← The Lawfare Podcast
The Lawfare Podcast
Scaling Laws: Ethan Mollick: Navigating the Uncertainty of AI Development
01:06:21
Loading summary
Transcript105 lines
- [00:01]
Lawfare Host
The following podcast contains advertising to access an ad free version of the Lawfare Podcast. Become a material supporter of lawfare@patreon.com lawfare that's patreon.com Lawfair also check out Lawfare's other podcast offerings, Rational Security Chatter, Lawfare, no Bull and the Aftermath.
- [00:32]
Ethan Mollock
If you.
- [00:32]
Lawfare Host
Work as a manufacturing facilities engineer, installing a new piece of equipment can be as complex as the machinery itself. From prep work to alignment and testing, it's your team's job to put it all together. That's why it's good to have Grainger on your side. With industrial grade products and next day delivery, Grainger helps ensure you have everything you need close at hand through every step of the installation. Call 1-800-GRAINGER Click grainger.com or just stop by Grainger for the ones who get.
- [01:00]
Lawfare Announcer
It done who says luxury has to come with a steep price tag? Quince brings you hotel quality sheets and linens without the markup. From breathable bamboo bedding to spa worthy bath towels, Quint makes every corner of your home look and feel luxurious. And by partnering directly with top artisans and ethical factories, Quint delivers premium quality at half the cost of similar brands. Upgrade home for less at quint. Go to quint.com homeupgrade for free shipping and 365 day returns. Quint.com homeupgrade it's the LawFair podcast I'm.
- [01:42]
Kevin Fraser
Kevin Fraser, the AI innovation and law Fellow at the University of Texas School of Law and a Senior Editor at lawfare. Today we're bringing you something a little different. It's an episode from our new podcast series, Scaling Laws. Scaling Laws is a creation of lawfare and Texas Law. It has a pretty simple aim, but a huge mission. We cover the most important AI and law policy questions that are top of mind for everyone from Sam Altman to senators on the Hill to folks like you. We dive deep into the weeds of new laws, various proposals, and what the labs are up to to make sure you're up to date on on the rules and regulations, standards and ideas that are shaping the future of this pivotal technology. If that sounds like something you're going to be interested in, and our hunches, it is. You can find Scaling Laws wherever you subscribe to podcasts. You can also follow us on X and BlueSky. Thank you.
- [02:43]
Alan
When the AI overlords take over, what are you most excited about?
- [02:47]
Kevin Fraser
It's not crazy, it's just smart.
- [02:50]
Alan
This year, in the first six months, there have been something like a thousand laws.
- [02:53]
Kevin Fraser
Who's actually building the scaffolding around how it's going to work, how everyday folks are going to use it.
- [02:59]
Alan
AI only works if society lets it work.
- [03:01]
Kevin Fraser
There are so many questions have to.
- [03:03]
Alan
Be figured out and nobody came to my bonus class.
- [03:06]
Kevin Fraser
Let's enforce the rules of the road. You're listening to Scaling Laws from the University of Texas School of Law and Lawfare, exploring the intersection of AI innovation and the law.
- [03:20]
Alan
Ethan Mollock, welcome to Scaling Laws. Thanks for being here.
- [03:23]
Ethan Mollock
Thanks for having me.
- [03:24]
Alan
So I want to start by getting your sense of where we are in the AI growth curve. So for context for our listeners, we're recording this on May 23rd and this week has seen a lot of high profile AI news like it seems like every week in the last two years. So just to give some examples, anthropic just dropped Claude 4, its biggest and best model. Google released, as far as I can tell, dozens of pretty impressive models and applications. At its annual IO developer conference. OpenAI announced that it's spending I think $6 billion in equity to buy Jony I've that was the former Apple designer, his startup I guess, though it's unclear whether they have a product or even an idea. But Jony I've is impressive. So I guess he and Sam Alton want to build some AI hardware. Something something. So I think it's fair to say that we're still on at least some exponential part of the curve both as far as models and applications go. And I'm curious if you agree with that and maybe more importantly whether you think that's going to continue or we're about to hit some I don't want to call them walls, but I don't want to go all Gary Marcus on us in the first two minutes. Right? But if hit some slowdowns.
- [04:37]
Ethan Mollock
So I feel like pedantic on a Scaling Law podcast saying this, but I think that the product releases are probably less interesting than what they're telling us about the scaling loss and especially the reason why we're kind of in the place we are is it's now there's now scaling laws and not law. So these laws just to make it clear for everyone, not empirical proofs, they're observations. But the first scaling law is that the bigger your model is, the smarter it is. Right? So until very now hyperscaling this idea of building larger and larger data centers, training your data, your AI and larger and larger amounts of data that has been the driving force for AI development and why they've been getting better. But there's Sort of a cyclical nature to that, because you need to build a new data center and get a lot of data together. And we're seeing the new releases of models are much larger than the last ones. But then OpenAI showed us there was a second scaling law, and there's probably more as well, which is the longer your model thinks about a problem, the smarter it is. So there's now two ways to make your model smarter. You can either spend a lot more computing power training it, or spend a lot more computer power where it thinks through the answer to each problem. So to kind of go back to the fundamental kind of question, is this still, you know, growing? Well, the first scaling law, I mean, the scaling laws also say, by the way, you need like 10 times as much compute to get to 2010 or 20% improvement. So they're always sort of a little bit, you know, they decay. Right. You need more and more effort to move the needle by the same amount. And so I think there was a lot of worry that maybe the first scaling law would run out. Now that we have other ways of making models better, it feels like there's a lot of room left for improvement. So I have a very different feeling. I think Silicon Valley does today versus how they would have felt, you know, if you asked them six months ago, before we knew there were other ways to scale models. So this is a very long, professory way of saying that I don't see a wall right now, and it's more that the labs don't see a wall. I'm just, you know, a conduit for what I'm seeing in the world. There's. And I don't see a reason we should suspect that AI development is going to cease sometime soon.
- [06:29]
Alan
Let's dig into the scaling laws question and the sort of specific ones we see. Right. Because I think it's actually worth spending some time on. So you mentioned the first scaling law, which is just model size. Then the second scaling law is we have test time compute or test time inference. You generate a bunch of thinking tokens, and the longer you do that, the better. The answer is, I've heard of another scaling law that, you know, basically parallel search, where basically you have the model generate a bunch of sets of output and then you kind of pick from them. I think that is part of what is happening with Google's new DeepThink model. Are there other scaling laws on the horizon? Because it is true that these labs seem very optimistic. If you talk to especially Dario Amade or Demis Sabis, they seem very optimistic that they're going to get to AGI or that they know the path to get to AGI. I think that was. Sam Alton made that sort of statement, famously a few months ago. Are they doing that on the basis of the current set of scaling laws and they just have to push them forward, or do they need a couple. Do they need, like, scaling laws 4 through 6 to figure out? And if so, do they have any sense of what those might be?
- [07:39]
Ethan Mollock
We're now laying our own uncertainty on top of uncertainty, right? Because, like, what's AGI? What's the end goal here? Is there some sort of, you know, reinforcing mechanism or feedback loop? I mean, there's a lot of questions that start to come up pretty quickly in this kind of. In this kind of situation. And so I think you're right. I mean, I think that there's arguments that the third scaling law is about parallelism. There's probably a fifth and seventh. Like, these are just multipliers at some point. Stop speaking scaling laws. It's like there's a bunch of levers we can move to make AI better. And when I talk about algorithmic improvement, right, there's a whole bunch of, like, papers lying around saying that, well, rather than just using LLMs, you can do this or you can combine LLMs with this clever mathematical technique and you end up with something better. Or tool use is another angle that people are pursuing. If you have AI use tools, does that change things? You know, even just adding more memory, does that do things? There's a lot we don't know. There's a lot of unexplored paths. And part of what gives me some confidence that the AI labs believe what they're saying is they don't seem to be exploring those. Right. There's a whole bunch of things they could do to add efficiency, to make these models move fast. They're not doing it. I think they feel they have a straight shot. Are they right or wrong? I don't know. Right. I think it's unlikely that we've hit a. That there's a dead end, a wall into all of AI in the near future based on the number of options available. And as you said, they're all confident. They're confident privately, they're confident publicly. You know, I think one of the most dangerous things you could see is dismissing this stuff as marketing. They might be wrong, but I don't think that they're just making this up. And to be impressive that they think they're going to get to AGI, they really believe it, for better or for.
