Scaling Laws: Is Meta's Oversight Board a Model for AI Governance?

[Intro]

Alan Rozenshtein: It's the Lawfare Podcast. I'm Alan Rozenshtein, associate professor of law at the University of Minnesota and a senior editor and research director at Lawfare. Today, we're bringing you something a little different, an episode from our new podcast series, Scaling Laws. It's a creation of Lawfare and the University of Texas School of Law, where we're tackling the most important AI and policy questions, from new legislation on Capitol Hill to the latest breakthroughs that are happening in the labs. We cut through the hype to get you up to speed on the rules, standards, and ideas shaping the future of this pivotal technology. If you enjoy this episode, you can find and subscribe to Scaling Laws wherever you get your podcasts and follow us on X and Bluesky. Thanks for listening.

Intro Voices: When the AI overlords take over, what are you most excited about? It's, it's not crazy, it's just smart. And just this year, in the first six months, there have been something like 1,000 laws. Who's actually building the scaffolding around how it's gonna work, how everyday folks are gonna use it? AI only works if society lets it work. There are so many questions have to be figured out, and- Nobody came to my bonus class. Let's enforce the rules of the road.

Kevin Frazier: Welcome back to Scaling Laws, the podcast brought to you by Lawfare and the University of Texas School of Law that explores the intersection of AI, policy, and of course, the law. I'm Kevin Frazier, the director of the AI Innovation and Law Program at Texas Law, and a senior editor at Lawfare. I'm joined today by my guest co-host, Kate Klonick, an associate professor of law at St. John's University and a senior fellow at Lawfare.

Today, we're joined by Kenji Yoshino. Kenji is the Chief Justice Earl Warren Professor of Constitutional Law at NYU and a member of Meta's Oversight Board. He recently authored a piece in Tech Policy Press with fellow board member Ronaldo Lemos, arguing that frontier AI companies should learn from the Oversight Board's approach to independent governance. But that begs the question of what exactly we should and shouldn't learn from the Oversight Board's track record, and Kate and I dive exactly into that with Kenji.

To get in touch with us, email ai@law.utexas.edu or follow us on X or Bluesky. Also, good news, we were recently ranked second in the Feedspot AI Policy Podcast rankings, which is super exciting. But as my dad and Ricky Bobby would say, "If you're not first, you're last." So if you like this podcast, please go leave us a review, hopefully five stars, and help us make sure we get that number one spot. And with that, giddy up for a great show.

[Main Podcast]

Kenji, welcome to Scaling Laws.

Kenji Yoshino: Thank you so much for having me.

Kevin Frazier: So everyone's heard of the Oversight Board, or presumably everyone who listens to Scaling Laws has heard of the Oversight Board. But just in case that their mind has been occupied with other things for a little bit, before we dive into how the Oversight Board model may apply to the AI space, Kenji, can you remind everyone, well, first and foremost, what your role is on the Oversight Board, and then just a brief overview of the Oversight Board's origin and jurisdiction?

Kenji Yoshino: Fantastic. Although in terms of origins, Kate might be the global expert on this, so I may have to defer to her and her excellent work. So the Oversight Board is a group of 20 individuals who are experts in some domain of social media regulation. And the board has been in existence for six-plus years, and our job is to evaluate Meta's content moderation decisions against both their own stated standards and against international human rights law.

So I'm a constitutional law professor in terms of my day job, so Kevin, I always think about this a little bit like judicial review, right? So, like, the community standards that Meta itself promulgates are like statutes. So the first order of business is to say, does this post kind of conform with the community standard or not in terms of the ultimate decision that was made about it? But then there's a kind of secondary question where, even if the community standard was kind of totally executed to the jot and tittle and was kind of perfectly done, the community standard itself could violate international human rights norms. And so, if you have, you know, a lack of notice, or if you have a violation of the rights of a protected group, or there's no proportionality or some such, then we could say that the community standard itself needs to be tweaked. And so it's really that two-tier decision-making that we engage in.

This is a really, present company excluded, extraordinary board. It is 20 people around the globe, and it's been tremendously exciting really. You know, I know, and we will get to, I'm sure, the criticisms of the board, my knock on the board, but it really has been one of the great kind of pleasures and honors of my professional life to be able to, sort of, engage with 19 other people from around the world to try and solve a global problem with regard to social media.

