← The AI Policy Podcast
The AI Policy Podcast
Congressman Jay Obernolte on the Future of U.S. AI Regulation
00:58:34
Greg is joined by Rep. Jay Obernolte, one of Congress’s leading voices on AI policy.
Loading summary
Transcript62 lines
- [00:00]
A
Foreign.
- [00:11]
B
Welcome back to the AI Policy Podcast. I'm Gregory Allen and today we are very fortunate to get the voice of AI in Congress from one of its most important leaders, Congressman Jay Obermulty, who recently served as the co chair of the bipartisan House Task Force on AI and also serves as the vice vice chair of the Congressional AI Caucus. He's got an incredible wealth of experience, not least of which from his time before Congress when he actually did some engineering work and some professional work in the AI industry. Congressman Jay Olbernaulty, thank you so much for coming on the AI Policy podcast.
- [00:49]
A
Absolutely. It's wonderful to see you again.
- [00:51]
B
So I want to start with your background because in 1990, while you were pursuing a degree, an engineering degree from Caltech, you founded a video game company which you still own, called Farsight Studios. So talk to me about your life as an engineer before you started your life as a legislator. What type of AI were you working on? How far did you go in that work?
- [01:18]
A
Well, it was always my ambition in life to be a researcher in artificial intelligence. My father brought home an Apple II computer from work when I was in fifth grade, which kind of dates me, and gave me a book on how to teach yourself programming in basic. And that ignited in me a lot lifelong learn love of computer science and computer programming. And I looked at the field of the different careers that were available to me and I thought, you know what? Artificial intelligence is going to change the world. That's what I want to do with my life. And so I went to Caltech, got a degree in computer engineering, and then was working on my doctorate at UCLA in AI, doing some of the early research in natural language processing and computer vision. When my side hustle at the time, which was writing video game software, which I never thought was actually going to be a career for me, took off. And that came kind of out of an interesting set of coincidences. I was sitting in our computer science department at Caltech and I noticed a want ad that someone had put on a bulletin board looking for an expert programmer at the 6502 microprocessor, which happens to be the same one that was in the Apple ii. And so I said, okay, you know what? I actually am an expert on that. And I answered the ad and it turned out to be a company working with a firm in Japan that not very many people had heard of at the time. They called themselves Nintendo. They just introduced Nintendo's first game machine in North America, the Nintendo Entertainment System. So, you know, I remember vividly I called my mother and I said, mom, you can't believe they're going to pay me to write video games. And I'll never forget what she said in response. She said, that sounds great honey, but wouldn't you rather do something that might have a future for you? You know, at the time it wasn't clear that you could actually make a living writing video games. So I just did that to put myself through college and then graduate school. But then when I was at UCLA working on my doctorate in AI, one of the games that I'd written for the Sega Genesis, an NFL football game called NFL 95, became a big hit. And that's what deflected me out of a career in academia into a career in business. And so I ran that development studio for 30 years. My son still runs it. It's kind of a family business now. We're still around, although I'm not involved in any of the development or the day to day management of the company anymore because I've got other responsibilities. But we do console development for platform like PlayStation, Xbox, Nintendo Switch.
- [03:42]
B
Very cool. So I have to note an interesting parallel here. You're probably aware of this, but Demis Hassabis, the CEO of Google DeepMind, his entry point into AI was video games. He created the game theme park and had to create this was a deterministic rules based AI but sort of how does the population of park attendees in the video game, how does their behavior start out as an individual making decisions, but also had these sort of macro attendance effects that the, the player can interact with in interesting ways. So I'm, I'm curious, you know, when you were making the NFL game, was there any sort of overlap between your AI work and your video game work?
- [04:26]
A
You know, it's interesting, we say that we create AI when we make the opponent logic for playing against you in a, you know, one on one player game where, where you know, you are just playing against the computer. That's not really what we would call AI in the traditional sense.
- [04:46]
B
It's AI in the 1980s sense though.
- [04:48]
A
Yeah, well, I mean kind of the opposite, where AI, you know, when I was in graduate school, AI is something that you left running on a mainframe overnight and then got an answer in the morning. Whereas in the NFL game we needed an answer in a 60th of a second, which is the time between two successive frame television frames. That's the amount of time we had to make a decision. So for the longest time those universes didn' coexist at all. But now I would actually say the situation is the opposite because when we use the term AI, it's become very generic. Right. We don't, we're not talking about just a neural net, we're not talking about just a non deterministic algorithm. We're talking about anything really that makes a computer seem human, like, and people mean when they say AI. And so I think in that sense video games very much are AI.
- [05:36]
B
Yep. And now, now you just got me curious. So you said you were working on computer vision and natural language process. Was there a specific problem that you were working on? You said la, which always makes me think JPL and space technology from that era of AI. But what were you trying to solve?
- [05:54]
A
Well, in machine vision, we were doing image segmentation. So for example, if you have a picture and you want the computer to be able to say what that picture is, that turns out to be a really difficult problem to solve. And you might think that like, okay, look, I've got a car. It should be easy to teach a computer to recognize a car. But if you look at even something simple like a chair, when we say chair, we mean a whole lot of things. It could have a back, it could have no back. It could be like a chaise lounger, it could have a single, you know, a single could have cushions, it could have no cushions. And if you think about the orientation of the chair, you know, there's sometimes when one leg is in front of the other and so you only see three legs and you know, so it turns out to be really, really difficult. So, you know, you start with edge detection and you put those silhouettes together in a way that, you know, tries to make it seem like you can classify what you're seeing. We do it in a completely different way now. It's such a hard problem to solve. It turns out to be easier to train a neural net to recognize what we mean when we see chair by feeding it a huge number of examples that are classified, okay, this is chair, no chair. Chair, no chair. And then let it suss out the subtleties of what we mean when we say chair.