- [09:13]
Kevin Fraser
Worse, and to move us on a slightly different track because there is this temptation, I think, among folks who talk about AI. It's kind of like when you meet a long distance runner, they always want to tell you about the next long race they're going to run and how they're going to do a 50 miler next. And then they've got this sweet 100k lined up and you're like, great, whatever, that's cool.
- [09:35]
Alan
This is why I try to not. This is why I try to not talk to long distance runners.
- [09:39]
Kevin Fraser
This is, you know, why. Why you're going to get invested in ski join. This is the winter, Alan, that I'm counting on you becoming the ski join champion of Minnesota. But we have a similar inclination, I think, with AGI, where we're constantly just looking at the finish line. And something that I appreciate, something that I appreciate about your writing, your scholarship, your public statements, is saying, whoa, if we just paused right now and stopped all AI advances and just used the tools that are available today, it would still take us decades to meaningfully implement all of the AI advances we've already achieved. So I would love your take on whether this debate and conversation about when we're going to achieve AGI, what AGI even means, what's the difference between ASI and AGI and all these just kind of semantics. Is that useful? Is it something that we should spend our time thinking about? Or is this more practical focus of just implementing today's existing AI really where we should be spending more of our attention and policy focus?
- [10:50]
Ethan Mollock
I mean, the most academic answer of all is it depends which is what I kind of feel like. I think the issue here is, you're exactly right. One thing that gets obscured, you know, Tyler Cohen said, oh, three is AGI. I mean, and I wrote a whole piece on this, like, for some jagged frontier version of AGI where it's. Where it's clearly superhuman across many characteristics and not others. Definitely. Right? Like, we already know that you'd probably rather like, if you're doing differential diagnosis, you 100% want to, as a doctor, you 100% want to be using AI to do this, right? Like, if you're not, like, it's already better than humans across most of that field, but not across other parts of medicine, Right? So part of this is figuring how it works, building the systems around it. And I really do believe that the systems now are so much more disruptive than people believe. Right. There is if you talk To a product manager. They believe product management has to change. You talk to a marketer and you could see already marketing is going to have to change. That's year long, years long projects to figure out how that's all going to work, right? Law is clearly going to change. Like this does things, you know, talking to lawyers, I am not one, but you know, who've used deep research before it's even connected to good law libraries, right. And it's like this is producing, you know, 40 hours of associate work pretty quickly and takes an hour to check. What do we do with a billing model on that front? So disruption is already baked into the system right now I would just say we could focus on that alone because what ends up happening is a conversation about what the future holds. You know, the hype question. But the truth is that the labs keep showing amazing things and, and you know, I'm surprised by some of these releases and I think they're surprised by what they could do because they don't even know, right? So I mean there are limited testing of all of these things too. So I think that we do have to pay attention to where things are going. I think the question I never weigh in on because I don't have no insight into it at all, is, is like, is a super intelligence possible machine that's generally better than humans at everything and kind of wakes up one day and decides what it's going to do with us, right? And that is the big fear everybody has. That's the big policy question, the nuclear weapons style question to deal with. And I, I just have no insight onto that at all, right? Other than there's enough smart people worried about this that we should probably be worried about it. Right? But what ASI means, what AGI means, what that means for employment, I mean we're figuring this out as we go and I think in some ways it's already baked in.
- [13:04]
Alan
I want to go back to something you said sort of earlier in the conversation where you said, look, talking about the models is a little more interesting than talking about product rollout. And I generally tend to agree, like I'm not myself, like a Silicon Valley product market fit kind of person. I more interested in the models, I guess myself. But it does strike me that I wonder if, and I'm curious what you think about this, whether it's actually the products that are currently the bottleneck in terms of seeing AI diffused throughout society.
- [13:31]
Ethan Mollock
Right.
- [13:31]
Alan
And just let me have a trivial example. I am positive, right, that Gemini 2.5 and probably Gemini 1.5, to be perfectly honest, is probably smart enough to help me with a lot of annoying email tasks in the sense of just like, help me respond to this. Can you just categorize some stuff? And yet we don't have that because Google is, I don't know, bad at including this stuff usefully in Gmail. So this strikes me as an example of where it's not the sort of sheer IQ or capabilities of the model is the bottleneck, but rather the integration, kind of last mile integration to make it easier for, for consumers. And I suspect that for a lot of knowledge work, right? And you know, you talked about law, obviously that's what Kevin and I used to do before we became a podcasters or kind of the business consulting stuff that you have a lot of experience with. That seems to me to be the bottleneck to see real AI adoption. So I'm curious what you, what you think about that.
- [14:34]
Ethan Mollock
I mean, yeah, I mean, the prioritization of this stuff is terrible. I mean, it's a chatbot. A chatbot should not be. The way you do real work is like, you know, is like there's a reason we don't use discord as our universal format for all, you know, for getting all work done. Like, it's, it's an insane tool, right? It's built for individuals to use. It's supposed to, it's a friendly, helpful assistant and yet we're having it write essays for us and like, and people are just bolting on, you know, like things into this, into the system. We at Wharton, at Journey of AR Lab at Penn, I run like we've been building tools on the API that like, let it listen into conversations and give you feedback. I mean, the capability set is very, very large. The UX is holding us back, right? The experience is holding us back. Which is why again, I'm so confident that these systems could do a lot. I think they could do email, I think if they were actually integrated into spreadsheets in a useful way. And we know how to do that, but no one's bothered even just the reward functions of these systems are set up really around math and coding and not around, you know, writing or engagement or the many other things that you could start to see them do. So I think we're held back by this. I mean, I think large part though, the underlying belief of the labs is intelligence solves all these problems. If we make the AI smart enough, who cares, right? Why would we waste any time building an interface when we could spend all of that compute Time building a better model. And I think.
- [15:51]
Alan
And then, and then the model will build the interface.
- [15:53]
Ethan Mollock
You're just going to ask clock or just. It doesn't matter anymore. It just will know what you want, right? It's heard this podcast that we get off of it and it's like rearranged my emails and dealt with the whole situation, right? Like the implementation is like you could do science. Why, why do engineering is sort of where, where I think a lot of them are. But there is this genuine belief that smarter models solve everything. And you know, there is a little bit of truth to that or at least some truth we already know, which is the companies like Microsoft that early on kind of committed to the ability level of models. So they built Copilot, which I'm sure a lot of your listeners have used in one way or another, was built around the idea that like GPT4 is a limited system and its capabilities are limited. So they've kind of, they've isolated the AI and only does very narrow things. And now Microsoft's kind of stuck in this Copilot model when the AI is sort of floating everywhere now and can do a lot more work than it did. So and I see this in companies all the time, right? They, they committed to building a AI system around the limits of a llama, you know, three model or llama two model back in the day and they built all of this cool infrastructure to support the fact that it wasn't very good at this stuff. And, and now you could just ask, you know, Claude or ChatGPT and it just does the whole thing now. So there is this kind of weight equation problem too that I'm sympathetic for the labs about. And what I worry about, right, of course, is that you decide to create a product today as a startup and then OpenAI or Gemini crushes you later. I mean that's just like the number of times my students I teach entrepreneurship, the number of times my students have launched ideas that help you do virtual clothes try ons and that those have received funding is quite large, right? Not just by students, but lots of people. And then Gemini just sort of Google just sort of announced, oh by the way, we have virtual clothes try on is just a thing that the system does now, you know, congratulations. And like, you know, the scream of dying startups echoes through the void. So I do think that you're absolutely right. The interfaces hold us back. I don't know if they're not making a good bet with like let's wait till the smartness plateaus before we start building the interfaces and building off of.
