not real.
• 4:10
  LLMs aren't even AI, it's totally just made up, but also for some reason we need to regulate it very heavily.
• 4:16
  um And then there's people uh who think that we're gonna have AGI soon who think we need to regulate it urgently.
• 4:28
  And I just thought all those perspectives were wrong, but then a lot of the regulation skeptic people like...
• 4:35
  Their argument for being regulation skeptic was basically the Gary Marcus argument.
• 4:39
  They were like, well, we just don't think AI is gonna be that big of a deal, right?
• 4:42
  Or like, still to this day, a lot of people rely on the crutch that AI is not gonna get much better than it is right now.
• 4:50
  um And that it's not gonna be that transformative and the world is gonna be the same.
• 4:56
  And it's like, no, that isn't true.
• 4:59
  It is gonna be transformative.
• 5:00
  The world is gonna be totally different in 10 years.
• 5:03
  And that's actually the reason that we need to be careful about regulating it, because we have no idea what to do.
• 5:08
  But no one was making that argument in an honest way.
• 5:12
  And I don't know if it's because they didn't connect the dots well enough or I don't know what the deal was, but I didn't feel like it was there.
• 5:18
  It was being articulated that much.
• 5:21
  And so um it felt like there was a niche to carve out there.
• 5:27
  And you already mentioned the California SB 1047, a lot of what you were writing about before going to the White House were these state efforts on AI regulation.
• 5:36
  We have, though, seen something of a sea change, right?
• 5:38
  1047 was vetoed by Gavin Newsom.
• 5:41
  The Virginia bill was vetoed by Governor Youngkin.
• 5:45
  The Texas TREGA legislation was scaled back.
• 5:47
  Even the Colorado legislation was adopted but was paused, and there's debates about scaling it back.
• 5:54
  So is the threat of
• 5:57
  patchwork of state AI regulation still something we should be concerned about.
• 6:01
  I think it still is, but in a different way.
• 6:07
  So, in 2022, 2023, and the first half of 2024, what you saw a lot of was states trying to do comprehensive AI regulations.
• 6:20
  So this would be like, Colorado has a bill called SB205, which focuses on algorithmic discrimination.
• 6:29
  That's a bill that's...
• 6:31
  quite explicitly modeled on the European Union AI Act.
• 6:34
  And I don't think anybody, um anyone, I I saw it, but I don't think that many people in policy making circles really saw how badly the Europeans had messed up.
• 6:44
  Now the Europeans themselves acknowledged this, right?
• 6:47
  Like they don't do it in public, but they do it behind the scenes in places like the White House, um where their elites will come and say, we know we messed up.
• 6:57
  We know we passed an absolutely idiotic bill and no one should be, you
• 7:00
  mimicking any aspect of this.
• 7:02
  And we're trying to, you know, they're trying to save face, right?
• 7:06
  They're embarrassed about what they did.
• 7:07
  um And, uh you know, but Colorado didn't realize that.
• 7:12
  And so Colorado dove right into the pool with the Europeans and they realized that the pool was not full of water, but instead full of macaroni and cheese.
• 7:19
  And now they're unsatisfied with their condition.
• 7:23
  everyone else who was a little bit more prudent, um who took the time to wait,
• 7:30
  is now looking at these people that have, these states that have passed quote unquote, well, just Colorado has passed the comprehensive framework.
• 7:40
  It was being considered in like 20 states at the beginning of this year.
• 7:43
  um So the threat was, I think, very real up to the beginning of this year.
• 7:48
  But everyone looked at Colorado and was like, that's not going well.
• 7:52
  And so now people are doing, uh at least for now, what I would describe as more targeted bills.
• 7:58
  um
• 7:59
  Those targeted bills are also very damaging.
• 8:03
  um Illinois and Nevada now have statutes on the books that as I read them um are essentially bans on LLMs.
• 8:13
  um A common theme is that policymakers do not know how to draft statutes with respect to um this technology.
• 8:20
  And so ah they do so quite incautiously.
