← The AI Policy Podcast
The AI Policy Podcast
The AI Diffusion Rule is Rescinded, AI Executives Testify Before Congress, & AI Adoption in the IRS
00:53:13
In this episode: Trump on AI Diffusion, execs testify to Senate, IRS adoption, Copyright Office report, AI & the papacy.
Loading summary
Transcript34 lines
- [00:00]
A
Hey, AI Policy Podcast listeners, Before we jump into today's episode, I wanted to provide you with a quick note on the status of the AI Diffusion Rule. This episode that you're about to hear was recorded on Monday, May 12, and as usual, the news just keeps coming. As of Tuesday, May 13, BIS announced that it officially, quote, plans to publish a Federal Register notice formalizing the rescission and will issue a replacement rule in the future. So as we're reading it, this is a takedown and replace approach to the AI Diffusion Rule. This will be a welcome news to many countries that found themselves in the Tier 2 list on the original framework, including several of the Gulf states with big AI ambitions that President Trump is meeting with this week. As you will hear at the beginning of this episode, this is almost directly in line with what we predicted and we wanted to give you that additional point of clarity. And with that, let's just jump right in today's episode. Welcome back to the AI Policy Podcast. My name is Brielle. I'm the Associate Director of the Wadhwani AI center here at csis and I'm joined today with Greg, the Director of the Wadhwani AI Center. In this episode, we'll cover the end of the AI Diffusion framework, AI executive's testimony before Congress, what AI might mean for the irs, and the US Copyright Office's latest report on generative AI training. We'll also touch briefly on what AI Policy might look like in the new papacy. So we have a lot to cover. We're just going to go ahead and jump right in. Greg. On May 7, 2025, the Trump administration announced its intention to repeal the Biden Era AI Diffusion framework. As our devout listeners will know, we talk about this all the time and this debate over this framework is something that we'll continue to cover on the podcast. But today I want to hear what exactly the Trump administration proposed last week.
- [01:53]
B
Yes, so this is really big news and I want to say that you can't always believe everything you even in mainstream news publications. The reason why I say that and why I'm grateful to folks who listen to this podcast is because the details matter a lot here. And so it's worth pointing out, like what there is some evidence for, what there is strong evidence for, and what there is no evidence for when we talk about what the Trump administration did and what it actually means for the future. So essentially, the Trump administration had four options available to them as they were thinking through what the future of export controls should look like. And specifically with this AI Diffusion rule, which the Biden administration sort of crammed into its last week of the presidency as a pretty significant export control policy shift. So what were those four options? Basically, they could keep it completely unchanged, get rid of it entirely and pretend it never happened. They could change it, or they could punt, which is basically, say, delay the implementation of the framework, give themselves more time to make a decision. So the reporting from a lot of journalistic outlets is saying that what the Trump administration has decided to do is to kill the diffusion rule. And I don't think that the evidence that we have right now actually says that. So we've got a few things. What do we actually have is like concrete evidence of what took place here. Number one, you've got the official statement from bis, the Bureau of Industry and Security at the Department of Cop Commerce. So on May 7, they released a statement stating, quote, the Biden AI rule is overly complex, overly bureaucratic, and would stymie American innovation. We will be replacing it with a much simpler rule that unleashes American innovation and ensures American AI dominance. So what does that mean? Number one, that means it rules out, keep it unchanged. Obviously, something is going to change. It also rules out number two, which was kill it entirely and go back to the way things were. You know, just before there was any diffusion rule whatsoever, they could change it, which they did not do. So I think what we know is that they've punted, right? They're not going to implement it the way it was, but they have said they're going to do something else that is going to address similar types of concerns. So obviously this is something. But I think it's wrong to say that they've, quote, unquote, repealed the diffusion rule because there's a lot of other signaling going on around right now that the Trump administration is interested in tough for export controls. We've talked about some of these data points on the podcast. Number one is the executive order that the Trump administration passed in its first week stating that they wanted, you know, tougher export controls to close loopholes. That was backed up just recently by the skinny budget request that the Trump administration put out, which included something like a 50% budget increased request for the Bureau of Industry and Security. So it's not like they're looking for looser export controls. Right there, you increase the budget by 50% because there's something about tightening that you're interested in. So that's kind of where we are, is it's being reported as a repeal of the diffusion rule, but many of the Concerns, you know, the diffusion rule was intended to address, such as chip smuggling to China, such as loopholes in existing regulations. You know, the Trump administration has said that they still care about all of those things and they want to address it. And there's, you know, stuff like the budget request that success, they're going to address it. So however they address those concerns, they're not going to address it with the diffusion rule as written, but they're also, you know, not going to abandon all chip export controls. So we're sort of left wondering what is going to come next.