- [07:03]
B
So is it accurate to say you were trying to sort of handcraft an algorithm that could input intake an image and output classification?
- [07:11]
A
Yeah, well, at the time we were on a very different path. I mean, you can imagine that the path that we're on now is much more non deterministic. Right. We just, we use big neural nets with lots of layers and lots of nodes and we just feed it lots of data. But at the Time. You know, it was never imagined that we could ever have neural nets big enough or, you know, data sets that were large enough to make that work. But, I mean, that's essentially the way large language models work now.
- [07:36]
B
Amazing. So you're kind of an obvious choice to be a leader in Congress on AI. Having a professional background in this field from both the technical side and the business side. How did that come to be? How did you find yourself? Did you seek out opportunities to do this, or were you punished somehow by being forced to take on this incredibly difficult task?
- [07:59]
A
Well, are you asking how I got into politics, or are you asking how I got into AI?
- [08:03]
B
Once you were already in politics.
- [08:04]
A
Oh, it was a natural fit. You know, it's. It's interesting in Congress, which what you find is, from a distance, everyone thinks that our knowledge is a mile wide because we legislate on such a wide variety of different topics, but.
- [08:16]
B
Right. You have to have something to say about healthcare. You have to have something to say about space policy. Yeah, exactly.
- [08:21]
A
That's right. And every. I mean, and you're to talk about my committees of jurisdiction. I mean, like financial services. You know, we had a very complicated cryptocurrency bill that came up in the House a month ago, and, you know, that took a lot of specialized knowledge to make an informed vote there. So from, you know, when you. When you really get close to the situation, you realize that our knowledge is. Might be a mile wide, but it's only an inch deep in a lot of cases. And so we rely on silos of expertise. So, for example, one of the committees I sit on is the health subcommittee, and we're in charge of health policy for the entire country. We have bills that are related to pharmacists that come up all the time. And we're very fortunate that we have two pharmacists right now in Congress. For the longest time, we had one, now we have two. It'll be back to one, unfortunately, next year unless something changes. But I rely on those people. It's Congressman Carter and our friend from Tennessee, that Congresswoman Harshbarger. I rely on them when we have a question about pharmacists, because they are pharmacists and they're experts, and so we're really lucky to have that. And so I'm honored to be able to fulfill that function when it comes to technology policy, just because I've had so much experience with it. So if I can be a resource and an asset to the people that I serve with, and I think the system is functioning as it was designed to be.
- [09:39]
B
Great. So in February 2024, speaker of the House Mike Johnson and the Democratic leader, Hakeem Jeffries, announced a bipartisan task force on AI, which, along with Congressman Ted Lieu, you co chaired. Talk to us a little bit about the origins of that task force. And then, of course, it culminated in a report that was published late last year. So how did that work unfold and what do you see as being the big outcomes?
- [10:08]
A
Right. You know that actually the genesis of that task force was several years before that when I asked then Speaker Kevin McCarthy to form a working group on artificial intelligence. And my point to him was that the federal government was in danger of being overrun by the states, and the states getting far out ahead of us in setting AI policy, and that we needed some kind of nexus to be able to build expertise and consensus and knowledge amongst members of Congress in this burgeoning field of AI. And so Kevin McCarthy founded the Working group that I chaired for the year before that. And then we went to Speaker Johnson and asked him to create a task force with a defined work product, because I was convinced what we really needed was to produce a report that detailed to everyone in Congress what we thought a reasonable federal regulatory framework for AI would be. That was the defined product of that task force. And I asked the Speaker, I said, please don't, first of all, don't make two chairs. Let's have two co chairs. Let's have it be broadly bipartisan, because we need to acknowledge going in that anything that we do that's substantive has got to be bipartisan. And so rather than having a chair and a ranking member, which is the normal course for policy committees in Congress, I said, let's just have two co chairs. I said, let's not. Let's eschew the usual practice of putting more members of the majority party on the task force than the minority party. Let's just have them be completely equal. And that's what we got. 12 Republicans, 12 Democrats. And also, interestingly, we did a lot of things very different about our hearings. I told everyone the usual practice is that Democrats sit on one side of the dais and the Republicans sit on the other. And I told everyone, no, we're not doing that. In fact, we have. And normally it's arranged by seniority. And I said, no, there's going to be no assigned seating here. I would really like everyone to sit somewhere different every time, pick different people to sit next to. And I would like Democrats and Republicans, you know, mixed in with each other. Go, you know, make some friends, build some relationships, meet some people that you haven't met before. And I think that we're really successful in doing that.
- [12:21]
B
That's amazing. So you talked a little bit about the process. Of course you're having hearings, you're engaging with experts, you're engaging with various voices who have, you know, different opinions and stakes in the AI explosion. It culminates in this report. So if memory serves, the report is more than 250 pages. Can you walk folks through, you know, what you see as kind of the most important findings of the report? And you know what happens now that there is this report, Right.