• 8:24
  um so as a result, I think you're kind of looking at um
• 8:30
  ah you know, states that have passed these like dramatically problematic laws.
• 8:34
  I don't know how that's gonna go.
• 8:35
  This is the Illinois and Nevada stuff.
• 8:37
  California is doing the same thing.
• 8:39
  ah Yeah, so there are a lot of bills under consideration, you know, the bills I'm referring to are bills that ban chat bots from engaging in mental health services.
• 8:51
  But the way that mental health services is defined is such that if I say um to a chat bot,
• 8:57
  I'm feeling down today.
• 8:58
  Could you please tell me something to make me feel better?
• 9:01
  That functionality is banned.
• 9:03
  And I think that's both stupid and cruel.
• 9:04
  um But ah yeah, we've seen a transition to no less ill-considered for the most part.
• 9:15
  um No less uh problematic, well, somewhat less problematic, no less ill-considered.
• 9:23
  um
• 9:26
  more narrowly targeted bills.
• 9:28
  Okay.
• 9:29
  And just so people understand what's so bad with the European Union AI Act.
• 9:34
  After all, everyone except Meta and XAI signed the code of practice.
• 9:38
  Well, everybody signed the code of practice, which is a different thing from the AI Act, right?
• 9:44
  um So the AI Act, just for the listener's background, um the drafting of that began several years before ChatGPT, which should write in and of itself be an alarm bell,
• 9:57
  because what exactly was it about AI that needed regulation in the years before ChatGPT?
• 10:03
  ah That needed comprehensive
• 10:07
  continent-wide regulation.
• 10:09
  I fail to recognize what that is.
• 10:12
  um We didn't have it in America, and as far as I know, the republic stood.
• 10:17
  um so there was this sort of framework that was being developed for basically narrow machine learning systems.
• 10:33
  um They weren't really dealing with social media.
• 10:37
  through this law, that was more the Digital Services Act, right?
• 10:40
  That they wrote an entire bill to do social media.
• 10:43
  And then they wrote an entire bill to do Apple and Google.
• 10:46
  That's called the Digital Markets Act.
• 10:47
  um And then the AI Act is about like facial recognition algorithm, originally, right?
• 10:55
  That's its DNA is about like sort of narrow systems, not like these general purpose chatbots.
• 11:00
  And then toward the final innings of the AI Act, really, the bill was substantively already drafted.
• 11:06
  um Chat GPT comes out and all of a sudden the Europeans are in a terrible position because they're like, my God, like we didn't anticipate this coming and now we have to like write
• 11:16
  a whole new thing.
• 11:18
  Right?
• 11:18
  um So they wrote this general purpose, the set of general purpose provisions into the AI act, which are um some of the most half-baked provisions I have ever seen in my policy
• 11:32
  career.
• 11:33
  They are impossible to comply with.
• 11:36
  because they were written hastily.
• 11:38
  And basically it says, you have to be transparent and do a good job.
• 11:43
  And we'll tell you what good job means later.
• 11:45
  um With a general purpose chat bot, like, okay, thanks.
• 11:50
  That's really useful.
• 11:50
  uh The code of practice then was mostly written during a time period when the Europeans had realized how badly they had messed up.
• 11:59
  And so the code of practice really actually walks back um many aspects of the...
• 12:05
  uh
• 12:06
  the AI Act, for example, the AI Act general uh purpose, the general purpose guide or parts of the AI Act identify several dozen risk categories that they wanted companies to
• 12:19
  address.
• 12:19
  uh know, each one of these, by the way, each one of the risk categories is thousands and thousands and thousands of man hours for a company to address.
• 12:27
  uh Europe has no companies that would be subject to these provisions.
• 12:31
  So basically just
• 12:34
  basically just American companies and they identified several dozen.
• 12:37
  they spent a lot of, they allocated a lot of capital on behalf of American companies, which we don't really appreciate in the Trump administration.
• 12:45
  or I shouldn't say we anymore, but the Trump administration does not exactly appreciate and I don't appreciate as an American.
• 12:52
  But so the general purpose provisions are like one thing, but actually there's this whole other part of the AI act that is um on deployers.