- [17:49]
Kevin Fraser
That idea of we're just operating in a space of so much uncertainty as to what the models may be capable of in three months, in six months and nine months, you can have students, you can have professionals, you can have startups saying, hey, actually, what's the point of learning prompt engineering? Right? For example, everyone was saying, you got to learn prompt engineering. It's the most important skill you're ever going to learn. And now everyone's saying, don't learn prompt engineering. It's not useful. All these models are too sophisticated. And when you even go and talk now to folks like Daniel Cocatello, who we had on the Lawfare podcast, and you read the AI 2027 report, and now you're talking about AI super coders, training models, and then having this just recursive loop where they're getting more and more advanced as a result of having these AI supercoders. How do we operate in a space where we may see levels of AI advancement that are just simply beyond our ability to comprehend right now? How much weight do you put on these sorts of forecasts that are saying we don't even realize how quickly things could shift in the next six months or nine months?
- [19:03]
Ethan Mollock
I mean, part of the problem is like, they're not really forecasts, right? Because we can't forecast anything. So there are science fiction stories that might be. Right, Right. They're provocative provocations to get you to think about these issues. And it's just interesting to hear 2027 referred to as literally, our vision ends right now. Right? Like the labs have another six months that we don't have six to eight months. Right. But they don't have six, eight months on what the implications are. They just have six to eight months of model stuff. They don't know what these things are good or bad for either. We don't know. Like, I just. It's a hard thing to deal with because, you know, you started with saying, what do you know about prompt engineering? And what, what are the results of this? How do we start thinking about what people should learn or do? And that requires taking a bet on where things are going and we instantly run to a wall there, which is what? Let's say we have super intelligent machines. What does the world look like? I have trouble even imagining the world where we have 03 Gemini 2.5 Claude 4 level models that we actually exploit the way we're supposed to. Right. Where they're actually, you have whatever the magical Jony. I've device is because it's possible to imagine a bunch of devices that just like at this point, the AI could listen to everything you're doing, process it 40 different ways, take action in the world to help you. There's a failure of imagination about what these systems can do. So we're trying to do this ambidextrous thing of thinking about the far future, which might be four years away, and also trying to adapt to today. And we have systems that need to adapt like people need training today. What should I learn to be good in a world of AI? And the answer tends to be adaptability, basic management skills, more than a technical set of skills. So we're dealing with second order change, right? Not just the rate of change is changing, not just things are changing, but the rate of change is changing. I don't know. It's a hard question to answer.
- [20:44]
Alan
Do you, do you just. On the last point before we move on, do you find that answer compelling in the sense of. I've listened to so many podcasts and read so many articles about people making this point of. The key is soft skills and critical thinking and emotional intelligence. And in some sense that's true. It also feels like a giant kind of tautological cop out. I can't tell if when people say that, they really mean it or they feel like they have to say something, but they actually don't know what to say. Because, my God, if AI really progresses at the speed that we're talking about, who knows what skills will save us? But you got to say something to be polite. When someone asks you a question, you can't just say, I don't know, pray to the machine overlords that they'll tell us.
- [21:23]
Ethan Mollock
Well, I mean, there's two things here. One is if we do get to the level where we have the machine overlords, who cares about our skill set, right? Because. Because it's like, you know, we'll do whatever. I mean, humans all adjust, we'll find meaning and whatever. Right? But I think that the more profound immediate question is, you know, is what skills we're telling people to get. Assuming a world that stays somewhat normal, but AI keeps getting better. Right. And is it great at everything? And I do think there are answers and I think that your list of things mixes together answers I would give and answers I wouldn't give. Right? So I don't think emotional intelligence, a critical thinking. I think those are. Those are attempts to drive bright line skills where like, oh, humans can do this and AI can't, or to Deal with a world of AI bs. We need critical thinking or only humans are emotionally intelligent. Neither of these things are true. Right. The AI is showing emotional intelligence. It's showing creativity. Our attempts to do bright lines of machines can't do this because they're the human skills is a really bad way to approach things. On the other hand, I think that there is value to be said in prompt crafting is less important than management. Turns out AIs respond really well to management ability. If you can articulate a problem, be able to be able to give feedback when that problem works or doesn't work well, to be able to, you know, to envision what steps you want to see. That stuff is all actually very useful. And it turns out experience is also very useful. We have a paper showing that, that junior consultants did much worse using AI or giving recommendations than senior consultants because they get a document back and it looks good and they don't have the deep knowledge to understand what's good or bad about it. So I don't. I think it's important to both say there is a little palliative, kind of like, oh, humans are still valuable. We still need your insights that I don't think we have as much evidence for. But I think, at least in terms of using AI today, there are things that make you better at using AI. We haven't articulated all of them yet, and that we don't want to dismiss those by just saying a lot of things people say about what you need to do in order to survive in a world of AI are kind of fake.
- [23:16]
Kevin Fraser
And I love this point and this whole conversation because one of my favorite things, when folks say leaning into the sort of emotional intelligence thing and that AI doesn't have the same empathetic capacities, number one, has anyone gone to a doctor's office and experienced a doctor with absolutely zero bedside manner? I'm raising my hand right now because, yes, that's been the case again and again. And number two, as you point out extensively, if you just use the tools, if you actually go spend 10 hours using these tools as you recommend and just ask, hey, be nice to me. Guess what? The model's gonna be nicer to you. It is capable of taking those emotions on and responding to them or, you know, be an A hole, also fully capable of being an A hole to you, it can take on those emotions.
- [24:05]
Alan
That's Grok's entire gimmick. That's. That's its product differentiator.
- [24:10]
Kevin Fraser
I just say act more like Alan, and all of a sudden it's ready to go. I'm just kidding, buddy. I'm kidding, buddy. So thinking about some of this systemic change leads me to one of my favorite examples, which is education itself, where we have some really antiquated systems in education. My favorite is why the heck do we still have age based grade cohorts? Theoretically, with AI you can now track instantaneously how each student is progressing in a specific subject matter, change their curriculum, change their courses and allow them to progress in that way. They can still do resource, excuse me, they can still do recess with their age based colleagues, but there's no real rationale as much for making their actual learning grounded in whether they're along with their peers or part of the same age cohort. That was a poor way of Tina, your use of AI in education because there are just such rigid thinking among academics, among educators about how and when to use AI in certain contexts, specifically in the classroom. Can you walk us through your own approach to how you instruct your students to use AI and how you have perhaps drawn some red lines with respect to your use of AI in the education setting, for example in grading.
- [25:43]
Ethan Mollock
So there's a lot there and so let me try and answer it in a couple segments. First off, we at the current state of AI, going back to the interfaces question, it's not built well for education and we know that also because it doesn't match what we know about pedagogy. So if you ask the AI to help you with something, it often gives you the answer that's not tutoring. In tutoring, as we know, you want to force people to try and answer the question themselves, guide them to an answer, adapt to what they're doing. So you know, if you just, we, we already have subcontrolled experiments that if you just ask AI to get, you know, if you just give AI to students, they ask for answers, they don't go through any intellectual effort, they think they learned something and they do worse on tests. On the other hand, if you configure AI to work like a tutor and give it people classroom support in the right kind of way, you get fairly big boost to learning outcomes. So there is a little bit of sophistication here beyond just like everyone should use AI to boost their learning because it doesn't do that by default, just like it doesn't do a lot of other things by default that it could do. System prompts can change that. A good prompt makes a difference. Adding other tools makes a difference. So I, I and you know, and in general Also, there are topics that we don't want to teach with AI. Like there, there's a. When, when the calculator came out in the 1970s, it caused a crisis in math education. And as a result, now we probably have. The math education is probably one of the best thought out areas in education because we had to reconstruct it from the ground up. So there's things you don't do with calculators, things you do do with calculators. There's a careful plan that was, you know, that was of. Of how to think about pedagogy in that front that we haven't had to do in other fields. Right. We haven't had to do in law, which has done things the same way for a very long time. Maybe that's the right way to teach, maybe it isn't. But there's not really a theory of learning behind it. It's just that this is, this is the approach we've always used, you know, to teach law or, you know, business school. Cases are literally taken in the 1920s to make the business school look more like a law school. Like that was a professionalization move, not a pedagogical move. So our pedagogical world is a mix of things that we know work and things that have seemed to work for the last 2,000 years, like essays, even though we don't know why they work. And now they broke. Right. So there's a lot of crises happening all at once in education. We also know that there is a big role for kind of student to student contact and classroom stuff. Right. Do we want everyone to be as advanced as possible or is there reasons we want to have people in classrooms of the same age range, even if they're much smarter? Right. We've all seen people accelerated through five classes and it doesn't necessarily do well for them either. So there's a lot of complex questions there. So I'm going to put the education policy piece aside, except to say there's obviously huge benefit to AI. Educators are pushing back against that, that I think in ways that may not be productive, especially when you start looking at education in places where people don't have access to good classroom education. Big gains. Okay? So that aside, AI in my classroom, when AI first came out, when ChatGPT first came out, when 3.5, I kind of went viral first for my syllabus which said, you can use AI, but you're responsible for the errors. That was great. For the three months to ChatGPT was using 3.5 and then as soon as it switched to 4, it became as good as my very, very smart students, when they're not putting full effort into something which I don't go back, revise.