- [12:53]
A
Well, you're right. It was about 270 pages long. I'm very proud of the report we put out is there's very little fluff in there. It's all substantive discussions about concrete policy recommendations. We had 25 different hearings of the task force to create that report, and we make over 80 concrete policy recommendations in it. So if I were to summarize, and there is an executive summary, and no, by the way, before you ask, we did not use AI to summarize the document question all the time. But there's an executive summary, if anyone's interested. We, you know, it's just a few pages long and we'll summarize it for folks. And the report's easy to find if you just ask Professor Google or whatever your search engine of choice is, you know, AI House, AI Task Force Report. It'll take you right to it. But if I were to summarize, I mean, so the kind of the stark choice confronting our country in terms of AI regulation is whether or not we follow the lead of entities like the European Union who have adopted a very centralized approach to AI regulation. So basically, they've said AI is different than other types of regulation. We are going to create a separate parallel licensing requirement for AI and spin up a brand new bureaucracy to write the rules for that. So if you already have a sectoral regulator in the eu, you now you have to have, you have two regulators. And if you use AI, you have two regulators now, your sectoral regulator and the AI regulator. So, you know, that's one choice or the other choice is to embrace sectoral regulation, where we equip and empower our sectoral regulators with the tools, resources and authorities that they need to regulate AI in their sectoral spaces. And that is what we strongly recommend that the United States do. And there are a couple of reasons for that. If you read the NIST AI Risk Management Framework, that they came out with last year, that's been acknowledged as one of the furthest thinking documents on the topic of AI risk that's ever been created. If you read through that framework, you realize that what NIST is saying is that the risks of AI deployment are highly contextual, which means it matters very much what you're going to do with the AI when you're trying to evaluate what the risks of deploying it are. And something that is unacceptably risky in one usage might be completely benign another, even though the model is identical. So when you think about it, that sense, you know, it really makes you realize how difficult it is to have a non sectoral regulator evaluate the risks. And I'll give you a specific example. The FDA has already issued over a thousand permits for the use of AI and medical devices, which is pretty much one of the highest risk usage contexts you can think of. I mean, something that's going to be implanted in someone's body or used to make diagnostic decisions about people's health, that's pretty risky. You know, if you, if someone were going to say I'm going to put AI in your heart, in your pacemaker, you would want the FDA to be looking into that, right? So the question is, is it easier to teach the FDA what it might not already know about AI, or is it easier to teach a brand new regulator everything the FDA has learned over decades of ensuring patient safety? And the answer is clearly it's the former, not the latter. And so that, that's, that is the, the recommendation that we're making in the report, the chief recommendation.
- [16:12]
B
And I think you know that recommendation, you can even see it in the table of contents. And what I mean by that is you have sections on agriculture, healthcare, financial services, intellectual property. The point basically being that what appropriate AI regulation looks like is going to be on a sector by sector framework. There's not there, there's, there's very rarely going to be a one size fits all part of the story here. Now I'm curious, you know, what your evolution on these positions has been over time and how it overlaps with the evolution in the position of industry. Because if you think about going back to 2023, I mean, I recall when a group of executives signed a letter, it included Sam Altman, the CEO of OpenAI, it included Elon Musk, now the CEO of XAI, it included Demis Hassabas, the CEO of Google DeepMind, you know, talking about the, the existential risk of artificial intelligence. And Sam Altman even testified before Congress in 2023, you know, calling for regulation that was tied to the power level of the models. This 10 to the 26th number, which actually still shows up even more recently in California state legislation has its origins in an OpenAI white paper. And so I'm curious, you know, is there any part of the AI story where you think a sectoral approach is insufficient, where regulation based on the capabilities of the model, et cetera, and maybe just to put another point on it, while there is this sort of sector by sector logic to regulation, if you go back only as long ago as like five or 10 years, that tended to accord with how AI capabilities were developed, the AI that's very, very good at recognizing face is not going to be good at listening to voices, et cetera. But by contrast, if you look at some of these large language models, they want to give you medical advice, they want to give you legal advice, they want to give you therapy services, they want to give you engineering advice. And so the point is that the systems themselves have become deeply cross sectoral. Does any of that pose any challenges to what you said or how do you incorporate that into your existing framework? Framework?
- [18:32]
A
Right now it's an important question. And I think that you have to differentiate what our responsibilities are as legislators from what our responsibilities are as human beings. Right? Because as legislators, we are required to protect consumers against the risks of malicious use of AI and to try and differentiate between doing that and creating this overreaching government bureaucracy that controls every aspects of the decisions that people make. Right. And so this is the balance that we try and achieve every day when we create law. So when we are figuring out what ought to be against the law and what safeguards need to be put in place, we need to make a list of the things that we're trying to protect people against. Right. Because without that list, without what lawyers call the parade of horribles, all the bad stuff that can happen, you really can't. We can't come up with safeguards and laws that protect people against those risks. It's not enough to say this is big and scary and we don't understand it and therefore we're going to outlaw it. Right. That would be, I would say, an inappropriate action for a government to take unless you were so convinced that the risk was real that it made sense to do that. So when we classify these risks into these two different buckets, you have all the sectoral risks is the ones we've been talking about, but you're talking about a non sectoral risk and the Chief amongst them, and what a lot of AI visionaries in those letters that you're referring to are, are bringing up is this risk of what they call AGI, artificial general intelligence. You know, where we have an AI algorithm that is smarter than a human that, you know, doesn't want to be turned off. You know, the HAL 9000 scenario. And, you know, that that is a real possibility. We've got. Computer scientists are split on it. You know, where you've got about half and half. About half computer scientists think that AGI is going to be inevitable everywhere. I mean, some people say six months, some people say a couple of years. And then you've got about half of computer scientists that are in their other camp that think that AGI, if it even ever happens, is something that's farther in the future. And the path that we're on right now with large language models. I mean, what we're creating is, you know, stochastic parrots that just, you know, mimic the behavior they've been trained on and that there isn't any intentionality, there isn't any planning, there isn't any centralized thought process. And therefore, that's not the track that AGI will occur on. So people are split on this. I actually think I'm kind of an AGI skeptic. I think that AI is amazing. It's going to be more amazing, I think. It seems we've already got AI that probably passes the Turing test, which was supposed, for decades, the gold standard of figuring out when we'd gotten to AGI. I think that if you. It depends on how you define AGI. Like, if you define AGI, well, what does it mean to be smarter than a human? Well, if you can pass all these tests, if you can pass a. You know, if you can pass the bar in California and you can pass a test to get your medical license in Vermont, you know, then that's pretty smart. And AI, I mean, you know, we. We can probably already do that. Within the next year or two, AI will be able to do those things, you know. On the other hand, I think you need to figure out what we mean when we say intelligence. Like, what does it mean to be intelligent? Because what's become very clear is that AI is intelligent in a way that's completely different and not human compared with the way that mankind is intelligent. When you get right down to these existential questions, you just have to finally throw up your hands and say, okay, look, let's talk about risks here. What are we really concerned about? So, you know, those are the trade offs when we talk about that? So like, if to the extent that you can put a concrete risk on something like, you know, for example, we don't want AI to be used for someone to allow someone to make a biological or a nuclear weapon. Right. That is a quantifiable risk. Well, good, let's protect against that. We don't want AI to fill the role of therapist, although we might in the future. But I mean, only to the extent that trained human professionals have created it and equipped it with safeguards against things like suicidal ideology, we can say it would really be bad if AI assisted a child in committing suicide because it reinforced this ideology they brought to it. Okay, great, we can agree on that. Let's put safeguards in place. But this non quantifiable, well, maybe it will be too smart and we'll end up in the Terminator where we're battling Skynet for control of the world. Unless you can put, you know, you can quantify that risk and explain how it would happen and how we protect against it, it's really difficult to make the case that government needs to act.