• 13:03
  the deployer side provisions, which require algorithmic impact assessments at a million different um points in time and require audits and require all these things.
• 13:14
  And by deployers, I hate the word deployers, but we really mean are like, schools, your local bank, insurance companies, ah electricians, right?
• 13:27
  We mean just like random businesses in the economy.
• 13:30
  um
• 13:32
  that are trying to use AI.
• 13:33
  And because the Europeans and the policymakers in Colorado and other states as well really have no idea how the modern world works, um what they did was they wrote a law that
• 13:42
  applied to a lot of software that has already existed for 20 years.
• 13:47
  And so now there's a deadweight loss associated with software that was perfectly fine, that wasn't causing any problems.
• 13:54
  Businesses were using it.
• 13:56
  And now there are audit requirements associated with that software and there are algorithmic impact assessments.
• 14:01
  and other types of paperwork that are entirely unproductive.
• 14:04
  And this all dawned on the Europeans after they passed their regulatory fever dream, after they woke up from it.
• 14:10
  Now they're realizing how bad of a mistake they made.
• 14:13
  So that's sort of what I mean.
• 14:15
  And by the way, those deployers side provisions are not in effect yet.
• 14:18
  So we have not seen the deadweight loss, but the estimates that I've seen suggest that you should essentially model the deployer provisions of the AI Act as a roughly 20 to 30 %
• 14:29
  tax.
• 14:30
  I'm sorry, 15 to 20 % tax on AI.
• 14:34
  It's not exactly a tax the government doesn't collect it, but if a business spends $100,000 using ChatGPT, let's say, they have a customer service chatbot that does it, you
• 14:45
  should expect that that business will have to spend whatever they spend $20,000 on compliance costs with the AI.
• 14:54
  Okay, we could spend a lot more time on this, but I want to get on to your time this year in the White House and the Office of Science and Technology Policy.
• 15:01
  um First of all, what did you find that people who haven't spent time in government most fail to understand?
• 15:11
  Well, it depends who, you know?
• 15:13
  um What I would say to Silicon Valley ah is Silicon Valley basically models the average DC staffer and politician as kind of a sort of uh idiot child.
• 15:26
  um And like they don't understand that we are in fact savvy.
• 15:31
  The idea has occurred to us that um you may have an agenda.
• 15:37
  ah that it's actually not just technical information that you're imparting.
• 15:41
  Like we kind of got past the idea that you guys were like helpful nerds, know, 30 years ago, 20 years ago, right?
• 15:49
  um It's a pretty, it's actually a pretty tough town, Washington.
• 15:53
  um And there's a lot of extremely savvy and intelligent people here.
• 15:58
  And, um you know, I think sometimes Silicon Valley thinks they can play Washington for fools in ways that they can't.
• 16:05
  And what that often translates into is that Silicon Valley's messaging in DC and their tactics in DC are quite crass and come across to those of us who have been here for a
• 16:18
  while and understand policy and are a little more sophisticated.
• 16:21
  They really come across as kind of the Beverly Hillbillies coming into town.
• 16:25
  It's like, yeah, you're like new rich, it's like you're, it's like you have all this new money and you're coming in and you're throwing it around.
• 16:32
  And you kind of look like idiots to anyone who's been here for more than about 45 minutes.
• 16:36
  That's what I'd say there.
• 16:37
  To young people, what I would say is that I can only speak for, you know, the White House reinvents itself every four years, but I'm gonna speak for my experience in this White
• 16:46
  House, but it was a dramatically less bureaucratic place than I expected.
• 16:50
  And that was really cool.
• 16:52
  That was amazing.
• 16:53
  It was a really cool experience.
• 16:54
  People have no idea how small the White House is relative to the scale of the government.
• 16:58
  Yes, no, that's right.
• 16:59
  And the White House has a small budget.
• 17:00
  um This White House in particular, the Trump administration keeps a very small White House.
• 17:05
  The Office of Science and Technology Policy where I worked was about 25, 30 people um when I was there.
• 17:13
  So quite small.