- [05:45]
A
So after the BIS released their statement, has have companies or the market reacted in any way to this news?
- [05:53]
B
Yeah. So Nvidia stock is up, I think it was out. It was up about 3.1% shortly after the diffusion rule came out. Other semiconductor stocks, of course, Nvidia is big in any semiconductor stock index, but like the Philadelphia Stock Exchange semiconductor index is up 1.7%. So these are people who are optimistic. But it's also worth pointing out that like, sometimes stocks, you know, move just on vibes, right? And so if you think, okay, this is going to be perceived as good news, you know, you buy that stock for the day trading kind of an aspect of it, but with like the long term implication of this is obviously going to be whatever the long term implication of the policy is. So I don't know that we necessarily know a lot based on that short term jump.
- [06:33]
A
Okay, so then on May 9, the senior white House policy advisor, David Sachs, posted on X the rationale behind this policy change. He gave a little bit of an explanation. I was wondering if you could walk us through that.
- [06:44]
B
Yeah. So David Sachs, who is often referred to as the White House AI Policy czar and the White House AI Crypto policy czar, put out a pretty long post on X that is definitely worth reading at length. I mean, it's not an especially long thing. It's like 500 words or so. One thing I should point out here, which is his, his X account says personal views only. So on the one hand he is obviously, you know, a part of this administration, but on the other hand, right. There's this disclaimer that he's not claim mean to, to speak for the administration. And I think that's kind of where we are. The diffusion rule, as we've been talking about on this podcast for months now, has been a interagency fight. There's different communities within the Trump administration who have different opinions on this. David Sachs has effectively, you know, revealed, right. That in that interagency debate. He was obviously a strong opponent of the AI diffusion rule. So I'm going to read some long quotes from this relatively short post because I think it's really worthwh. First, he said that this was an overreach of export control authority. So he was saying that this rule marked an unprecedented and arguably unlawful expansion of export control authority. Under the Export control Reform Act ECRA of 2018, the President is empowered to restrict exports of dual use technologies that have both civilian and military applications. And then it goes on. But the diffusion rule went significantly further. It required nearly all global sales of high end GPUs, even to trusted allies to obtain export licenses or to fit a narrow set of license exemptions. So here I think you have to say I don't think the Trump administration is going to agree entirely with what Sachs just wrote. This is where I think personal views, you know, are important. When have you ever heard the Trump White House say that they do not have the legal authority to take certain actions? And it's also worth pointing out that even if that one legal interpretation was not necessarily correct. Right. So he's talking about the interpretation of the authorities under the Export Control Reform Act. We just published a huge paper on the nature of export control legal authorities, not just here in the United States, but around the world. But in the case of the United States, you could also access export control authorities through other legal means, such as iepa, the International Emergency Economic Powers Act. So the, the, as the Trump administration has demonstrated with their actions on tariffs, they have an exceptionally broad view of the legal authorities here idea that GPUs are not dual use enough for them to have the authority to regulate them on export controls. Which isn't precisely what David Sack says, but which is sort of what he's gesturing at in that quote. I, I have a tough time thinking that that's actually where the Commerce Department is going to land in their interpretation of their own authorities. The next was really about just whether or not this was a desirable state of affairs. The problem with the bureaucratic allocation of compute. So here's another quote. Quote, the rule imposed detailed numerical caps on how many chips and how much computing power foreign entities could acquire and operate. This was a radical departure from market based allocation principles, placing the U.S. government in the position of rationing compute power globally. It effectively turned Washington into a central planner for the global AI industry. I mean, number one, centralizing power as it relates to trading kind of stuff. Well, that's kind of what the tariffs policy is about. They Want, you know, more planning of what trade policy should look like in the United States. Caps on chips and attacking the