- [28:47]
Kevin Fraser
The syllabus and just lead to a complete. Did you go back and revise the syllabus?
- [28:52]
Ethan Mollock
Absolutely.
- [28:53]
Kevin Fraser
Complete riot.
- [28:54]
Ethan Mollock
Yes. Completely changed it. And now my classes are 100% AI based. There's AI simulations, you teach the AI things, there's a co create a case with the AI, you get debriefed by AI. So like I teach an entrepreneurship class, though I'm lucky I can do something like that, right? You know, instead of launching a product at the end of a six week class, you do it in the first day. Like I can do those kind of things because of the kind of class I teach. If I was trying to teach English writing, I would approach things very differently. I think AI is less useful for that than making people write and maybe getting feedback. Right. So there is differences and there's opportunities available, but we have to think hard about what we're trying to accomplish educationally. And I think it's a mix of banning AI and fully embracing it. And again, the interfaces are fighting us. Right. A chatbot is a very bad way to do teaching at scale. So I think the transformation is there. But you're seeing the same problem you're seeing in every other case, which is rushing to, yes, we should throw out all the lawyers or throw out all the teachers. It's too early for that kind of work and probably always going to be too, like always a terrible idea. And you know, there's a technical. And we've, we've been through this before. Technology alone doesn't solve these problems. But for the first time I'm hopeful that it can solve a lot of these issues because what the models can do is very impressive off the bat. So we're in an interesting set of changes and I think thinking about how we embrace this in universities while also not saying we need to throw everything out because, you know, is also important.
- [30:14]
Lawfare Host
Right now the headlines are chock full of data breaches and regulatory rollbacks, making us all vulnerable. But you can do something about it. Delete Me is here to make it easy, quick and safe to remove your personal data online. We all want an easier way to deal with data breaches and I'm telling you, you should get delete me. The fact is we're all at risk. How many times have you gotten an email or a letter saying your data has been stolen? It's unsettling but the good news is Delete Me can help. It's not just a one time service. You give them the information you want to remove from the public Internet. And Delete Me is always working for you, constantly monitoring and removing the personal information and you don't want out there. I am somebody with an active public presence. I antagonize people, I bother people. But privacy is important to me. I don't want people knowing things about my private life, about my family, about all sorts of things where they can kind of mess with me. And I keep hearing about these data breaches in the news. A recent Coinbase data breach, for example, a hacker obtained names, addresses for phone numbers and the last four digits of users Social Security numbers plus masked bank account numbers. They vacuum it up, they create a searchable profile and they sell it to whoever wants it. Thankfully, DeleteMe can help you preserve it from those sites and keep your information from falling into the wrong hands. So take control of your data and keep your private life private by signing up for Delete Me now at a special discount for our listeners. Get 20% off your Delete Me plan when you go to JoinDeleteMe.com LawFair20 and use the promo code Lawfare20 at checkout. The only way to get 20% off is to go to JoinDeleteMe.com lawfare20 and enter the promo code lawfare20 at checkout. That's JoinDeleteMe.com lawfair20 code lawfare20.
- [32:44]
Lawfare Announcer
Out here, it's not only the amazing views, but the way time stretches out a little longer, how laughter bellows louder among friends, and how the breeze hits just right at the summit. With all trails you you can discover and experience the best of nature with over 450,000 trails worldwide and navigation right at your fingertips. Find your outside with alltrails. Download the free app today and find your next outdoor adventure.
- [33:13]
Lawfare Host
This is Car Tracks with Turtle Wax. Your car says a lot about you.
- [33:18]
Ethan Mollock
So if we asked your car what.
- [33:20]
Lawfare Host
It would say about you, what would it say?
- [33:23]
Ethan Mollock
Listen, you dropped one of those tiny.
- [33:24]
Lawfare Host
Cheeseburgers under the seat like last week.
- [33:27]
Alan
And now we're both dry heaving at the stench.
- [33:30]
Ethan Mollock
Do us a favor, grab some Turtle.
- [33:31]
Lawfare Host
Wax and let's get to work. This has been Car Tracks with Turtle Wax. You are how you car ever feel.
- [33:41]
Ethan Mollock
Like you're carrying something heavy and don't know where to put it down?
- [33:44]
Alan
Or wonder what on earth you're supposed.
- [33:46]
Lawfare Announcer
To do when you just can't seem to cope? I'm Hesi Jo, a licensed therapist with.
- [33:50]
Alan
Years of experience providing individual and family therapy. And I've teamed up with Better Help to create Mind if We Talk, a.
- [33:56]
Lawfare Announcer
Podcast to demystify what therapy's really about.
- [33:59]
Ethan Mollock
In each episode, you'll hear guests talk.
- [34:01]
Alan
About struggles we all face, like living with grief or managing anger. Then we break it all down with a fellow mental health professional to give you actionable tips you can apply to your own life. Follow and listen to Mind if We Talk on Apple Podcasts, Spotify, Amazon Music, or wherever you get your podcasts. And don't forget, your happiness matters.
- [34:24]
Kevin Fraser
Well, I'm glad. Hopefully they won't throw us professors out too soon, which is a good sign. But your test for a lot of use cases of AI is often asking, what's the best available human? And comparing that AI to that best available human. And yet you've noted from a grading perspective, it's probably true, if not 100% true, that using an AI model would actually be a better way or perhaps a more accurate way or a faster way of grading your students. And yet you've drawn a red line and said, no, I'm going to insist that's something I do. Why? And why have you drawn that line? And why do you think that perhaps that's something other professions should learn from?