• 17:16
  So the big thing that you worked on was the AI action plan that the Trump administration released in July.
• 17:22
  First of all, just talk about what are the basic elements of that plan?
• 17:28
  Yeah, so the AI action plan was tasked to the Office of Science and Technology Policy and a couple other White House components um by the president in his first few days in office.
• 17:43
  He rescinded President Biden's executive order on AI and then he signed a new executive order that um tasked us with an AI action plan.
• 17:55
  The AI action plan is supposed to be, you know, like a real statement of administration policy.
• 18:01
  Um, but the important thing about it from, you know, from our perspective is that it's not like a broad nebulous strategy document.
• 18:11
  It's instead a sort of very concrete set of recommended policy actions, um, to federal agencies.
• 18:22
  So it's things that executive branch agencies can do.
• 18:26
  with existing laws and existing budget.
• 18:29
  So no new money and no new laws have to pass.
• 18:33
  In some sense, that is a constraint, but I think it's also important because I think with such a broad policy area as AI, you do need to constrain yourself in some way and this
• 18:45
  ended up being useful.
• 18:46
  And so the action plan came out July 23rd.
• 18:50
  We had about six months, the president gave us about six months to work on it.
• 18:53
  um
• 18:55
  I, when I came into the office of science and technology policy, it was a blank sheet of paper and, um you know, I and others on the team brought it to uh its completion.
• 19:08
  And um it has three primary pillars.
• 19:13
  Pillars are innovation, infrastructure, and international security and diplomacy.
• 19:17
  Very, very quickly.
• 19:19
  Innovation is about not just making sure that
• 19:25
  Um, you know, we don't over-regulate the technology itself, but also ensuring that there is not regulation that stands in the way of positive and productive AI deployment and by
• 19:37
  businesses and also by government and also within the sciences.
• 19:42
  Um, there's a lot of stuff that falls into innovation, but more or less the way to think about it is like, what are all the things government can be doing to lay the groundwork?
• 19:51
  for America to lead the world in transformative AI adoption.
• 19:55
  Infrastructure is all about data centers, power, and the workforce that's necessary to build them.
• 20:00
  And then finally, you have the uh international security and diplomacy, which deals with um international alliances of all different kinds.
• 20:11
  So first of all, export promotion, which is one of the signature efforts of the action plan.
• 20:19
  It's the idea that
• 20:20
  America wants to build uh infrastructure for AI, AI data centers, and also export the full AI stack to allies and partners all over the world so that we have as many people over the
• 20:33
  world as we can who are uh building their economies on the American AI stack.
• 20:41
  have uh those also things in there though about export controls and about security, about
• 20:47
  managing and mitigating uh potential risks from AI that could emerge from foreign uh adversaries.
• 20:56
  And that's both countries and also groups, terrorist organizations and things like this.
• 21:02
  So that's sort of a high level sketch.
• 21:04
  Great.
• 21:05
  And obviously there's tremendous differences in many ways between this administration and the prior one.
• 21:11
  But what exactly did the Biden administration, in your view, get wrong on AI policy?
• 21:16
  Because you look at the specifics of the Biden executive order and the Chip Act and things like that.
• 21:21
  There's a lot of areas where there's continuity, I would think.
• 21:24
  So what do you see as really the biggest pivot?
• 21:28
  Yeah, uh think there's a couple things.
• 21:32
  First of all, on a rhetorical level, the Biden administration spent a lot of time telling other people what to do ah and what not to do with AI.
• 21:44
  So they came out with this, I remember this absolutely obscene document called the Blueprint for an AI Bill of Rights, which uh someone's...
• 21:54
  in that administration decided it would be a wise idea to spend taxpayer dollars on.
• 21:59
  And that person, whoever it is, should not be very proud of themselves.
• 22:01
  um Hugely wasteful, uh silly document that just goes around and it's like a bill of rights.
• 22:07
  We already have a bill of rights.
• 22:09
  Why do you need an AI one?
• 22:10
  We don't have a computer bill of rights.