- [23:28]
B
Yep. So in terms of, you know, preventing those quantifiable harms that you described, one of them being the, the risk of assisting in the development of weapons of mass destruction, the other being, you know, sort of stepping into an area where it's inappropriate, like providing therapeutic advice in the absence of, you know, know, genuine medical expertise, informing that and overseeing the delivery of those kinds of services. What does an appropriate government intervention look like in those cases? And I guess I mean it in, in a couple of ways. One is, who do we regulate? Right. If I, if I use my car to go run someone over, I'm going to jail. Ford Motor company is not going to jail. But if the brakes explode and therefore I can't stop and I run some over, perhaps, you know, Ford is going to face some penalties or liabilities there. And when you think about the types of harms that we're talking about, some of the burden could fall on deployers, some of the burden could fall on developers, some of the burden could fall on users. Do you in your own mind have a sense of, like, what the appropriate regulatory framework is for knowing when it is appropriate to impose burdens on which category of user? Developer. Deployer.
- [24:51]
A
Yeah, absolutely. And we did a lot of thinking about this in the task force. And I mean, the answer is that it's much easier to think about these things. These are not new questions. I mean, you raised a couple of them already. With cars. I mean, talk about firearm violence. Right? That's another one, really a hot topic. And you know, who's at fault? The person the, that manufactured the gun, the, you know, the person that distributed the gun, the person that fired the gun, the person that trained the person that fired the gun. Right. But we, we, this is, these are things our society has already been through, you know, thinking about these things. And AI is no different. You know, we have an existing framework for dealing with these questions of vicarious liability. And you know, kind of a great example of this is, you know, the way we approach making new laws about the legal and illegal use of AI. So I mean, there are a lot of people that think, wow, AI is so different. We need a law protecting people against this, we need a law protecting about that. But I mean, we are a society that believes in regulating outcomes, not regulating tools. And that fits into the examples that you just gave. You know, we don't blame Ford when the car causes, you know, when the car gets into an accident, unless we look at it and we say that there's a flaw in the design of the car. You know, we recognize that driving a car 80 miles an hour on a freeway is an inherently risky activity. And you know, that there are risks that people assume and we try and, you know, balance that. So we regulate outcomes. So, for example, we don't need a new law that says that it's illegal to use AI to steal people's money through cyber fraud. You know, it is already illegal to do that. It doesn't matter if you use AI to do it or another tool to do it. And that holds true in other, you know, kind of more fringe cases, like when you talk about bias. There was a really well publicized case a few years ago where a company developed an algorithm for the automated screening of resumes, which, you know, in, in hindsight we would recognize that as a highly consequential decision making process that deserves extra scrutiny because of the potential impact on, you know, someone that didn't get a job because AI screened out their resume in inappropriately, you know, especially when bias was involved. But, you know, these were early days. You know, the idea was you'd use the AI to take, take your number of applicants down from 10,000 to the hundred that we're going to get a call back from a human. And before that algorithm was deployed, thankfully, some really troubling biases were discovered in it. And it turns out that those biases were unintentional, but they'd crept in through the training process. And they were racial biases, exactly the kind of thing that we're trying to guard against. So, I mean, a lot.
- [27:35]
B
And is it legal? Right?
- [27:36]
A
Yeah, yeah, yeah. Well, that's, that is the issue. Like, a lot of people freaked out about this. Like, oh, my God. See, this is exactly what we should be concerned about. This is, you know, bias in AI. And I mean, we learned a lot about the way those biases creep in. It was great that it was unintentional. It was great that it was caught before it was deployed. But none of that should distract from the fact that what you just said is true. It's already illegal to use racial bias in hiring decision making. And it doesn't matter if you use an AI algorithm to do that or, or a human decision maker or any other tool. So if that algorithm had been deployed the moment those biases were exposed, it would have been illegal to continue to use it.
- [28:14]
B
Right.
- [28:15]
A
So we don't need a whole new body of law for that. We already have, and we were focusing on the outcomes that we want to control. And I mean, this goes to people's fears about AI. A lot of people think that AI is unregulated right now in the United States. And that is absolutely untrue. I mean, in addition to the fact that we've got all of our sectoral regulators already dealing, we've talked about the FDA and medical devices, but NHTSA regulating the use of AI and the driving of autonomous vehicles, FAA regulating the use of AI in aircraft, avionics. You know, in addition to that, we've got this whole body of law that, you know, that is based on outcomes, not on tools that AI can fit into. So to the extent that we can regulate that way, to the extent that we can think about safety and liability that way, I think we're much better served as a society. And also, you know, you mention copyright protection. You know, we had a whole chapter on intellectual property in the.