- [35:05]
Ethan Mollock
So, I mean, I think there's two kinds of red lines to draw. The one I drew is a great example of a system thing, right? Like, there's an expectation for my students that I grade their papers even though I can tell them that AI is better at. Is better at grading than I am, and we'll give them more insightful comments. This is what the, you know, there is a social requirement for this, right? So the system that's holding us back is a social requirement. And, and that's completely fine, right? It's. They're getting worse results from me, but that, you know, and maybe I offer them both options, but, you know, and, and then let's take one step further in this. Let's think about another thing that has to be human, which is letters of recommendation. I write letters of recommendation all the time. The sign, the goal of a letter of recommendation is to set my time on fire as an indicator that I care about somebody, right? And in 45 minutes, I write a pretty okay, ish letter, right? Like there's, you know, it takes some time off of the students will give me some guidance of what they want to write, and I get an infinitely better writer letter if, if I take from a student Their resume, the job they're applying for. And I say, I'm Ethan Malik, write a good letter recommendation. But I'm not no longer indicating anything of value to the outside world. When I ask my students what they'd rather do, all their hands go up that they'd rather have the one minute letter than the 45 minute letter because it's more likely to get them a job. Right. And that's because the signaling still suggests Ethan Malik cared enough to write a letter for me. How long that lasts becomes an open question. I also had a student for the first time last year and they allowed me to share this. Send me a prompt that I should use to write their letter of recommendation. So literally they just emailed me, please use this prompt. Feel free to that is next level. But like this is a but you could see both sides of the social coin, right? My students expectation in one area that's supposed to be human and in another area that's supposed to be human. And this is what we're really navigating at a very micro level. And then there's a really interesting cases where maybe those kind of constraints are actually helpful. Like so there's a big deal where Hollywood actors negotiated a bunch of AI rules. And my sister's a producer in Hollywood. And it's been really interesting because those rules actually allow for creativity with kind of constraints. So for example, before she does a photo shoot with, you know, with actors, the AI won't. They'll use AI to mock up potential photo shoots, but they can never use the AI to actually produce a final photo shoot. Right. She's just done a movie with Michelle Pfeiffer. They can use Michelle Pfeiffer's voice and automated for internal testing that they're doing to figure out what ADR lines might work somewhere. But they can never show an external audience the voice of Michelle Pfeiffer. ADR by AI So there's a really interesting place Also, you know, when you think about lines, there's those three versions of it. There's no one should use AI to do this. We're embracing AI and it's completely destroying the underlying meaning of the system. And we're gonna have to grapple with that. And there's this mid level thing of like we're going to create these particular bright lines that we know will actually protect people in some useful way while allowing for AI to cut down on the grunt work that would have been miserable that nobody wanted to do. And I think that that is going to be a compelling problem in law and everywhere else.
- [38:07]
Alan
So I want to stay on the education league just for one last question. And this is kind of selfish because I now have you on the podcast. So I want to just tell you how I use it, then you can tell me if I'm doing it right. So this is just. You're, you're, you're just going to give you two minutes of Ethan Malik consulting here. So. So I teach a range of courses at the law school and I teach some doctrinal courses. You know, I teach con law and constitutional law and criminal procedure, and those are still very traditional classes. I don't really use AI that much. You know, I have my lecture notes and then the quirk of law school grading is that everything's decided by one final exam at the end of the semester. They used to be take home exams, but after ChatGPT came out, we all switched to in class exams. And they're not perfect. You know, three hour timed exam is, is, you know, not perfectly correlated, but it still kind of gets the job done in the sense of figuring out who I, figuring out who has what I, what I take to be what employers care about, which is some combination of like smarts.
- [39:00]
Ethan Mollock
Right.
- [39:01]
Alan
And conscientiousness.
- [39:02]
Ethan Mollock
Right.
- [39:03]
Alan
Because that's what you kind of need. Okay. And then I teach some seminars which are more discussion based and they have response papers and you have a final paper at the end. And for there I just assume that everyone's using AI and that that kind of undermines, I think, a lot of the purpose of grading. But in law school, there's kind of a norm that people care much more about how you did in those core doctrinal classes than they really care that you, you know, and how you did in those kind of seminar classes. So for me, that's been kind of my equilibrium with AI and how I intend to use it. But it strikes me that I'm actually, despite loving AI and defining my whole career to it, it does strike me as somewhat of a luddite approach because for the core of my doctrinal classes, I still teaching them kind of in the same way that I did in the past. I mean, I tell my students, hey, you should probably use a ChatGPT or Claude to like teach yourself some of this stuff on the side. I mean, if, you know, was you're reading cases. But I'm sort of curious what you think about that. Right. As like I just told you how I use it in the context of law school. And I'm just curious what you think about that.
- [40:04]
Ethan Mollock
So I think there's two pieces to what you're doing. One is assessment. It is a completely reasonable approach to assessment. And law school, by the way, if you graded active learning as opposed to an exam, that would also be reasonable. So both, you know, in person exams and active learning are really good ways to grade. They've always been good ways to grade. Tests have always been useful, even though we hate them. And turns out testing also makes people learn stuff. If I was going to give you one piece of advice, by the way, it would be more small tests. Along the way is the way you would actually want to run this. So you get feedback as you go.
- [40:36]
Alan
The problem and this, everyone understands this and no one does this because we don't have TAs. That has always been the difference between law schools and a lot of other places. We have to do all our own grading. And grading law exams is the worst thing in the world.
- [40:47]
Kevin Fraser
This begs the question though, as Ethan pointed out, give students the option, say, do you want to opt in to AI given feedback for recurring analysis?
- [40:57]
Ethan Mollock
And there's a bunch of papers actually talking about how terrible law school is for education because it doesn't give running feedback and that's bad, right? So I mean, just like all of education, we're dealing with a whole bunch of broken systems already. Like forget AI. I would tell you, okay, then all the answers in scantron forms, even short multiple choice quizzes once a week would not only be assessment tools and help people correct what they're doing and be formative, but also they, people remember things better when they're quizzed on it, right? And you're not pulling up one person in class and doing a performance with them. So I mean, law school's already broken, right? But all education, it lags the pedagog. Like we're not doing what we should do pedagogically anyway. Okay? So let's leave that aside right now, which is like the complaints about, you know, we don't get pedagogy, we just do the thing we always did. The systems are in the way. I would, at the very minimum for your, for your large school class, I would be building tutors. I would be, you know, I wouldn't trust people to just ask ChatGPT. I have a tutor for, you know, for different, different subjects where I have the AI do stuff. I would have people do a simulated, you know, like case discussion with the AI because not everyone gets to be part of a case discussion. So I would have a case discussion available in each case. And by the way, I'd be switching, I'd be making increasing versatility to realism by having some of those be. There's a, you know, you have another. The AI acts as the attorney on the other side, the axes, the judge and grills you on stuff. I would have people apply them to a fictional setting. So like you show up and the AI is now you don't know what case it's going to play out, but it's a, you know, it's about, you know, a case on Mars and now you have to make a decision about that. So we're changing structure and approach like simulation. I'd have the AI act as a befuddled law student and have the teaching others is very helpful. So have you teach the AI how to do something right? I'd have it, you know, so, you know, prepare me for the test kinds of models that are also pulling back maybe in rag style from previous tests. And these are relatively easy to build. And by the way, if you go to the general lab at Wharton, we have a whole bunch of Creative Commons prompts that we've released that are designed for teaching and you can modify for yourself. So I think you are leaving learning on the ground, even just keeping your classes structured the exact same same. The idea that you're not offering a bunch of tools that could help people learn are interesting. Then you make them assessments. If the TA is the problem, then, you know, why not have like you could use AI grading for that and say, look, the, you know, you'll be using this for feedback along the way. And by the way, you can actually throw in every part of what the AI has given the students feedback on and you throw that into the system and you say, what themes am I missing? Where am I, where am I failing as a teacher? Because people are not getting this, what should I do in my next class to close the gap between where I think they are and where they should be? And also on the teaching side, there's some early evidence that teachers who use AI to help them prepare for teaching do better as well, especially if you integrate your interaction with the AI. So I think you're leaving a lot of value on the table. To be very honest with you, I don't think the core problem of a cloud that the assessment is as big an issue as there are ways of having people catch up on learning and it's relatively easy to do. And if you don't care about how they're interacting with AI outside class, then these are easy things. You could take advantage of.
- [44:07]
Kevin Fraser
So, as you mentioned, Alan's leaving so much value on the table. And I'm kidding again. All of the legal profession is leaving so much value on the table.
- [44:15]
Ethan Mollock
All of us are. Oh, I just want to be clear. This is not a college. I do, too. Right. We have many things to do as academics, and the dumbest thing we could do is jump on every educational fad as we go. Right. So there is a reason to have some momentum behind what we do. And it frustrates me when people say, well, college is dead. It is obviously not that we'll figure this out. Actually, I'm not worried as much about college as I'm worried about post college and what happens after you graduate. Like, that makes me much more nervous because so much of the learning is on the job learning that is being threatened by this. So I'm not worried. Like, this is not a call out to Alan, who asked my criticism. I think there's more we can do. But that's okay. We will figure this out. And we're not expected to, like, veer wildly as, like, every new pedagogical technique comes along. It is okay to keep doing things we've done for a couple hundred years as the default, but I think the crisis is not going to be in schools. I mean, it is right now because everyone's cheating, but they were always cheating. It's going to be outside of schools.