caps on chips. While there were certain caps under certain conditions, they didn't apply in all cases. What I mean by this is, for example, if you were a country in, you know, Eastern Europe, like Latvia, right, you're a tier two country. Well, you can buy 50,000 chips with no additional work or you can buy an unlimited number of chips as long as you get the companies who are going to be using those chips certified as national verified end users or whatnot. So you know, those detailed numerical caps and how they can operate. There's some truth to that. But it's also true that there was like kind of an unlimited option that was available to these folks. But it is true that, I mean Sachs is definitely correct that it put a lot of pressure on the US government to think through these issues and decide who was going to get what chips under what conditions. The main condition that they wanted was security restrictions, which gets kind of to the, to the next Sachs quote that I think is interesting. Quote. Yes, we must take aggressive steps to prevent advanced semiconductors from being illegally diverted. Diverted here is a synonym for smuggled into China. But that goal should not preclude legitimate sales to the rest of the world as long as partners comply with reasonable security conditions. And so in the diffusion rules, you know, theory of the thing, you can sell unlimited chips to the Middle east, you can sell unlimited ships to South America or Africa or Asia or wherever, as long as it's not China or another D5 country. But the main restriction was security conditions. And Sachs here is, you know, saying that whatever the, whatever the end state that we come out on, security conditions are going to have something to do with it. Okay, there's, there's two more things in the Sax post that I, I want to highlight here. The next was on alienation of U.S. allies, quote, unquote quote the rule also strained relationships with key allies by arbitrarily dividing countries into compliance tiers, labeling many friendly nations as second class partners. This kind of regulatory hierarchy undermines trust and risks pushing allies towards non American technology alternatives. Here Sachs is just correct. And I think that what's important here is perception, which is important even if perception does not align with reality. So I just pointed out, right, that there are ways the diffusion rule where countries like the UAE or whomever can get an unlimited number of chips. But it is definitely the case that the Biden administration, which only had a week right before they left office, did not in that week, go on a massive international roadshow in which they did a magnificent job explaining this rule. Basically, they announced the rule and then they had to leave because, duh, like, the White House administration was over. And so what that means is that perception of this rule was really bad. Um, I mean, you have countries like India, which the United States is trying to court for various reasons, one of which is as a hedge against China, because the India China relationship is very strained. And the fact that, you know, India was put in Tier 2, even though there's a way to explain, hey, you're going to be able to get all the chips you want, you just have to go along with these security conditions. It's just true that India didn't see it that way. Right. They felt insulted. And there was a bunch of other countries that felt insulted. And so that alienation of US Allies point is true. I think one could also say that the tariff policy has alienated U.S. allies. But Sachs doesn't get into that. You know, the Trump administration hasn't gotten into that. Okay. The final thing that he talked about that I thought was interesting was the, quote, lack of due process. So, quote, the diffusion rule was issued just five days before the end of the Biden administration without a meaningful public comment or review period. Given its sweeping scope, its retroactive elements, and the global compliance burden it imposed, this rollout was deeply flawed from both a procedural and a practical standpoint. End quote. Well, this is one of the key complaints of US Industry related to not just the diffusion rule, but a lot of the Biden administration's semiconductor export controls. And so it's definitely true that industry doesn't like this. It's definitely true that the government can make mistakes that industry can point out to them. If you have that kind of process, period. The flip side is that all of that time that you're giving to industry, you know, to notice and comment, is also time that you might be giving to China to think about what they want to do to minimize the impact of this. You know, my favorite example of this is the export controls on high bandwidth memory. So on the one hand, you know, you want to talk to industry and say, like, hey, if we were to restrict high