And going back to the origins of the board, and Kate, you know, if you have anything to add, please do. We didn't really begin from the city on the hill. We didn't begin with idealism. We, we began out of utter crisis. So this is a moment when Cambridge Analytica was happening, if you can cast your mind back there, or where the Rohingya massacre was being attributed in part to Meta's lack of care with regard to its policies. And so we were really created in order to create some guardrails for the organization. And so oftentimes, this is a point of irritation for everyone on the board, we're kind of conflated with Meta, but we really regard ourselves as being independent of Meta and being Meta's watchdog. So what have I left out?

Kevin Frazier: Well, there's, there's only so much we can dive into for the sake of a single podcast. So I think you've done an excellent job of setting the initial scene. One thing that I'd love to explore a little bit further, as you raise in your recent tech policy piece from July 7th that you issued with your fellow Oversight Board member Ronaldo Lemus- Lemos, excuse me, was the fact that you mentioned in that piece as well, as you explained here, that the Oversight Board is continually a sort of work in progress. And of your accomplishments or of the actions you all have been able, able to take, you've issued more than 200 decisions, and you've also issued more than 300-plus recommendations.

And so as we're exploring this as a potential model to apply to AI labs, and more generally the idea of private oversight or independent oversight of any other entity, can you delineate the difference between when you all issue binding decisions and when you all issue recommendations, and to what extent those recommendations have actually informed Meta's practices?

Kenji Yoshino: Great. Yeah. Thank you so much for saying that, because I should have included that in my intro. So the original commitment that we have according to the governing documents that Meta created was that on leave up/take down decisions we are binding. And so if someone says, you know, "I think all trans people should kill themselves," which was one of the cases I'm proudest of where we said that actually is in complete violation of our suicide and self-injury policies as well as at the time your anti-trans policies, so you, you have to take it down. They've always abided by that. So in that sense, a leave up/take down decision is binding, and as you say, there have been 200 of those.

Sometimes we all get a little bit hot under the collar. I will say, you know, that when we, we told Meta that it needed to take down the Hun Sen post. So this is the then prime minister of Cambodia inciting violence, in our view, against his political opponents in a long speech that he gave. You know, Meta had a very different view, you know. And, you know, in fairness, it was a, you know, it was a totally arguable position to say it was one clip in a, in a long video, and he said, you know, "We'll just take the bat to them, you know, if, if they don't, if they resist us," and is taking the bat to someone really incitement to violence or not, et cetera, et cetera. But if you looked at the whole record, we believe that this was an incitement to violence, especially against past history and the like, and so we said take it down.

And that was the moment, you know, I think every kind of governmental official or quasi-governmental official, even an administrator, has those kind of make or break moments where you make a decision, you kind of hold your breath. You know, Justice Scalia used to say this about Bush v. Gore, where he, you know, issued the opinion, and he was so proud of the court that the country followed. And this is a moment, not to compare ourselves to the Supreme Court or to that momentous decision, where we did hold our breath because we thought, you know, Hun Sen could react to this by shutting down the entire internet in the country, or he could respond to it by kicking Facebook out of Cambodia. And so it was really something that cost the company something, but the company still abided by that. So, I'm incredibly proud of both ourselves and of Meta for abiding by that agreement to say if it's a leave up/take down, it is binding, and Meta has never betrayed us on that.

With regard to the recommendations, there is a re- requirement to respond to the recommendations, but not to take the recommendations. So they're forced into dialogue with us, but they're not forced to do anything. That said, I hope because we are kind of both reasonable people on, on both sides they've taken over 75%, by our count, of our recommendations. So usually you can argue or make the case for some kind of change. You asked for an example. You know, labeling of manipulated media was something that kind of shockingly they weren't doing, you know, until we pointed it out in the Biden manipulated media case, and so now that is just, you know, standard there in the Meta social media universe. And so that's, you know, Instagram, you know, and Facebook and recently we've, we've moved into Threads and, and other areas as well.

So, and if I, if I could say one more thing, like in the, when I was actually interviewing for the board, and it seemed like there were like 12 layers of interviews. I've never gone through so many interviews. So it's the final point, telling tales a little bit. I was, I was like, you know, take me or leave me, but like, I'm not doing any more interviews. Like, you know, you basically know me better than, you know, distant family members know me. So, you know, at this point, like we really have to fish or cut bait. So they took me, thankfully, and I'm very glad that they did. But one of the things I kept asking was, "Are we binding? Are we binding? Is there real teeth? Is there legitimacy here?" Because I don't wanna be some kind of heat shield for Meta. I don't wanna be some kind of ornamental. You know, it's a lot of time. You know, I don't wanna be some kind of, you know, ornamental kind of, thing, you know, that they're just attaching to make themselves look better. And the insistence was that we are binding, we are binding the up/down decisions.