- [45:06]
Kevin Fraser
It reminds me there's a great Portlandia skit where they run through every different new parenting book that comes out, where first it's, you know, speak to your child like an adult and they hand the kid the car keys and say, you figure out your life. Good luck. Then there's, you know, the ignore it forever. And anyways, a little bit more seriously with this notion of we all have these new incredible tools at our disposal that are empirically shown to improve outcomes when used correctly and when used competently. And yet many of us, as you pointed out, just aren't doing it. And so this may not seem like a huge deal to some folks in an education setting in law school. Right. The difference between me using a great GPT versus me just teaching as normal may not be massive. So we can forgive that. But now let's move into the context of medicine where we know that there's a new tool that can spot that tumor with 99% accuracy, but a doctor says, you know what? I just don't want to use it. I just prefer to not go there. And patients should just trust that my Expertise will be enough. When are we going to transition from a policy perspective to saying a failure to use these tools is something that should be a civil violation of law or perhaps a criminal violation of law? When do we need to start saying failure to adopt these tools actually amounts to some sort of malpractice, depending on the context.
- [46:41]
Ethan Mollock
I mean, part of the problem here is we have a technology that is wildly useful across a wide variety of fields. The amount of research going on into how useful it is is vanishingly small. Right. Like I tweet about this stuff. What happens? There's a bunch of really good medical papers suggesting that AI is good at medicine. And you probably, as a doctor should be using as a second opinion, because it's pretty good at this diagnosis stuff. And certainly as a patient, you should be using second opinion. But. But we don't know if it introduces bias. We don't know if it increases over treatment rates. We don't know if it leads to better patient outcomes. So because we don't know those things and because there's no good interface, what you're telling me is that ChatGPT, HIPAA compliance, which you can but like that, we haven't built, we haven't tested this enough. And to me, the big policy tragedy is, you know, along with the set of policies that we should be thinking about preventing risk and danger and everything else. There is no giant R and D effort across governments to try and nail down these areas where we can get improvement in the world. You know, how does this work to help, you know, explain and work with regulation? How does this work to help, you know, to help simplify paperwork or other kinds of issues? How should we be used as education? How should we use this in medicine? We need to establish this stuff. I mean, in a personal sense, I think a doctor should absolutely be using ChatGPT as a backup. As a societal sense, we don't have enough evidence yet for me to be able to make that point. And we don't know what the blowback or counter effects are. And those are solvable problems with today's models. But we're counting on the academic institutions who are behind on this to do this or trusting the labs themselves. And to me, that's the big thing. On law, there are basically two or three very good empirical papers looking at AI in law that use controlled studies to do kinds of stuff, some of.
- [48:25]
Alan
Whom are written by my. I just want to plug Minnesota, some of them written by my colleague Daniel Schwartz.
- [48:29]
Ethan Mollock
I was about to say that those papers showing that O1 preview leads to 10 to 30% improvements in quality and speed. I mean, that's a huge deal for law. It feels like the entire profession is hinging on four or five more studies like that, that show and we're not doing them right. You would talk about education. It's weird that I'm like giving you random suggestions for stuff. I would like to see more tests happening and the tests are all kind of being done randomly. This feels like an emergency level for academia and it isn't happening at the level it does. So that like the answer is probably already. A lot of things should be shifting over to AI, but like to be able to establish some rule that you must use, it really requires us to do some, you know, evidence based work. And that to me is kind of the most baffling thing is like as an academic, this feels like they actually care. Like the stuff we do matters a lot right now. Right? How good is this as a constitutional law? How good is this as acting an advisor, as a. When should you use this as your attorney? Like, doesn't that seem like the most urgent problem in law and like we don't have an answer to it? What is the hallucination rate? Is that dropping? What? Like it's insane to me that we're all just like, well, someone will figure that out. That's us. That's us. We're supposed to do that.
- [49:39]
Kevin Fraser
Yeah, well, I love that charge because it gets, it attacks the narrative right off the bat that there's no utility for academics or that the scholarship perhaps won't matter as much because we can just generate it. We need to reorient our scholarship, really focus on the issues that matter most. But what I love too is just thinking about your doctors. You walk in and saying, yeah, tell me about your AI policy. How frequently are you using? It would be just a funny change in perspective.
- [50:09]
Alan
Well, I mean you already, you already have clients of law firms saying, look, I'm not paying, I'm not paying for a fifteen hundred dollars an hour associate if they're not using AI because that I'm leaving, you know, so many billable hours on, on the table. Ethan, I hate to turn this yet more into sort of therapy time with Professor Mollock, but again, I cannot resist. So I want to ask you about your thoughts on the thing that I am worried the most about with AI again in the short term. I mean, the Skynet question is something we can deal with in five years and that is the issue of cognitive deskilling. Because I feel that I can just feel that in my own work, you know, sometimes, you know, my kid didn't sleep well last night, so I'm tired and I have a bunch of stuff to do. And so maybe I ask the AI to help me prose out these three bullets into a paragraph and it's actually pretty good and only get better, but that'll make me worse at that. And obviously there's a sort of concern there for kind of individuals, but then there's a broader societal concern. I mean, you talked about one thing you worry about is not so much with education but with employment. You know, especially if AI kind of replaces that sort of first five years or first 10 years of white collar jobs, you know, you're going to have a whole generation, I think of my children, for example, a whole generation potentially of people who didn't get, how do I put it this way, who didn't get to spend their 20s adding mostly net value to an organization that nevertheless paid them a salary because they knew that they were training them to be useful going forward. So there's this kind of cognitive deskilling concern at the point, it seems to me of AI use. And then there's also a cognitive deskilling concern at the point of replacing a lot of labor. And I'm just curious how you navigate that sort of especially kind of in your own personal use of AI.
- [52:01]
Ethan Mollock
Right, so the good news, bad news, right? The good news, positive view of this is like, all right, so we lose some skills. Do we need them anyway? Right. Automation always removes some skills. My kids can't drive stick, right.
- [52:11]
Alan
Like I can't do long division. That's actually fine.
- [52:14]
Ethan Mollock
That's right. I mean, well, I mean there was, you know, and when, when, when cell phones, there was actually worry when cell phones came out that we would not being able to memorize long strings of numbers were going to be a problem. There was actually like a New York Times editorial about this. Like, I mean, so that's the positive one. But on the other hand, it's also none of these things aimed at intellectual work the way they have now, as broad based as they are. So I actually think this is a massive concern because we've built society, especially white collar, high end, expensive work is an apprenticeship model. Same model we've used for like 4,000 years for high end work. Right. You're the apprentice priest, you know, in ur, you're, you know, or the scribe. And now you're, you know, and now you're an apprentice lawyer or banker, whatever, right. We teach you generalists in school and they become specialists when we send them off. And the way we teach them is, you know, this is why it's a good times when we worry too much how we're teaching people in school. I'm like, well, it's still better than what we do when we send people off for their actual apprenticeship, which is they, they get paired with someone random who either yells at them a lot or doesn't. And they kind of learn work, you know, you know, ambiently. And that's obviously breaking like the talent pipeline broke this summer. So if you talk to like in any job, every intern who is smart was using AI to do their work because why on earth would they ever, you know, like turn in their own work when the AI is better than them and they want a job. And every middle manager has realized that they can just turn to AI rather than interns because it never, you know, asks for an excuse or was, you know, drunk the night before or whatever and gives them higher quality results. And so that the skilling has already started. Right in terms of building the next set of skills. And I think the answer, I know this is going to be some crazy from professor. The answer is more education. I think that we have to start doing more formal approaches that doesn't have to be in universities, but more formal approaches where being a big law firm might have an advantage. Because you could think about how we formally teach a client classes to keep you moving forward as you're doing this apprenticeship. Mentorship is going to have to be much more serious about doing kinds of conversations in the room rather than give me the work and I'll criticize the work, but that changes the cost structure of these things. So I think cognitively skilling super big issue. I do think it's solvable, but it will require actual effort to solve. Like if you go to large companies, learning and development has become like a checklist thing. Like it's a reward you get, you get to go to a seminar somewhere or you get a one update thing. That's not actual classes. Like pedagogy is not a concern. It's how happy were you taking this thing? We're now going to start thinking seriously about L and D learning and development as an actual high end, high class skill we're dealing with. And we're not ready for that yet. So I think there will be a period where we figure this out again. Unless the AI just does all the cognitive work for us. I mean one fun activity is you can always ask the AI what does the future of work look like. And it will always give you this answer of like, you know, Joan wakes up on the beach and checks on the five agents doing the work. And then you ask it, well, why does Joan need to check on the agents? It's like, well no, you're right. Joan wakes up on the beach and like is told the work is done. Like, why is Joan after checking the work at all? No, you're right. Joe, Joe sleeps in and you know, wakes up in the morning and there's her ubi, payment has been given to her and you know, like, so there.