bandwidth memory, you know, what would. What would that do to you? You know, would you find that to be devastatingly dumb? Would you find that to be inconvenient? Would you find that to be good? Let's talk about, you know, what are the better and way worse ways we might implement a policy of restricting high bandwidth memory? Memory well, unfortunately that leaked, right? That leaked in, I want to say July of 2024. So by the time that the policy, you know, went through, by December with all of this industry consultation, China had six months of buying all the high bandwidth memory it could possibly, possibly buy. And so that lack of due process concern, on the one hand, that's legitimate on the other hand, some of the strongest opponents of normal due process have been Republicans. Republicans in places like the House Select Committee on the Chinese Communist Party in places like the Senate Foreign Relations Committee where Republicans have sort of said like hey, this, this quote unquote due process for export controls isn't working. And so while obviously you want some kind of mechanism to effectively consult industry, you also need a way to do that where you're not, you know, tipping your hand to China. And so it's interesting to hear Sachs, you know, echoing that kind of a complaint, which is certainly one that Nvidia, the Semiconductor Industry association and have said.
- [16:29]
A
Sure. So taking into account of course that this is all on X. So there's some mix of true administration lines with also David Sacks opinions. I'm hearing a couple of different things here. You're talking about legality, you're talking about the perception of the, whatever they choose to move forward with, keeping in mind security and low bureaucracy and then clarity on procedure, but being able to move quickly and stay ahead of the competition. So in your sort of outline of killing, changing, delaying or keeping being, we're in the delay stage. I'm kind of hearing that maybe you're thinking they're just going to change rather than completely kill.
- [17:06]
B
Yeah, I think this story ends with change, but I do think that change is probably going to be big. I think they're, they're going to want some kind of mechanism whereby they have some degree of confidence that the chips that they sell to the Middle east are not just going to stay there for an evening on their way to China, but they, whatever the mechanism by which they, you know, know, feel that they have confidence that that's going to be the case, it's not going to be the diffusion framework. One thing you know, I'm really curious about is whether this TEARS approach will survive because the TEARS approach has certainly been the most politically controversial in other countries. And I want to point out that like a lot of countries have the public export control policy and then they have like the secret process private export control policy. The United States is remarkably open about its export control decision making process. The policies by which they will evaluate stuff, the criteria, etcetera just compared to other countries, including, like, European countries, for example, a lot more of our stuff, like putting countries into tears is just all out there for anyone to see. So you can have secret tiers. Other countries do have secret tiers that they don't talk about. But, yeah, I think the public version of the tiers, there's just, it seems pretty likely to me that that's going to go away. That's going to win you a lot of goodwill with industry. That's going to win you a lot of goodwill with allied countries and their leaders. But, you know, whether or not these security restrictions remain in place and the conditions under which those security restrictions apply, I think something like that is probably likely to stick around. You know, that said, this is the Trump administration, and if you talk to anybody who tells you they know with certainty what Trump administration is going to do, yeah, they're probably wrong. Right. There's, there's a great deal of volatility. The Trump administration likes to project volatility because they like to keep everyone else guessing and, and having their attention focused on them. So while I, I feel like what has just happened has vindicated my earlier predictions, that doesn't give me a ton of confidence that my next round of predictions are going to be right. I, I think one thing that's just out there that's worth commenting on here is On April 29, Reuters reported that an unnamed senior Trump official wanted to replace the AI diffusion rule with a policy that is, quote, stronger but simpler. So what exactly that means? Who knows who that person represents, you know, in the interagency fight that we've been talking about, who knows? But I think that that probably does reflect, like, what they're, what they're thinking about here. Like, what is the problem to be.