• 22:11
  um And just sort of presumes AI of being guilty of all sorts of things with absolutely no evidence for it.
• 22:18
  And then tells the American people that they should feel a guarantee to
• 22:23
  all these things that are by no means guaranteed by our constitution, by no means guaranteed by our laws that have passed, but that have been passed by Congress.
• 22:32
  And that would be implicitly fulfilled by AI companies.
• 22:36
  And they have no authority to do that.
• 22:38
  And they put this document out and they just basically told a bunch of other people what to do.
• 22:42
  ah Same thing is true.
• 22:45
  I think there was actually a fair amount of sensible stuff in president Biden's executive order on AI.
• 22:50
  But what I would say is that the
• 22:52
  way that they talked about it and the posture with which that executive order was released was very much focused on control.
• 23:00
  The idea that we need to control this technology and it's our decision, you know, is open source okay, we'll let you know in six months was kind of the mentality.
• 23:09
  And it's like, from my perspective, um I didn't ask for my government to make that decision, actually.
• 23:16
  um I really didn't.
• 23:18
  um You know, I don't especially care.
• 23:20
  what the federal government thinks about open source because it already exists.
• 23:24
  And like, if you need to stop it, well then okay.
• 23:26
  Like you're gonna have to sell me on that though.
• 23:29
  It's not like you have some decision-making power here.
• 23:32
  So there's a profound arrogance.
• 23:34
  The diffusion rule, very similar.
• 23:36
  We went from, know, we're gonna keep a small number of very advanced AI chips from uh China's military to we're gonna regulate the global diffusion.
• 23:50
  of a multi-trillion dollar chip industry.
• 23:52
  uh That's where we went, that was about two, two and half years uh in the Biden administration.
• 24:01
  Just quite a rapid trajectory.
• 24:04
  um There are a lot of sensible things the Biden administration did at the same time, right?
• 24:09
  Like I don't dispute that.
• 24:11
  And I know some people who were the architects of president Biden's AI policy and a lot of them are really, really intelligent and patriotic people.
• 24:20
  You know, don't mean to, I think they were part of a system that couldn't help itself, but be ham-handed and arrogant.
• 24:28
  I don't think they themselves are, those, people that I know.
• 24:31
  um You know, so where are we different?
• 24:36
  The Biden administration had a lot of fantasies about, you know,
• 24:42
  the training of AGI.
• 24:44
  They thought it would be like, there's one day where we wake up and the day before there was no AGI and then we wake up and there's AGI.
• 24:50
  And AGI was like made in a specific building that's gonna exist somewhere in the world.
• 24:56
  um And that, you know, it's gonna come out of that building, right?
• 25:01
  Almost like physically.
• 25:02
  um And so they were very concerned about the security of that building and the security of the chips in that building.
• 25:10
  And...
• 25:11
  They talk, there's just all this foot and you know, the entire strategy of the export controls and the diffusion rule, 10 to the 26th flop reporting requirement in the
• 25:20
  executive order is all based on this idea of, you know, AGI getting trained one day.
• 25:28
  You know, but there was very little that they asked about like, what happens after that?
• 25:34
  It's just like, oh, it's like a nuclear bomb and now it exists.
• 25:36
  And so no one's gonna mess with us ever again.
• 25:39
  I think that's a very, very, very um sort of childish way of thinking about the way that AI will diffuse throughout the world.
• 25:50
  And so the AI action plan from the Trump administration, maybe one of the most commonly used words in that whole document is adoption um or deployment or diffusion.
• 26:04
  um We really care quite deeply
• 26:09
  about the idea of deploying AI in transformative ways, in regulated businesses, within government, as I said, in small businesses, all over the economy.
• 26:22
  Because that's what matters.
• 26:24
  It doesn't matter if you have a really, really performant model and if we call it AGI or whatever, it doesn't matter if no one uses it.
• 26:34
  Or it doesn't matter if we do it and then a year later,
• 26:38
  China builds, or some other country builds a competing product and their society is more supple and more open to adopting AI in a transformative way, then um we lose.
• 26:53
  And it didn't matter that we were six months first or a year first to the model weights.