- [55:14]
Alan
Is her Johnny I've neural implant, you know, that's right, drips dopamine into her.
- [55:20]
Kevin Fraser
Joan simulates waking up at the beach, but is really just sitting in her bed.
- [55:27]
Alan
Is you know, Joan who has fully integrated with the cybernetic machine. God.
- [55:35]
Ethan Mollock
It is worrying that, you know, in the end that the matrix may have been right and our battery power ability might be our best one. So.
- [55:42]
Kevin Fraser
Oh my goodness. Well, we've talked a lot about known unknowns and known unknowns that unfortunately the folks best suited to address us, academics and our colleagues are not necessarily moving ahead on as fast as we'd like. And we've mentioned quite a few unknown unknowns about what the heck's going on. And yet we find ourselves at the state level, at the federal level, at the international level trying to develop policies around AI use cases. Some very niche trying to grapple with how for example AI should be used by specific users in specific contexts. Others looking at model level regulations, concerned about those sorts of paperclip esque scenarios and more low tail but high severity risks. If we were to drop Professor Malik into Congress and I apologize for dropping you in there, what do you say? Do you say, you know, these are the principles you should be leaning on. These are the studies you should be issuing and underwriting. What's your first kind of high level guidance to policymakers with grappling this, this tricky issue.
- [56:57]
Ethan Mollock
So this may seem weird for an academic who's on Twitter a lot, but I am not an expert on these topics and everything I'm saying now is said made up without a lot of expertise.
- [57:06]
Alan
So speculate wildly. That's kind of part of our brand.
- [57:10]
Kevin Fraser
That's a good prompt hallucination.
- [57:12]
Ethan Mollock
Okay, yes, that's fine with no regard to truth and my citations are made up. So you know, I, I'm, I'm a social scientist, right, at a business school and I find a very compelling case for how to generally approach law and AI is Josh Gans, who's an economist at the University of Toronto, has a really nice paper outlining that in an area where we don't, where, where we don't know what's going to happen, where harms and benefits are both emergent, that preemptively regulating can actually be problematic because you don't actually know what the harms are and you can actually have bigger harms as a result. And that the right answer is very fast responsive regulation to problems as they occur. Right. You don't want to pre regulate electricity and say electricity is bad or you can't have over this wattage because we don't know what that's good or bad for yet you do want to be able to respond, okay, you need safety wiring right now, right? Like it turns out you need a grounding plug for anything over this voltage. And we can't anticipate whether you need a grounding plug or not. We just don't know because we don't know the usage cases. So the short answer would be I would love to see the exact opposite of what, what I think legislation is built for, which is we need fast moving bodies that regulate. So like there are harms emerging and I don't know enough about the anti, you know, deep fake bill that just got passed to say whether it's good or bad. But like that is a clearly an emergent harm that we want to deal with right away. Like the models are good enough now for deep fakes like how do we deal with that kind of regulatory approach? And so I think fast emergency responses strike me as the most important place to be. I think it's really hard to talk about model limits because we've already passed 10 to the 26th flops and that was supposed to be like kill us all and it clearly doesn't matter, right? GPT4 level models, if you look back historically would have been good enough to trigger all kinds of warnings things. So it's hard to like preemptively say this is the bright line. I think we can look at testing and it probably would be good to have more Red team testing that's officially done by government in some way or another. And right now we're counting the labs self regulate. But I also think that's a fantasy world, right? Like that's clearly just not going to happen. So I think we have to separate what will be from what what is.
- [59:20]
Alan
So what about the use of AI in the government itself? Because you can imagine that as sort of as a massive force multiplier. Make the government much more efficient. Right. I mean, obviously we've had this DOGE experiment which has not gone terribly well, but it would be nice if in principle, the government was more efficient and AI may be a useful part of that. Again, maybe not the way that DOGE is currently thinking about using it, but in principle. And so I'm curious, from your studying of how AI is used in organizations, and I assume that you are focusing here really on private companies, what's your sense of how the government can use AI most effectively? Again, keeping in mind that the government is not the private sector. Right. There's no competition between. We only have one federal government.
- [60:08]
Ethan Mollock
Right.
- [60:10]
Alan
We don't have the same concerns with command and control in the government or command and control in private industry that we might have with the government. And of course, the government operates more slowly. It has procurement issues. I'm just curious how you think the government should and use AI and what lessons, sort of it can learn from how private organizations are currently using AI.
- [60:34]
Ethan Mollock
So the framework I've been talking about is to use AI in organizations. You can't just say use it, right. Like that doesn't get you there because then people individually use it. Some do and some don't. So you need actually leadership, I would say leadership lab and crowd. So you need a leadership like you need to articulate what you want to do. And there's lots of things we could imagine. And I've seen there's some experiments happening in Pennsylvania on this as well. They had a chatbot instance. But, you know, how do we use this to streamline, you know, streamline the regulatory process? How do we use this to increase our data capacity? How do we use this as a customer facing tool to give, you know, so they could, so people. We could route around damage and it fills out all the forms for you for a process that we can't otherwise get rid of. There's a lot of things you can imagine using it for. But we do need some articulation from leadership level that doesn't have to be from the president. Right. It could be. Or even from a, you know, a secretary. It could be at a lower level. But we do need some articulation of what we want the future to look like with AI. Right. Rather than just randomly use it. And then we need to get tools in the hands of people using it with appropriate safeguards. And then we also need to be doing R and D work at the government level to figure out how this is being done. Right. The government does. You know, I would Love to see an efficiency lab approach inside of, inside of government, where we see each group thinking about, okay, how good is this at helping us with, you know, with filling out these forms that take a huge amount of time. And I think there's also this additional pressure on government is really interesting, which is in the US in private companies, there is a lot of reluctance to show AI use because efficiency gains translate to people being fired a lot. And I think the interesting thing about government is in theory, you can imagine a version which is every department's understaffed. And the benefit of getting AI to help you with something is you get to do more interesting work and you get to do less tedious work and you're serving people by doing this. And that actually is a catalyst for good AI adoption when done right. Because as opposed to I'm inventing myself out of a job and then will I get fired? I'm inventing myself out of this, you know, part of the job. I'm helping people and I get to move on to something more interesting. And, and I think, and then the other thing I would just say on the government side is the current capabilities of systems is already really high. Like the government spends a lot of money on reports and spends a lot of time on people creating reports. Deep research creates really good reports. What are we doing with that? Does that help us? Where does that help us? Where does that help us? So reinvention also becomes important.