- [19:51]
A
Solved that sort of mimics the way that they think about a lot of policies coming out of this administration. So I could imagine any number of officials saying stronger but simpler. Okay, well, let's move on to the AI executives that testified before Congress. So on May 8, CEOs and top executives from OpenAI, AMD, Cora Weave and Microsoft testified before the Senate Committee on Commerce, Science and Transportation. I want to talk about their messages to the Senate. It so what were some of the priority policies for the expert controls and how did they react to the repeal of this diffusion framework?
- [20:30]
B
Yeah, so the way this happened. Right. Is the committee hearing happened after it was clear that the diffusion rule was going to be repealed. So you got a really rapid reaction. Now, one thing I should point out, and this is not me knocking any of these executives. I just think it's worth for everybody to sort of keep in the back of their mind, you know, corporate America usually does not like to project strong opposition to any move that the, the White House is making. You know, usually you want to keep your disagreements private and your agreements public. Right? Because that's how you kind of build goodwill. You know, if you just think about policies like dei. A lot of companies were for that when it seemed like the Biden administration was for that. A lot of companies are against that when it seems like the Trump administration is against that. So anytime you're evaluating, you know, any executive statements, you got to kind of ask yourself, you know, are they showing the side of themselves? Are they showing the side of their beliefs. Right. That agree with the administration? Because that's what they want to project. Right. Is that they're being a good partner to this administration. So with that, you know, minor caveat, which doesn't mean that anyone was lying, doesn't mean anybody was wrong. Just, you know, sort of a mental model for evaluating private sector statements when they're testifying in this kind of thing. Sam Altman, the CEO of OpenAI, he said, quote, I was glad to see he was talking about the diffusion rule rescinded. I agree there will need to be some constraints, but I think if the sort of mental model is winning diffusion instead of stopping diffusion, that directionally seems right, but that doesn't mean there's no guardrails. So that was, I thought, a really interesting thing. Previously, you know, other companies like, like Anthropic, had been loudly in favor of the diffusion rule. And OpenAI had not expressed a strong opinion on the diffusion rule specifically, but had expressed public support for export controls of AI chips as a means of affecting the US China AI competition. Brad Smith, the Microsoft president, he had what I thought was actually a pretty interesting argument, which I think is the strongest argument against the diffusion rule. And, and so I think it's worth quoting at length, quote, the number one factor that will define whether the United States or China wins this race is whose technology is most broadly adopted in the rest of the world. 18% of the people in the world live in China, 4% live in the United States, 78% live somewhere else. The lesson from Huawei and 5G is whoever gets there first will be difficult to supplant. We need to export with the right kind of controls. We need to win the trust of the rest of the world. And again, that's that. That comes back to right the trust of the rest of the world getting stuff adopted everywhere else that, I mean, Microsoft has really been on the receiving end of all these countries who are frustrated with the diffusion rule. Right. Because Microsoft was almost certain to be certified as a universal verified end user, which means that countries can buy unlimited numbers of chips as long as those chips are going through Microsoft somehow, you know, to data centers in their countries. Same for Google, same for Amazon, et cetera. But the basic point is, right when the governments of the world are complaining about the diffusion rule, one of the people I'm sure they were complaining to was Microsoft. And so they were really on the front lines of receiving that alienation that we talked about earlier in the conversation. And so you can see why Microsoft is kind of like, look, we want diffusion. And again, the word is a little bit confusing here because it's also the name of the rule. But diffusion in terms of diffusing AI technology throughout the world, we want that to take place. We want American leadership to occur through diffusion. And what the diffusion rule tried to do by, you know, controlling the flow of chips, not just to, you know, adversaries like China, but also to allies, just really rubbed everybody the wrong way. And we're looking for some kind of controls that can address some of these security concerns, but do so in a way that's less alienating to our allies.
- [24:36]
A
Sure. He's basically saying that there's 70% of the world that is maybe a little confused and it just makes it harder for them to be able to get the technology there, there. Even if the rule itself was on the track that we need to be on to win this race. Is that a correct?