• 26:59
  Who cares about that?
• 27:02
  As you've written, though, there are also tensions on the right in these areas.
• 27:07
  There's a lot more we could talk about in terms of the Biden administration, but I want to understand going forward.
• 27:13
  There are those in the Trump coalition who want to attack the AI companies or who are particularly concerned about the DEI issues and some of those topics.
• 27:26
  So the AI action plan
• 27:30
  thread of that needle within the administration, but going forward, how do you think those tensions get addressed?
• 27:36
  Well, think there's gonna be, I think one of the really key areas of tension is gonna be on the kids safety issue, um where Silicon Valley um is not very familiar institutionally
• 27:53
  with the concept of tort liability, the concept that their products may cause physical harm or loss of property to someone in the world.
• 28:06
  and that they may have liability for that harm.
• 28:10
  um That is not something that happens.
• 28:15
  You can't really sue Google for much of anything.
• 28:21
  Google Chrome can't physically harm you.
• 28:24
  um If you Google for, ah how do I uh make a bomb?
• 28:34
  You know, you can't sue Google for giving you that information, right?
• 28:36
  Google is a sort of neutral conduit, but it's different when you have a chat bot that is controlled by a company that is making um these statements or sometimes taking action on
• 28:49
  your behalf.
• 28:50
  And especially as they take more action, I think this liability issue is going to become very significant.
• 28:56
  I talked to executives at AI companies who kind of just looked at me funny when I said this a year ago.
• 29:02
  um
• 29:04
  And now there have been some real tort liability lawsuits in the last six months or so against AI companies, including against open AI, like not only, especially there's a
• 29:16
  recent one, a very tragic one involving a 16 year old named Adam Raine who killed himself sort of in dialogue with chat GPT.
• 29:26
  um This is not a legal advice.
• 29:31
  I'm not a lawyer.
• 29:32
  But like, I don't know, I read that complaint and I'm like, I've read a lot of complaints.
• 29:37
  I've read a lot of tort complaints in my life.
• 29:38
  Like, I think open, I think they got open AI.
• 29:43
  I think open AI is, I think there is some trouble on this one.
• 29:46
  um I don't think that like, there's a sort of contingent on the right that is sort of the, you're like, okay, sort of more libertarianish, right?
• 29:56
  We used to call it libertarian, now we call it new tech, right?
• 29:59
  But it's basically the same thing.
• 30:00
  They say, they claim like there's a difference.
• 30:02
  And the difference is that they say, libertarians don't say nice things about Donald Trump and New Tech right does.
• 30:08
  That's the difference.
• 30:08
  um But like, they have no idea.
• 30:11
  They don't really even know what common law liability is.
• 30:14
  They don't know what it is.
• 30:16
  They don't want anything to do with it.
• 30:17
  But unfortunately it's here for them.
• 30:19
  And um I think as a general matter, these folks tend to, as I said, they have pretty bad political instincts and they tend to overestimate the power that they have.
• 30:28
  um
• 30:30
  So uh that's gonna be interesting.
• 30:32
  And then there's a lot of people on sort of the right, right?
• 30:34
  Like social conservatives and whatnot who come from, uh they've been fighting oftentimes battles on social media, not on social media, but about social media with respect to the
• 30:45
  impact of those technologies on kids in our society more broadly.
• 30:49
  They've been oftentimes like urging regulation, urging social media bans of various different sorts.
• 30:57
  I'm agitating for it at the state level and sometimes having been successful.
• 31:02
  Those people are continuing that same battle pretty much with AI.
• 31:12
  um you know, it's a little bit, I mean, it's a classic example of you find what you look for, right?
• 31:23
  So like,
• 31:24
  there's already this huge contingent of our society that's advocating for child safety issues.
• 31:29
  And um so we found child safety issues, no surprise.
• 31:34
  If that same group people had been advocating for like pet safety issues, we would be talking about AI providing veterinary care and providing bad veterinary advice, right?
• 31:43
  It isn't to say that the kid safety issue isn't real.