- [62:55]
Alan
Do, do you have a sense from how AI is being used in the private sector of what it does to the ability of the folks at the top, you know, the C suite, to control their organization? And the reason I ask is that there's a huge debate, it's a very live debate right now happening in government and in the courts about what is sometimes called the unitary executive. The question about, you know, whether the, how much control should the President have over the executive branch itself? And that debate is very interesting and it's quite important, but in certain, certain ways it's kind of theoretical because even if you gave the President the theoretical power to control everyone, the President can't control all 3 million federal bureaucrats. But I can imagine a world in which you have AI systems that are embedded throughout the government or embedded throughout any organization that are tuned to reflect the preferences of, let's say, the person at the top. And so they both efficiently monitor the organization and feed information up to the top decision makers. And also they can sort of automate some of the top decision making itself, because maybe for low level decisions you can not make a decision before asking the AI which has been tuned to the preferences of the president or the CEO or the director, you know what that, what that, what that is. And that could be a way of actually operationalizing for good or for ill, high level control over the innards of any organization. And I'm curious if you have a sense of whether that might be happening or is a fundamentally a decentralizing technology and it empowers people lower level in the organization or maybe it does some combination of, of both.
- [64:21]
Ethan Mollock
So I mean that, that's literally, that that is literally the goal of AI companies is to give everyone the ability to do this. That's what an agentic system is, right? Is I give it instructions, it follows those instructions, right? And so what like that, that that's the goal of the AI organization of AI companies in large part is to create everyone, make everyone the executive over everything that they control. Because I can tell the AI, you know, make me money and it goes and does that. And I don't think they've thought through the implications for government in the same kind of way of like, you know, do I still, is this still giving instructions to humans? What is like everything you're talking about technically is possible. We actually can, the AI can, you know, identify emotions. They may not let it do that, but it absolutely can identify emotions. You can have every web camera reporting on what everybody is doing all the time. Like we are 100% in a world where 1984 is a completely doable situation with current level LLMs, right? No one's done it yet. Mostly they haven't realized they can. I probably should be talking about a podcast, but like this stuff is very doable. And creating a control system that is a panopticon is very doable. Creating a system where you must consult the GPT of the government. I don't know if it's a good idea. I think it probably isn't. And we don't know when it'll fail or not. But it's a doable thing, right? And you know, in some ways we do this already. There's a directive for the top to get rid of something or add something and then that gets carried out by every bureaucrat according to their own belief system and what their manager says. So this is all doable right now, do we want to do it? You know, like is anyone doing the R and D effort to make it happen? What is the consequences? Those are much bigger decisions. But I don't think there's a technical problem at this point, like I think that is a doable problem. And when you think about the agentic future that's coming towards us pretty quickly and you know, we don't know how far agents will get. But if you haven't played with Manus or for most people who listening, just take O3 and ask it a bunch of hard questions and open up its thinking and you'll see it, you know, it does some research, it writes some code, it does some stuff in the world or use operator. You'll see this is the goal that the AI companies want. Systems that autonomously carry out your will like you're delegating to a person. So I think that's already going to happen one way or another. These questions of who's in charge become very salient. These questions of who gets to tune these models becomes very salient. I mean if you go zoom out a little, right? Every organization, every has a similar look to it. And the reason they have looks, the reason that they are, they are cascading downward in the way that they are, the way that they have that, you know, the span of control is very similar that at the lower level managers control 5 to 12 people and no more than that or 3 to 512 people. It's all based on the idea that the only intellect we could apply to any decision was human intellect. And that came in human sized packages with human limitations. And that's why the government looks the way it does, that's why the military looks the way it does, that's why private organizations look the way they do. As soon as you have another form of intelligence that you can apply to situations that don't come in human sized packages and that can be different levels, you start to change what's possible in organizational structure. And that change is going to be felt first the private industry, but there's no reason it shouldn't impact the government for better or for worse. And guiding actively what that looks like. You know, we talked about what I would have government do. That is what I'd be thinking about a lot is like do we want the idea that everyone has a government bureaucrat in their home is 100% a doable situation? You know, what does that look like? Is that something we need to regulate against or have a constitutional amendment against? I think those are going to become very salient questions very quickly.
- [67:37]
Kevin Fraser
I'm looking forward to the third amendment scholarship on can you quarter an AI system in someone's home? Is that a violation? And to your point too, I think getting the government to think beyond merely having inventories of AI use cases and getting to the empirical benefits or drawbacks of using it. Where under OMB memo M2521, basically the administration told agencies, hey, we want you to track all these use cases. But tracking of just, hey, we used it to do X, Y or Z, but not looking at the actual impact isn't really going to be the sort of analysis we need there.
- [68:15]
Ethan Mollock
And also not bringing it up a level either of like, okay, so there's all these disparate use cases. What would be the tool we would build, right, that would let us execute on this use case in the right. In the right way? What, like how do we universalize this? How do we test it to make sure it's good? Like, you know, you've got the crowd, but you don't have the lab, right? And you have the leadership piece. So the crowd's doing work great. They're more efficient individually. But what happens, by the way then is if you don't watch, that is people just don't show you their AI use because they become 90% more efficient or whatever. And why would they ever give that work back to their employer, right? Unless they're like, they just don't have the motivation to do that.
- [68:51]
Kevin Fraser
So just a closing question for you before you hopefully have a bunch of fun long holiday weekend plans, perhaps including reading a incredible amount of AI research. How do private entities, and the government in particular, avoid a sort of tech lash to the ever greater integration of AI into critical decisions? Is it a sort of deference to the user or deference to a member of the public? So, for example, hey, you have a dispute, do you want to go through this antiquated, really long and expensive agency adjudication process, or do you want SEC AI bot to adjudicate your claim or your dispute? How do you avoid this sort of tech lash that could delay AI's adoption?
- [69:40]
Ethan Mollock
I mean, I think that part of this is taking the, you know, this is made my private industry view, but like taking the user customer perspective, right? Which is. So you want them to have agency over the kinds of decisions that they're making. And when they use AI and when they don't, you want them to be able to appeal to a human in the right kinds of situations where that's happening. But they want efficiency and they want use. Like, I would separate things for, I want a complicated ruling from, please just get me my passport. I don't understand why I can't renew it, right? And I think one of the things that we're seeing with AI is less tech lash than we, than we've seen with previous forms, including forms of AI. Right. One of the problems with AI is that it labels both algorithmic, you know, systems which, you know, large scale machine learning, which is still a thing that there's a lot of regulatory regulation about, but it's also lapsed with this generative AI. But people like generative AI, like we were talking about doctors earlier, they don't, they don't have the same aversion, you know, algorithmic aversion. They had to previous forms of AI because it's personable and it wants to help you. And that feels really good. Right. And it adapts to your level. And so I think that there is a world where we use this in ways that are very good and where people feel good about working with it because they get a helpful answer, but they know that they also is backed up by somebody who cares about making sure they get good answers. And that again, the theme that we've had through this whole conversation is deliberateness. We can't treat AI just as something happening to us. We have to actually be deliberately shaping what this is in our fields and professions. And that means in government and law, making proactive decisions. Because the goal of the AI labs is to eat everything we do. Right. Machines will do it all. We get to make decisions about that, but we have to start making those choices.
- [71:15]
Alan
I think that's a great place to close. Ethan Mollock, thanks for coming on the show.
- [71:18]
Ethan Mollock
Thanks for having me.
- [71:22]
Kevin Fraser
Scaling Laws is a production of Lawfare and the University of Texas Law School.
- [71:27]
Alan
It is produced by Goat Rodeo.
- [71:29]
Kevin Fraser
Thanks for listening.
- [72:13]
Lawfare Announcer
Got a new puppy or kitten. Congrats. But also, yikes. Between crates, beds, toys, treats, and those first few vet visits, you've probably already dropped a small fortune. Which is where Lemonade pet insurance comes in. It helps you cover vet costs so that you can focus on what's best for you and your new pet. The coverage is customizable, sign up is quick and easy, and your claims are handled in as little as three seconds. Lemonade offers a package specifically for puppies and kittens. Get a'@lemonade.com pet. Your future self will thank you. Your pet won't. They don't know what insurance is.