That said, I will say that if I were to measure our impact under kind of truth serum, I would say that our biggest impact has been with the recommendations, not with the up/down decisions, because so many of the up-down decisions are just like knuckleheaded things that somehow flew under the wire, right? So, you know, the trans case, again, going back to that would be a really good example. If anyone had really properly looked at that, they would very clearly see that it violated their community standards. They didn't really need us to tell us to do that, right?

But the recommendations tend to go to the heart of how things like that happen. So we would ask things like, "Are you aware enough of what transphobia looks like?" Because there's a kind of malign creativity going on there where it wasn't this naked, "Oh, all trans people should kill themselves." It was a curtain that was hung and the caption was, you know, self-hanging curtains, right? So if you put it all together, it was pretty clearly, you know, trans people should kill themselves. But our question was, are you doing enough internal training of these moderators so that they know that this kind of malign creativity goes on? Because it's always this escalation war where you create a policy and then someone tries to evade it.

So even though the thing that I was focused on before I stepped on the board was just this hard line, it is so ordered, like, do you have to carry out the order? Ultimately, I think what, you know, Joe Nye calls soft power has been much more effective in terms of us moving the giant battleship that is Meta in one direction or the other.

Kevin Frazier: It's worth commenting and kind of laughing about the company that once said, "Move fast and break th- things" being described as a battleship. But I'll, I'll, I'll leave that for another podcast to dive further into. Kate, I know you had some additional questions.

Kate Klonick: Yeah. I just, you know, after years of watching the Oversight Board get set up but then of course, watching it from afar and kind of, hearing, you know, back channel through people that were at the Oversight Board, people who left, people who are still there kind of a lot of the things that you're saying, Kenji, really resonate. And I, I think they have the ring of truth to me because I hear them from so many different people, and I hear them kind of in, like, mostly the same way.

One of the main criticisms of the Oversight Board has been that it takes such a small percentage of cases, which I always just think has been so misguided. Because in the practicality, like, maybe you take, like, one, one you know, one out of every 100,000 cases if that, that actually, like, a lot of these things are in error. Like, the actual error. Like, they just enforce their, their, you know, the Facebook enforced the rules incorrectly and are happy to kind of, to, to reverse themselves and, like, in, in the small instance.

But what I have been really surprised by and I wouldn't have predicted is the value of recommendations. To send this signal, and also to consolidate stakeholders on the other side of things, which I just think is, and I, I, I put that down as a flag because I want to actually talk about that role that it could have in AI governance. Because there is, there's all of this talk about multi-stakeholder governance. There's all of this kind of, like, ideas, and it sounds so soft, frankly. And it is, because you just have people that happen to have access to these groups and this, like, way of doing things, and it's people who can set up an organization or people, and, like, some people who have connections to Facebook or some people who don't.

But the Facebook Oversight Board, the Meta Oversight Board ends, ended up being a really, I think, overall, one of the things that I've been most struck by is when I go from stakeholder to stakeholder that engages with you, and a lot of people were skeptical of it, that at the five-year anniversary, a lot of them were like, "No, we actually really love it." Like, "We actually really love it."

And that, shockingly, was not just stakeholders that had been skeptical. I heard that from people that were at the upper echelons, the top 50 people in Meta, saying that they had opposed it the entire time that I was inside the company covering it, and I heard about how much they were trying to kill it. And now they have, like, had this about-face. Because actually what it is is this really wonderful way to surface more kind of reflective norms that represent not just human rights norms, but norms generally and law in such a comprehensive, transparent way.  And, you know, the otherwise how are, how are they gonna do that? The policy team loves you guys 'cause you basically do all this incredibly hard thinking for them, and you're the right people to do it. Like, right? You know, you're thoughtful. I think you are. Like, you're thoughtful, like, really well-reasoned people with great train- diverse training in these areas from diverse parts of the globe.

But that being said I do think that one of the things that I'm super interested in is that you said at the beginning that the entire point of setting up the Oversight Board, and I think this is kind of something that people miss, was that you were, there was a response to this idea that Facebook was in this reputational hole, right? Facebook had no other choice in 2018 when Mark announced that he was going to start setting up this board than to try to kind of dig themselves out of this, and he had very little to lose by setting this up and kind of being responsive to this. And it was imagined as a group that was gonna go forward and maybe be a model for all of these other platforms for governance, and then it didn't happen. I talked to a lot of places. TikTok you know, TikTok, a few people informally at YouTube, a couple places I probably can't mention, but like some, a lot of places that considered setting up oversight boards, and the main thing was like, "We just are not hemorrhaging, like reputationally, the way Meta is. We don't have to do this."