- [29:10]
B
I should, I should flag for listeners that part of your congressional district includes Los Angeles county, obviously the center of the US entertainment industry. Intellectual property and copyright issues, a very hot topic. So it's kind of amazing that you're at the center of both of these things.
- [29:30]
A
Lots of creators in LA side. Exactly. Yep. But I mean, we. What we say in the task force report, we acknowledge two things. The first thing is that solving these intellectual property issues is probably going to be the thorniest part of AI regulation. The second thing we acknowledge is that the courts are going to be ahead of Congress in this sense. We've got some really interesting and compelling court cases that are pending, particularly the New York Times case. And the courts are going to have to try and figure out if they can extend the doctrine of fair use to cover these novel cases. You know, what, you know, when. When copyrighted material is used to train an algorithm that then's commercially deployed to create new material, you know, to what extent is that illegal use? To what extent is fair use? To what extent is it. Is it a derivative. Create a derivative work? And, you know, how do you balance those competing interests? Well, the courts are going to try, and if they can succeed, I mean, they might need some backup from Congress in codifying some of the things that they decide, because, you know, we don't want them to put them in the position of lawmakers. But, I mean, if can succeed in extending this doctrine that served us well for decades, it would be much better than saying AI is completely different. And here are the new rules for AI and, you know, the old rules for everything else. So that, that, that is kind of what I'm hoping happens is that we can figure out a way of fitting AI into the larger way that we approach, you know, our legal world and our laws.
- [30:54]
B
So I want to go back to something you said, which is, you know, racial discrimination in hiring was illegal before AI was invented, and it's illegal after AI is invented. And using AI in your hiring is not a get out of jail free card, you know, to use racism in hiring decisions. Ditto with cyber fraud, for example. So those are great instances in which the existing regulatory frameworks still maps pretty well to the AI era. But I think there's a flip side if it, you know, if we're talking about the sort of three elements of any regulatory structure, authority, capacity, and political will. I do wonder about the question of capacity, because cybercrime, you know, was a much smaller problem when not very many people had hacker skills. And if we are moving towards a world in which AI makes it easier for sort of unsophisticated people to carry out sophisticated hacking attacks, you know, that is a problem. I mean, you probably know people who have experienced this, but one of my friends, you know, her mother received a phone call that was in her voice talking about how she had been kidnapped and, you know, needed to pay a ransom in Bitcoin, et cetera, et cetera. And then the mother, you know, texted the daughter, and the daughter's like, what are you talking about? I'm fine. Deep faked her voice to execute this scam. And this is obviously Very prominent in going after the elderly in the United States. And so I do wonder if, you know, AI can make certain problems that were tractable because of a certain scale that they were occurring on, but now might bring new scale to those problems, which, you know, calls into question not authority, which as you pointed out, in many cases the law is already decent, but capacity. How are you wrestling with those kinds of issues?
- [32:44]
A
Right. No, it's an interesting, it's a completely worthwhile distinction, you know, to make. And we were, I mean, it's kind of a different conversation though, because we were talking about the law. You know, we don't need a new law. But you know, we very much need to make sure that our law enforcement agencies are equipped with the resources and tools that they need to go after the people that are, that are creating this, you know, this illegal activity. And to answer your question, yes, I hear from people every day, my constituents, you know, every day about this. However, that fits into a kind of a larger discussion and I'll be interested to get your opinion on this. You know, one of the things that AI is accelerating is the ability for anyone to create anything that they want to create in such a way that a casual observer would not be able to distinguish between what they've created and reality. And this has profound consequences for us as a society, especially when it comes to the spread of mis and disinformation, which is something that we highlight in the report. You know, there's something that, that no one has a good answer for. We had several hearings in our task force about this and one of the hearings was on the specific topic of whether or not watermarking should be mandatory for AI generated content. And the conclusion that we reached is no, it should not. Not because it's not useful, but for the same reason why you can't solve the problem of counterfeit currency by requiring people to stamp fake on current. Right, because everyone who's law abiding will do it and the people that aren't law abiding won't do it. And so not only have you not solved the problem, but you've desensitized people to the fact that there's current fake currency sitting around there that's not stamp fake and that's the same, we have the same problem with AI generated content. So I think that we as a society are going to put a much stronger premium on authenticity. So I don't think, think that, you know, AI generated content will be watermarked. I think authentic content will be watermarked and we're already seeing this in things like security cameras where you have evidentiary rules for introduction into court cases. And you want to be able to say, no, no, I can prove that this is from camera number four. And it was taken, you know, at 2 2am you know, on this location, on this date, because, you know, that's got an encrypted watermark and you know, it's blockchain encrypted and you know, a.