• 31:45
  It's just to say that like public policy as a general matter is the...
• 31:50
  It is the practice by which we decide whose harms we care about and whose we're gonna pretend don't exist.
• 31:56
  um But we should never pretend like we care about everybody's harms, because we don't.
• 32:00
  We care about a very, very small number of the harms that happen in the world.
• 32:05
  But they're agitating about that, and I think they're right to, by the way.
• 32:09
  um So that's gonna be an interesting conflict, how that all emerges.
• 32:15
  there's a sort of more pro-regulatory instinct on the right, and then there's a...
• 32:18
  a very anti-regulatory instinct.
• 32:20
  And I think the pro-regulatory people are uh oftentimes uh motivated more by animosity toward the people, toward their perception of the people in Silicon Valley.
• 32:32
  And I think that the anti-regulatory people have a very simplistic model of the world.
• 32:36
  So neither of those is going to work.
• 32:37
  And so you have to figure out like, um what's um a compromise here?
• 32:45
  Are we in a position?
• 32:47
  I want to ask you about one compromise may not be the right word for one approach that you've advocated for, which ties into the issue of liability, which is private governance
• 32:56
  mechanisms, kind of hybrid.
• 32:59
  And I know this is not something that was directly in the AI action plan, but can you talk about uh going forward, how you see that option developing and why is the right way to go?
• 33:11
  Yeah, so I mean, I think we're already seeing it happen.
• 33:14
  m You know, the great thing about private governance, which is the idea of things that implicate governance being done not by the formal public government, but instead by
• 33:25
  private organizations.
• 33:26
  um There's nothing radical about private governance.
• 33:29
  It's been done for a long time.
• 33:32
  And I mean, if you wanna be a real, know, uh pedant about it, families are a form of private governance.
• 33:39
  right, being a father is a form of being a private governor um or a mother.
• 33:44
  But um the, I think,
• 33:52
  you know, there's what I would call private governance is the idea that like, there is some service that can be provided for which there is supply and demand and which can
• 34:03
  therefore be provided at least reasonably well in a traditional market process without government interference.
• 34:12
  A good example of this would be like insurance where it's like, I have a risk.
• 34:18
  I would like to lower that risk and I would like to share that risk with other people and I will pay someone for the service of taking on some of my risk.
• 34:29
  That's what insurance is.
• 34:30
  That didn't emerge because someone in government designed it.
• 34:33
  That emerged uh centuries ago because it was necessary to facilitating trade and facilitating risky ventures.
• 34:45
  So you're starting to see these sorts of things happen.
• 34:47
  Like there's a nascent AI specific insurance industry that is developing.
• 34:53
  And that to me is a very exciting thing.
• 34:55
  um Then, at the same time, ah there's um private, there's what I would call like a more formal way of dealing with this, which it might be to say, the government says, hey,
• 35:12
  there's some technical thing about AI that-
• 35:14
  worried about, but we don't want government to be the one creating the standards because we don't think government has the technical capacity or we don't want to spend the money
• 35:23
  or whatever else.
• 35:24
  What you might be able to say then is, hey, if a private body were to do this in a way that we, the government approved, you know, private body could, maybe it could charge the
• 35:35
  companies or maybe it could receive private donations, act as a nonprofit, then maybe we would be comfortable privatizing.
• 35:43
  really more privatizing as opposed to private governance, more like privatized governance, we would be comfortable privatizing some aspects of basically technocratic rulemaking for
• 35:53
  this specific thing.
• 35:54
  And in exchange for a company adhering to the standards put out by that private body, the company would get some benefit.
• 36:01
  Maybe the benefit is um more favorable treatment by courts in liability conversations.
• 36:09
  Maybe it's something else, but that's kind of the idea.
• 36:11
  I think both of those things are very promising.
• 36:13
  The first one is already happening.
• 36:15
  The second one, whether or not it catches on, we'll see.
• 36:20
  To be continued, think that's a great place to stop.
• 36:22
  pleasure speaking with you.
• 36:24
  Thanks so much for having me.
• 36:25
  was a great, it was a blast.