So fast-forward to like the AI movement, and so I'm really kind of curious like what you think. You know, Anthropic has set up this long-term long-term benefit trust. I, I, I really take your point, but what has changed? What do you think has changed? What do you think this mo- about this moment has possibly changed that would create the incentive to do, for lack of a better term, the right thing. Like, to really set up these, these responsive arms of these incredibly, incredibly important public-facing companies.

Kenji Yoshino: Yeah. And we've talked as well about Jack Balkin's piece, you know, “Free Speech is a Triangle,” right? And so thinking about not just governmental actors and speakers, but these companies as being incredibly important, you know, actors in the speech landscape. So, you know, everything that you said was music to my ears. You know, thank you for your kind words.

But I also totally agree with you that my biggest worry is that we are not gonna see the groundswell of activity that we saw to create the Meta Oversight Board until there is an analogous crisis in the AI domain. So, you know, one could argue that there have been some crises in the AI domain. I'm thinking about the allegations about, you know, chatbots encouraging self-harm and suicide and, and the like. But I think it's gonna take something of an even larger scale, you know, whether it's, you know, national defense or something really wide scale that is gonna be the wake-up call that says "We, we need to create a board."

So it's just, you know, sadly human nature. Like, we can have the most idealistic and best intentions, but I don't think we could get our act together institutionally until the kind of rubber hits the road. And so as I said earlier, we were not born out of idealism. We were born out of utter crisis. And my sad prediction is that we'll take a similar, you know, event, right, in the AI domain for people to wake up and say, "Well, wait a minute, who is, you know, holding people accountable here?"

Kevin Frazier: So, Kenji, there's so much we can dive into with respect to whether the Oversight Board maps well into the AI space. First and foremost, I think the most obvious issue is content moderation is one thing, right? And as you noted and as Kate noted, calling a, a, a strike or a ball in the context of should you be able to say you want to kill trans people, for all intents and purposes as we've discussed here, pretty dang easy to decide that question. Yes, that is a horrible use of the platform and is prohibited under international law and the community standards. For something like should we develop an AI agent that's going to replace 10,000 jobs in a period of three days, this is obviously a over a over-exaggeration, but for those sorts of decisions, having anybody that has the requisite degree of expertise and background and knowledge and time and resources is incredibly difficult.

So you and Rinaldo highlighted three things that you wanna see from anybody that is in this kind of independent regulatory space for AI. First and foremost, you highlight having a credible body of overseers who are both independent and representative. Second, you say that they need to have meaningful oversight such that there's a credible external body of standards. And then third, you say that there needs to be transparency about what the decisions are, how they're made, and how they're responded to by the company. So we've got this nice three-part framework, and I think just about everyone listening to the podcast would say yes, yes, and yes. All of those things sound great.

But of course, as you all have experienced at the Oversight Board and as anyone who develops policy has experienced, the difficulty is that rubber meeting the road mentality. And for something where right now as listeners have heard me say on many occasions, we're still at the national anthem of this ballgame, right? We haven't even thrown the first pitch. Only 3.3% of Americans have paid subscriptions to AI. So we don't even know what problems and what crises may be ahead. You know, there, there's very real issues right now, but it's gonna get so much wilder.

And so I wonder, you said you're one of 20, and you all have a pretty substantial budget, all things considered. We can get into funding perhaps later. But, do we need five different oversight boards? Do we need 15 different oversight boards? Or do you imagine a sort of single entity being able to play this role, not only for a single lab, but perhaps across labs? Because one other thing that's interesting about the Oversight Board's origins was some degree of optimism that the model would then spread to other platforms, and that you all would become the oversight board for Facebook and for this other social media platform and so on and so forth. So let's just start with do we need multiple oversight boards, or is this a model that you think can scale in some other way to govern AI?

Kenji Yoshino: Yeah. That's so amazing. Thank you for that. So let me begin 'cause there are two things that tie together. So one is you said in our three criteria perfectly correctly that one of our criteria was that there has to be some, well, first of all, credible body of people, external body of law or standards that the company can't change on its own. And then finally, like transparency, right? And it's that transparency that did us in. And you know, Kate, you know, you know this better than anyone.