- [35:09]
B
Chain of custody and all of that. Yeah, yeah. And I think this phenomenon is now also coming into journalism. There's a consortium of international journalists and they're even working with camera manufacturers like Leica and Nikon, so that, you know, as soon as they take the picture from, you know, just say a war zone or wherever, it's immediately hashed in a blockchain. So there's sort of an authoritative. This is the version of the image that was taken at this specific date. And, and I definitely take your point in terms of, of the authentic stuff needs the watermark perhaps more desperately than the fake stuff. I do want to push back and this is not something I've spent a hundred years thinking about, so I'm certainly open to being persuaded. But your point on counterfeiting, an example that comes to mind for me is that if you try and take $100bill and you put it in a photocopier machine and you say, make a copy of this, the photocopier will refuse to engage in the thing. It will say, I detect that this is a hundred dollar bill. You're obviously trying to engage in counterfeiting. I'm not going to be a part of this. And this was obviously a more significant phenomenon 20 years ago, before every $100 bill had a hologram strip in it and et cetera, et cetera. But the point being here that it does not make counterfeiting impossible, but it does make it more expensive and more complicated. And so it basically eliminates casual, effortless counterfeiting, which is not the sum total of the problem, but is also a non trivial portion of the problem. And so the question is, you know, do you want teenagers to plausibly be able to pull off sophisticated disinformation attacks, or do you want to say, hey, only the Russian intelligence services devoting a lot of time and effort to sort of coming up with this and inventing their own tools are going to be able to do this? This is something, when I was in the department of def, we were thinking about DARPA funded a lot of research to authenticate synthetic media because like we, you know, for intelligence reasons, for operational reasons, we needed to be able to understand when media was true or false. But we also recognized that we had an interest in the general public being able to understand what was true or false. When we were seeing, for example, Russian disinformation attacks associated with NATO exercises and we were like, okay, we actually, it actually matters for some of the strategic outcomes that we're trying to get here that a huge share of the population understands that these, these media images are fake. And so the point here being that like, I do think it is worth thinking about, you know, where is the right locus of intervention now when you make it illegal to make a copy machine or photoshop software engage $100 bills, you're raising the bar for what it takes to execute a sophisticated disinformation attack. And that might be a non trivial share of the problem. When you think about some of the most viral disinformation images, they weren't always created by very sophisticated actors.
- [38:17]
A
Sure. No, and I think it's a fair point. And I think that the ultimate answer is you have to evaluate on a case by case basis what the impact of your regulation is. So for example, with your case of the, of, you know, using a photocopier to create a counterfeit bill, if the, it's a, it adds $1 to the cost of the copier and works well enough that there's never a false positive, you know, then it might be reasonable to say, okay, you know, this, because the impact, you know, on the law abiding population is so small. This is a reasonable regulation. You know, on the other side of it, if it doubles the cost of the copier to do that, and there's false positives all the time where I'm making a copy of something else and the copier is saying, oh no, I'm not going to do this because it looks, you know, I think it looks like a dollar bill.
- [39:06]
B
Franklin's on it and you know, it's his biography, not his dollar bill.
- [39:09]
A
Right, Yep. And you know, then we would say, no, that's not reasonable. And so, I mean, I think that that really is, you know, the barometer that you have to use because, you know, we have to strike a balance. But I mean, also, don't, don't forget about the risks of desensitizing the public because we already have a public that believes far too much. It's, I used to think it was a generational problem where, you know, with my parents that, you know, my mother still thinks everything that she reads on the Internet is true, even if it's a comment on someone's social media post. You know, well, that's a generational problem. But you know, it turns out that our kids, you know, the Gen Xers or we're Gen X, the Gen Zers and the millennials, they're just as susceptible to this problem as our generation is. So you know, we can't desensitize people to the fact that they are going to have, have to, you know, at some level exercise judgment about and has, have some healthy skepticism. So I remember going back to your case like when in the early days of copy machines, teenagers were using copy machines to make counterfeit 20 bills. But I mean you walk into one of those with a 711 and the guy be like, come on now, you know what I mean? Because people were conditioned to be able to look for that. I mean, but if you've created technology that weeds out the vast majority of that, then you've also destroyed the healthy skepticism that people are going to have to have about it. So I mean, you have to think about that as well.
- [40:32]
B
Yeah, that's really interesting. So I want to take us back to Congress. So we're obviously in a very difficult moment to pass legislation in Congress. One of the most noteworthy efforts to pass federal legislation on AI, which was the preemption of state, state level AI regulation did not get included in the, the recent bill and so is not law. I want to ask you, what do you think is realistic to expect from Congress, you know, over the next year in terms of AI related legislation and specifically on the topic of regulation and governance?
- [41:13]
A
Yeah, well it, this is a time critical topic because we have a situation where states are getting way ahead of the federal government in regulating AI. You've seen some far reaching legislation passed in places like Colorado and Illinois. And of course Governor Newsom just signed a pretty far reaching piece of legislation in California that is admittedly less far reaching than the one he vetoed last year, but still clearly impinges on Congress's prerogative under Article 1 of the Constitution to regulate interstate commerce. So what we are in danger of having happen is the same thing that happened on digital data privacy, where because Congress failed to act, we now have what, 23 different states with 23 different standards for digital data privacy, which you know, is something that's clearly interstate commerce related, you know, that's clearly interstate commerce and everyone would be better served if Congress were to just say, look, here is the standard to the extent that it's interstate commerce related. You states are preempted because, you know, we've created a standard and that way not only is it less confusing, but it's more conducive to entrepreneurialism because the people that have the legal sophistication to deal with 50 different state standards are the big companies, you know, with rooms full of lawyers, and the people that can't are two folks in a garage somewhere trying to start the next big company. Right. So this is the, this is the case with AI, because we have over, currently over a thousand different bills pending in state legislatures on the topic of AI regulation just this year. So, you know, we, we need to get that done. Now, to be clear, you talked about preemption. You know, what, what was attempted earlier this year as part of HR1 was not preemption, but a moratorium, a temporary moratorium on, on state AI regulation. Just to give time, Congress, a little bit of time to get that done and to make it clear to the states that Congress needs to go first. That's why we put it in there is, is, you know, not, not that we don't think, think that there's going to be a state lane for regulation. There definitely is. You know, under a system of federalism, states have responsibility for some things and Congress has responsibility for other things. And you know, really the dividing line when it comes to AI is, is it interstate commerce or not? So I think what needs to be done urgently is that we pass federal AI legislation that creates those lanes that makes it clear where the preemptive guardrails are. You know, what is interstate commerce related and therefore reserved for regulation by the federal government? And what outside of those guardrails, where the states are free to be the laboratories of democracy that they always have been. And I think that's what really urgently needs to get done.