But you know, I kept thinking rather naively like, "Oh, well, I, you know, I know a lot of people at Google,” you know, through other, through my, you know, diversity and inclusion work. You know, “I know, I know a lot of people at, you know, other companies that we can reach out to, LinkedIn, you know, PayPal. Let's reach out to those people.” And of course, we got sort of one door slammed in our face after another. And as I thought about it, I thought like, A, why would they join an organization that was seen as Meta's creature because Meta would-- first of all, there are confidentiality issues and like how good are the firewalls, et cetera, et cetera. And then second of all, there's this notion of any kind of credit really redounds to Meta in the first instance, and so why would you allow a competitor to take a victory lap around you?

But then the last thing that I didn't think of was that we were victims of our own principles because we thought, you know, this has to be a public good. It has to be a published set of opinions that anyone can read. Like we can't ask people to pay for it. We can't hide it. We can't just send confidential advisor opinions to Meta. We have to, you know, publish it to the world. And what we realized is that people-- and I know this because I've had direct conversations with people who say like, "We love your opinions. We totally free ride off of you. There's absolutely no reason why we need to sign up for Meta even without the first two issues because this is a public good and we could just, you know, download it and, you know, learn what we need to learn from it, and then we're done," right? So they get all the benefits without any of the downsides.

So in an ideal world, right, this would be baked in from the beginning. So the question of how many oversight boards there needs to be, in fact, is an outgrowth of that point which is to say if individual companies like let's say Anthropic, you know, says tomorrow, you know, we want to let other people into this public benefit trust kind of arrangement, and it really needs to be you know, a multiple company oversight Institution, our experience tells us that that's not gonna work, right?

So, if that's the model, then you're gonna need one per company, right? Which isn't great, right? So in an ideal world, you know, much as I hate criticizing my own institution, if the Oversight Board were starting today, it would not be solely Meta-funded. Like, everyone would pay into it, understanding that the Rohingya crisis analogs are gonna be spreading all over, you know, other platforms as well, right? And then it would represent many platforms.

Even then, I don't think it should be one body because I don't think, right, you know, one body should have that much authority or power, and there should be some give and take and bricolage. And of course, you know, I didn't say this earlier, but I'm a firm believer to one of the questions that you put to me in, in, in your brief, that self-governance and, you know, regulation are not mutually exclusive. Like, I, I absolutely believe that we're operating in the shadow of the law and that regulation is ultimately gonna need to be a huge, you know, factor, right, in the regulation of AI.

Kevin Frazier: Right, and I think highlighting also that there is some degree of self-governance that's already underway in the AI space. So we've had OpenAI's model spec, Claude's constitution as these documents that attempt to identify red lines for their models as well, as well as values to embed within these models. When you go and talk to the folks who were charged with drafting those documents, though, they'll admit, "Hey, you know, it was a pretty insular group in a pretty ad hoc fashion that went about developing these safeguards." And as you've flagged also, there's a bit of an awkward moment when you're charged with interpreting the own rules that Meta has promulgated and the community standards that they've promulgated.

And so, when we actually analyze to what extent is self-governance right now doing its job, I think that's a key question we have to grapple with because when folks hear Claude's constitution or model spec, they think, "Oh, wow, they're doing it. They're doing the thing. This is great. We're seeing action be taken." And one model that Kate and I were discussing was this idea of the long-term benefit trust, which all of the sudden is hotter than Kool-Aid here in Austin. And so I know we wanna explore that further, and Kate, I know you had some particular ideas to dive into with respect to LTBTs.

Kate Klonick: Yeah. So, Kenji, I kind of wanna supposition something and just see what you think of it, 'cause as a con law professor, as someone who's been on the Oversight Board, but there's this huge difference in my mind between platform governance and things like the Oversight Board, and even the metaphor that permeated the setup of the Oversight Board, which was the C- which was like these, the Supreme Court. And the longterm benefit trust, which is this idea that Anthropic set up in like late 2023 that was essentially, you know, it was a, I'm not gonna get into the details of it, but there was a stage, There were different types of people that were appointed from different types of places to be on their board with the idea that you would kind of like hold them to this public benefit type of idea, and it would be at the corporate governance level. So it would be at this level of where the money was being made, and so they would have kind of this like very kind of real power.

Now, what's really fascinating to me is that AI is kind of gravitating to this language around constitutions and corporate governance, whereas like Meta gravitated to kind of this very rights-based framework, and they're slightly different. And one is because, like, you know, you're really talking about like the users and user voice, and like that type of idea in a platform governance set- setting like the one that the Oversight Board does. But the longterm benefit trusts, the idea of governance at the A level isn't really about like your right to speak on a platform, right? Or your right to be heard or your right to be listened to. It's really truly kind of about the safety of humanity. It's about kind of people having, you know, this is, it's a much kind of more holistic, existential question.

And so it's kind of interesting to me, and I wonder if you have thought much about the fact that like there is this, there is this, you know, uptake in this idea that's more corporate governance based than it is kind of traditionally, you know, rights-based or things like that, and if there's room for both in the AI context, or you think that it's correctly based in the corporate governance the corporate governance kind of stature.

Kenji Yoshino: Yeah. Again, thank you for that. So I think there's definitely room for both, and in fact, the pieces of the oversight board model that I think are most transferable to the AI context are the three governance principles that I was talking to Kevin about. So I really view those points not to be about individual rights, but much more to be about what structure of governance would we need to assure that we get good outcomes. So that's sort of thought number two number one.

Thought number two is, I really don't like the sponginess of individual rights or ethics or the like. And the one thing that I have really loved about the Oversight Board is that we hang our hat on international human rights law. So I would like the rights piece of this to not just be as Kevin was saying, like a cloistered group of philosophers talk, I mean, it's, it's wonderful. You know, I have a, a STEM kid and a humanities kid, so I'm like delighted for my humanities kid that philosophy is like the most employable major now apparently in, in the world. But I'm kidding, of course. But I don't think it's a good idea for people to leave it at the level of ethics because it's too subjective and it's too insular.

Whereas when we look at international human rights law that is a diverse global body of law that we have iteratively constructed over decades and feels much more stable and much more like a common language of rights that we could all have. I think both are gonna be critically necessary, and if we're tilted too much over one way, then we're in trouble.

And I can't resist, I mean, the con law guy in me says, you know, agrees with my colleague Rick Hills and, and many others that and I'm sure, well, I suspect you tell me, Kate, that you would agree with this too, that, you know, even though we divide constitutional law into structure and rights, those two pieces are so intricated with each other that they're really impossible to talk about in a sophisticated way without, it's impossible to talk about one in a sophisticated way without talking about the other, right? So like I always say to my, my students like, "Well, then who protects your rights?" You know, because if it's the courts, it's one thing. If it's Congress, then we have a separate provision of the Constitution, you know, Section 5 of the 14th Amendment that guarantees equal protection rights and et cetera, et cetera. So the structural pieces and the rights pieces are always already kind of intertwined with each other in a way that makes me a, a little bit, you know, nervous or allergic to conversations that say, "Is this a rights problem or is this a structural problem?" It's always gonna be both.

Kate Klonick: Yeah, so I, I kind of wanna just follow up on that really quickly, which is such an, which I, I would put it slightly differently. I would be like, I would kind of be like it's an enforcement question. It's a question of how exactly, whether structural or rights-based the structure might dictate who is doing the enforcement or how it kind of takes place, but the actual enforcement of these types of things is a huge part of this. And one of the interesting things is, like we were talking about at the very head of the show, that Kevin asked, set up these questions about the binding nature and this question of how impactful these kinds of recommendations are. But one of the things that I'm super interested in is the, of course, the corporate governance structure of the long-term benefit trust.

You know, we watched with the, the blip, basically, as they call it, the moment when Sam Altman la- was fired and then was brought back from like the trust and like the entire board turned over and there was all of these types of things, these questions of kind of like how really can a public boar- benefit corporation be enforced?

And if this, if you, this was exactly the moment that like the trustees thought that he was deferring from the public benefit, and so they fired him. And so then there's a riot of like kind of the, the employees, and then he comes back. I, I really do wonder, like, what is so much better about the long-term benefit trust in your mind than just any type of public benefit corporation or any type of corporate board that has to in some way kind of balance shareholder rights in some type of capacity, is, you know, even if it's just kind of written on paper.

Kenji Yoshino: Yeah. I, I, I mean, I, I wanna be cautious here because I really respect Anthropic, but I, I have the same hesitation, which is we were talking about sort of open source code. I would really like open source governance . And the fact that the public benefit trust documents are not publicly available is really alarming to me, right? So the Oversight Board does, you know, publish all of its charters and, you know, underlying documents for anyone to see. And so the fact that this public benefit trust is being governed by principles that we ourselves are not privy to kind of violates my transparency norm, and so it, it puts up some of my antennas.