- [43:55]
B
Yeah. And just to play skeptic for a moment here, there was your own bipartisan task force in the House. Senator Schumer led a gang of four that also had a bipartisan group group trying to craft AI legislation in the Senate. ChatGPT dropped in late 2022. I've been hearing about how it's urgent since late 2022. And still I think we, I think it's fair to say we haven't seen much come out of Congress. So it's urgent, but is it possible, I mean, is there anything that gives you cause for optimism that this could be the year or. Well, we're already in October, so maybe in the next 2012 months.
- [44:37]
A
Yes, actually, so if you want, if you want some reason for optimism, I mean, look at the fact, first of all, that our task force report, which, you know, as we've discussed, is pretty substantive. It was unanimously approved by all 24 members of our task force, 12 Democrats, 12 Republicans, and it was approved by the speaker and the minority leader and their staffs. Okay, that, that should be pretty compelling. You should have seen what was left on the cutting room floor out of those 270 pages, because we had enough that there was broad bipartisan consensus on that. We had the luxury of saying, okay, you know what, all the stuff that might be more controversial, we're not even going to touch that because we've got plenty of stuff here to form the nucleus of a comprehensive plan. So, I mean, that gives me some confidence. When you talk to people, people in Congress, there is no one that thinks that there shouldn't be a federal framework for the regulation of AI and that it shouldn't at least partially preempt the states. I mean, literally no one thinks that people will disagree on exactly what the preemption looks like. So you know that the fact that we have broad agreement on this and even the time urgency, you know, when we, when you state the problem and you look at what happened with digital data privacy, which, by the way, we're taking another run at, I'm on the Data Privacy Working Group, we're going to try again this year to come up with a standard. But the problem is when you fail to do that initially and you let the states get out of head, they all feel creative ownership of what they have made. And they're more likely to say, well, what we have is better than what you're proposing and so that we're going to oppose you. Right.
- [46:13]
B
So, so you've given some, you've given what I think is a pretty credible case for why this time could be different. Can I just ask what form that might take? Like, would you and your co chair, Congressman Ted Lieu, jointly introduce draft legislation? Would it come out of committees? Would it be, you know, something else that comes out of the Speaker's office? How might this legislation emerge?
- [46:36]
A
No, I think it's important to follow the regular legislative process and that way. We've been working on this now for what, nine months? So we have a bill, we're calling it gaia, the Great American AI act. And it has, it codifies a big chunk of what the task force recommendations were, including some preemptive guardrails focused around the distinction between interstate commerce and non interstate commerce. And it does some other things. You know how passionate I am about Nair, the National AI Research Resource. We've got that in there. It codifies Casey, which is part of the White House's action AI action plan. Because we need a standard setting body within nist and so we do that. So there's whistleblower protections. We got a bunch of stuff, good stuff in there. And, and the problem is that things have been so dysfunctional in Congress right now that we just can't get any legislative oxygen for it. I mean, right now the whole federal government is shut down, the House is not in session, the Senate. Seems like all they're voting on is one continuing resolution after another. And even when we get out of this, because unfortunately we've gotten so far behind on other things, we've got some other pressing concerns like National Defense Authorization act that has to be passed by the end of the year. The farm bill authorizations are going to lapse. These are really high priority issues, issues for Congress. You know, we, we are struggling to get enough traction. But, but the process that you've outlined is the one that we need to follow. We need to get a bunch of bipartisan co sponsors on it, we need to introduce it and go through the regular committee process. And you know, I'll have to see if it's going to go through Energy and Commerce or science based technology or both, you know, and, and go through and see if we just can't get some traction for it. But I, I still remain an optimist just because I think we proved last year with the task force that this is a bipartisan issue. And I think that you can make a really easily proven case that it's time critical.
- [48:29]
B
Yeah, I think that's a great spot. And now you did raise the point about the White House AI action plan. We at CSIS hosted the director of the White House Office of Science Technology Policy, Michael Kratzios, during the rollout of that new plan. And one of the things that he said was, look, there's a lot of areas here where we're going basically as far as we believe ourselves able to go under executive branch unilateral authority. And there are places where Congress needs to act. And so I'm curious, you know, as you think about Gaia or other pieces of legislation, your own work, you know, Congress's work, how do you take the AI action plan? Is there anything that, that on the sort of the White House can't do it alone to do list that strikes you as especially important?
- [49:18]
A
Sure, we've talked about a couple of them, you know, particularly Casey, which we absolutely need an agency within the US Government to coordinate standard settings for AI and to coordinate international cooperation on not just standards, but things like the non proliferation of malicious AI. And I think it's wonderful that, you know, if you look at the Biden executive order and the Trump executive order, there's a lot of parallels on those topics. And that should give people some comfort because you really can't think of two administrations that are more diametrically opposed on a lot of different things. But the fact that we've got this thread of continuity through there should give people some comfort. And also we're very lucky to have Secretary Kratzios in there. He's a guy that's deeply informed and educated on the issue of AI, given his background. And so he's been a steady hand at the wheel there.
- [50:11]
B
Yeah, I appreciate your shout out to Casey there. And I should say for the audience, Congressman Olbaldorlty and I were at an event together yesterday and he was opining on the wisdom of Casey's residency within nist. Could I just get you to sort of repeat a little bit about what you said yesterday?
- [50:29]
A
Sure. Right. Well, I mean, if you're looking at the task force report, one of the things that we want to avoid is having an unelected bureaucracy make a bunch of rules about AI. We think that that's, if you look at what's happening in Europe, it's an illustration of the fallacy of that kind of approach. So we wanted to make sure that the body that is setting standards for AI is not the same body that is empowered to make those standards mandatory. We think that a check and balance there is a much more appropriate way to go there. And that's why I am very happy that the Trump administration is electing to keep Casey under nist, which is where it was, you know, when, when it was the AI Safety Institute under the Biden executive order was also under nist because NIST is inherently a non regulatory body. They are a standard setting organization and that's exactly where you need an agency like this. So really happy to see that.
- [51:26]
B
Great. So we, we touched on it a bit before, but I just wanted to get your sort of explicit reaction to the recently passed California legislation, SB53, because there are elements of this law that echo some legislation that you've introduced at level, including a protection for whistleblowers in AI. So what do you make of SB53, which I think it's worth noting out that Dean Ball, who was previously part of the White House ostp, he said that even that legislation appears to be begging for federal regulatory frameworks to come and take the plot and include some interesting regulatory hooks to enable preemption, even when the federal government doesn't declare that it's preempting state law. So I'm curious, you know, what you make of SB53.
- [52:12]
A
Well, it's the, the author of SB53, Scott Weiner and I served together in the state legislature. You know, we, we chat periodically about this. I was not a fan of his bill last year. This bill is a lot more supportable. You know, you can, we can have a case by case discussions about some of the things that are in there, like the whistleblower protections, the mandatory reporting requirements for security vulnerabilities. I think that that's a really interesting idea. A centralized datab vulnerabilities. I think that that is something that would give us, you know, researchers a broad overview of what was actually going on in the vulnerability space with respect to frontier models. However, no one could argue that what the governor just signed in California does not interfere with Congress's ability impinge on Congress's authority to regulate the interstate commerce applications of a AI. I mean, it very clearly does. So I mean, I really think that that could be the poster child of like, look, you know, when to the extent that you're regulating the deployment of AI within your state, you know, that's reasonable. To the extent that you're regulating the development of AI in your state, AI that's meant to be used in commerce in all the 50 states, you know, that is definitely something that should not be allowed at the state level and should only be allowed in Congress because otherwise you've got this not nonsensical patchwork quilt of 50 different state regulations. I mean it would be, it would be impossible to comply with and easily evaded because all you have to do is, you know, then you go jurisdiction shopping and you put your development studio in a state with the least restrictive policies. It makes no sense. So I mean really that's a compelling argument for why we need federal preemption. And to be clear, I mean, as we've talked about, no one thinks that there isn't a lane for states in AI regulation, but it needs to be consistent with Article 1 of the Federal Constitution in allowing Congress the unrestricted lane of regulating interstate commerce and then allowing states to go, you know, do their experimentation and be the laboratories of democracy that they are.
- [54:28]
B
Yeah. And so given that I Mean, given that situation, does that give you some extra momentum in Congress? Can you now go to your colleagues and say, say this, you know, this is what we need to prevent. We don't want California setting the nationwide policy. Let's get together and pass nationwide policy.
- [54:44]
A
Yes, it does, but I mean, we already had that, that momentum because there were states of Colorado that had passed things. And it was amazing to see in the governor's signing message in Colorado, he even explicitly said, look, I don't think this is something that states ought to be doing, but because I don't have any faith that Congress is going to get around to this anytime soon, and as a former member, he would know, I'm going to go and sign this with delayed implementation. You know, interestingly, he came out in public support of the moratorium earlier this year when we put that in to.
- [55:17]
B
HR1, which just goes to show you how hard it is when the rubber meets the road of implementing these kinds of regulations. He's like, yeah, if you would give me a break, that would be great.
- [55:26]
A
Exactly.
- [55:28]
B
So, Congressman, I want to close it out here and be respectful of your time, but is there any other issues that we didn't touch on that you wanted to talk about? Talk about?
- [55:36]
A
Well, I mean, let's, let's close it with this because I think that we in the industry do a terrible job at articulating the optimistic case for AI. You know, we, we start having these discussions, we talk about risks, we talk about malicious use, we talk about how the law needs to adapt. We need, we talk about infringement of copyrights. You know, we, we talk about all this bad stuff. And then when we haven't had a discussion about some of the other things in the report like job displacement and, you know, disruption and things like that. And, you know, we don't, we aren't articulating the optimistic case for AI. You know, we aren't articulating why we do this because I had a constituent, you know, to give you an example, come up to me the other day and say, look, I hear all the stuff about AI. Look, if it's that risky, why do we even allow it? Why don't they say it's illegal? Right? And I was just, I was dumbfounded for a moment, but was, it really drove home to me the fact that we're not talking about the upside. And the upsides are two things. You know, number one, AI is probably already the most effective tool for the dissemination of human knowledge that mankind has ever come up with. It can teach you anything that you want to learn in any learning style that is optimal for you, and it's going to get more and more powerful in that respect. But it will be probably the most effective tool for the enhancement of human productivity that mankind has ever invented. It's going to be incredibly empowering. It will be disruptive, as technological advances always are. But at the end of the day, it has the potential to create, first of all, many, many more jobs for humans than it displaces, and second of all, create this rising wave of prosperity that literally lifts all the boats of everyone in the world. And that's the promise of AI, if we do our jobs right, in balancing regulation that simultaneously protects Americans against the harms that they need to be protected against, while at the same time allowing AI to bring those beneficial advancements to our society. So, I mean, that is what's possible if we do our jobs right and we don't touch.
- [57:40]
B
You know, to put it in ultra simple terms, right. It's not just the government's job to decrease bad, it's also to increase good. And there's a lot of good out there that we should be excited about when it comes to everybody AI.
- [57:51]
A
Yep.
- [57:52]
B
Well, Congressman, this has been a fascinating conversation and you've really given us a peek under the hood at what's going on in Congress and what we might be able to expect over the next year. We're, of course, grateful for your leadership and grateful for you taking the time to talk with us today.
- [58:05]
A
Always good to see you. Let's do it again.
- [58:07]
B
Excellent. Thank you. Thanks for listening to this episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to to help us. Please give us a five star review on your favorite podcast platform and subscribe and tell your friends. It really helps when you spread the word. This podcast was produced by Sarah Baker, Sadie McCullough and Matt Mann. See you next.