All that said, you know, Kate, I mean, maybe I'll turn this back to, to both of you. I mean, to me, the question is always not “is the Oversight Board perfect?” Or is this “Benefit Trust perfect?” Because of course we're not, and they're not, and no one ever will be. It's really a “compared to what” question, and I want, I wanna take that seriously. I don't wanna say, "And that disposes of everything," because of course it's better that the Oversight Board exists than that it not. Because one could say, "Oh, well, the fact that the Oversight Board exists gives Meta more credibility than it should, and it protects it from regulation." I've seen no evidence of that. I've seen no evidence that the existence of the Oversight Board has performed any kind of shi- heat shield function for Meta. I've seen no evidence that it has slowed, you know, you know, the DSA from going online or other attempts to regulate social media. I don't think the existence of individual regulation is going to prevent these lawmakers from stepping in and saying there's certain problems that only we have 'cause they're collective action race to the bottom type problems that only we have the authority to stamp out. And so it's just an institutional competence issue.

But, you know, at the risk of oversimplifying, right, I do think that if the question is “compared to what?” Then the issue would be, is the Oversight Board better than regulation. And then the million dollar question is, are we preempting or even slowing down regulation? And the answer to me seems to be to be an emphatic no. So that seems like a false choice. And so then it really does become is the Oversight Board better for existing and dealing with the cases? And, you know, we only have like a keyhole into Meta's operations, but that is way more than anyone else in the world outside of Meta has, right? So if we're using that in order to, you know, shove our human rights values through there to mix some metaphors how can that be a bad thing, right, compared to it not existing at all?

Kevin Frazier: And that's why I think your metaphor for open source governance is so compelling and is particularly true with respect to AI because it's not only speech, but it's the economy, and it's our culture, and it's our politics, and it's the concentration of power, and it's all of these things. And so a sort of governance in layers has to be the solution where we're thinking about novel checks and balances.

But one other thing I wanted to call out with respect to the Long-Term Benefit Trust and similar models that I would just encourage listeners to dive into whenever these things get released is, number one, transparency, as you pointed out. If labs are complaining that a voluntary framework by the administration is clandestine and being kept secret from public observation, transparency on their part is welcome, and they are free to share as much information as they'd like. My radical own take is that I think labs should publish their org charts and show their decision workflows before reaching major decisions. That to me seems like basic common sense transparency that is necessary in this context.

But one other thing that you and Ronaldo point out is that the trustees on the, quote, "long term," and I'm doing scare quotes for those who aren't listening or watching, the Long-Term Benefit Trust has trustees that only serve one-year terms. And if you're only serving a one-year term, that is a short-term incentive structure, right, where you're just thinking about decisions you can make here and now versus incentivizing that longer-term perspective.

And something that I wanna touch base with you as well is how we're actually going to incorporate future impacts of AI on communities that aren't even using AI yet, right? You can go and you can look at where is AI actually adopted, and there are whole communities the world over that have incredible numbers of young people who are going to be the majority of AI users in the future, and yet our governing institutions rarely look like them. But when it comes to making sure that we are scrutinizing these self-governance mechanisms, I think we have to really go back to those incentive structures. And so long as those incentive structures don't inline with what we expect that governing institution to accomplish, I think it's some degree of governance theater in some aspects.

Kenji Yoshino: Yeah. Again, you know, hard agree with that, you know. Except for one, one caveat which actually stunned me, which is that I looked at a university study that said, like, 30% of undergraduates don't use AI, and it was for environmental reasons largely of saying, you know, every time you use it, you kill a tree, and so we're just gonna be socially responsible, which made me think, wow. So this idea that, you know, younger generations will inevitably have greater uptake of AI, I think is largely true, right? But I just wanted to so that I don't seem like too much of like a, a softball guest here, you know, enter that, that point of, of caution or disagreement.

But like I otherwise I, I entirely agree with you, which is to say, you know, these governance issues are they're really core issues. And so if you have individuals who aren't affected, are affected by AI but aren't at the table, right? You know, the, I just recently heard an AI speaker, you know, come in and say, you know, "If you're not at the table, you're on the menu, so watch out," right? And so I do think that there's a lot of that.

I think one thing I would throw back though is to say that, you know, I think we're having a kind of, and putting on my law professor hat now, I think we're having the law of the horse debate all over again. So Frank Easterbrook famously said about the internet, not about AI, that we don't need a law of the internet any more than we need a “law of the horse.” You both, of course, know this. And he said, you know, if it's, you know, if you're selling a horse, then the law of contract will govern it. If you've been trampled by a horse, then, you know, the law of tort will cover it. And so, you know, basically this new thing can actually be absorbed by different principles.

And I think that's actually a, a, a kind of heartening point, right? Which is to say I work in civil rights law, so this is a place where I'm most knowledgeable where I look at the huge effects of inequality that AI could have because every Fortune five hundred company that I know of is using AI in hiring, promotion, recruitment, advancement, you know, et cetera, et cetera, like Eightfold or Workday or, or what have you.

But disparate impact law is actually very good at screening this kind of stuff out. So you could have these case decisions like the Walters case that say like AI doesn't have any kind of intent, at least in the case of defamation, right? That's state court in Georgia so, you know, take that with a grain of salt. But disparate impact doesn't require any kind of intent. And so if you have a Workday or Eightfold and it disproportionately weeds out people with disabilities 'cause, say, you don't make eye contact enough in your video AI interview or age, right? Because it has these, you know, correlations that it's developed through its training data about age and competency or race or gender or the like, then you're gonna get nailed on disparate impact grounds, right?

So, the, the one area, so, and, and there's a case as, again, I'm sure many of your listeners know, called Mobley in, v. Workday in the Northern District of California, which everybody's watching very, very carefully to see whether or not the law, law of the horse premise of we don't need a law of the horse because we have disparate impact will play out. So that's one thing that could protect us.

But going back to your side of the ledger, Kevin, I was at a faculty workshop, and I'll, I'll just name her because I think she would stand by this and would be, and should be proud of this comment. But Cathy Strandberg, who works on privacy issues, said the one place where I, I think we do kind of need a law of the horse is privacy, right? Because the extant kind of privacy laws just didn't understand the kind of scale and the intrusiveness, right, that AI could engage in by, you know, scraping our data and, you know, doing the mosaic thing of putting together a portrait of us.

And one of the most fascinating lawsuits in this domain is the FICRA lawsuit, which I think is less well understood than the discrimination Title VII-type lawsuits, where Outten & Gold and Jenny Yang, the former EEOC commissioner, is bringing this, and she's saying, "If you're scraping my data to create a profile as to whether I'm employable, and you don't give me a chance to rebut that when you've collected it from things that are not easily and publicly available, like, say, my credit score, then the Fair Credit Reporting Act, you know, structures apply, and so therefore you could be liable on that basis as well."

So, you know, that's an interesting, I'm watching that case really closely as well to see whether or not it can secure us some gains at privacy. But I'm with Strandberg, Professor Strandberg, intuitively in thinking that privacy might be the one domain that's different from my neck of the woods, which is equality.

Kevin Frazier: Yeah, and I think what is really going to be a question for us to answer on this law of the horse inquiry is the extent to which AI's technical elements s- expose more fields than just privacy as things that require updates. Because one of my grave concerns right now is you can go to SF, and you can hear people using terms, talking about “spinning up agents,” raising their hand that they have one hundred agents working on their behalf, and using lingo that the rest of the country would not understand.

And what I worry is that too much governance being concentrated in one community or one geographic community is going to leave everyone else who used to be able to enforce the law of the horse, for example, stranded. And so if we don't have that sort of epistemic understanding being distributed enough across folks who can hold truth to power and contest decisions that are being made, well, then that's really where you run into some governance challenges.

And so I know, Kenji, you're studying these issues. Kate, you're studying these issues. I've got to get back to writing and have some students to go teach about this law of the horse and AI and all that jazz. So Kenji, unfortunately, we're gonna have to leave it there, but this was so much fun having you on. Thank you for joining and walking through your excellent Tech Policy Press article.

Kenji Yoshino: Oh, it was a joy to be with you. Thank you so much for having me.

[Outro]

Kevin Frazier: Scaling Laws is a joint production of Lawfare and the University of Texas School of Law. You can get an ad-free version of this and other Lawfare podcasts by becoming a material subscriber at our website, lawfaremedia.org/support. You'll also get access to special events and other content available only to our supporters. Please rate and review us wherever you get your podcasts. Check out our written work at lawfaremedia.org. You can also follow us on X and Bluesky. This podcast was edited by Noam Osband of Goat Rodeo. Our music is from Alibi. As always, thanks for listening.