- [24:52]
B
Yeah, I mean, I think, I think they're not out there saying, hey, let us sell all the chips to China. Right. I think OpenAI anthropic Microsoft, nobody on that score has been advocating for selling chips to China the way Nvidia has at times. Right. Because they sort of say, look, we need to sell to China, China to make sure that, you know, we get that revenue and that we make it hard for Huawei to grow. Microsoft is talking about, you know, blocking China's growth in the rest of the world. But they're not fundamentally, at least not in this claim here from Bad Smith, disagreeing with the overall restrictions towards China specifically.
- [25:32]
A
So a big part of the, the testimony centered around AI governance. So what approach did these executives take when talking about AI governance? Governance, yeah.
- [25:43]
B
So, you know, this was a pretty wide ranging congressional committee. Obviously the diffusion rule came up, which we were Talking about. Another big area of focus was on regulating AI And I think the two kind of things. It was interesting, the folks who were talking about this were mostly talking about what they don't want and being a little bit less clear about what they obviously do want. And I think the two things that came up was, number one one, they don't want the European approach. At the outset of the hearing, Committee Chairman Senator Ted Cruz of Texas issued a warning about, quote, taking the path of Europe. And in response to that, Sam Altman said, quote, I think that would be disastrous. I don't want to live in Europe either.
- [26:30]
A
What does it mean to take the path of Europe? What is the Europe approach in this context?
- [26:34]
B
So I think what they're. What, what. What Europe signifies In the Washington, D.C. policy debate is hard law, strict regulation of AI. So you can think of a soft law approach geared around standards and voluntary compliance. You can think of a no regulation approach, which some people cite the example of the Internet as a mostly unregulated industry in its early days. And then there's like the strict, you know, fill out these 500 forms telling me how you're definitely safe and you're definitely not violating privacy, you know, et cetera, et cetera. Whether or not that's a fair characterization of Europe, that sort of is the mental model that Europe conjures in most members of Congress, right, when they're thinking about what is the approach we want to take here. Now, Brad Smith, Microsoft President, had a really interesting rejoinder to that, which wasn't necessarily praising Europe, but he was kind of talking about the peril of inaction in the face of European action. So, quote, I think of the evolution of privacy law. The United States didn't adopt a national privacy law, Europe did twice. And most American companies of any size today across the United States work that complies, do work that complies with European privacy law. And so he's kind of talking about the Brussels effect here, and he's saying that, look, you know, if we're. If you're a big multinational company or even if you're a small company that, you know, just most tech companies are international, even if they're of a small size, you're affected by European regulations. And so that's really kind of interesting. He's sort of saying, look, whatever the American rebuttal is to the EU AI act, it probably shouldn't be do nothing. You know, we kind of need to do something because otherwise Europe is going to win this argument, not just in their home Market, potentially. And here I'm going farther than what Brad Smith said, but like a lot of African countries literally just copy pasted huge portions of GDPR and made it their own national privacy law. And so I think that's a really worthwhile thing to bring up. The other big focus of the discussion was around state and local AI regulation. So at the same time that, you know, Congress has, I think at this stage, it's fair to say, really given up on the idea of comprehensive AI legislation which was the dream of, you know, Senate, then Senate Majority Leader Chuck Schumer in the summer of 2023. Because there's no federal action, because there's no federal preemption, at least not at this stage. People are talking about a lot of state and local AI regulation. There was a bill that failed here in Virginia, there was a bill that failed in California, but there's other bills that are passing and there's a lot of bills that are being proposed. I mean there's literally hundreds of bills that have been proposed throughout America's state legislature. So Sam Altman is very concerned about this. He said, quote, it is very difficult to imagine us fig out how to comply with 50 different sets of regulations. One federal framework that is light touch that we can understand and it lets us move at the speed that this moment calls for. Seems important and fine. So there, there's a kind of a, basically saying, as I was, you know, sort of gesturing at with the Brad Smith quote a moment ago, do nothing is actually not helpful. So you know, just having something that takes place at the federal stage that might get in the way or persuade states to not go their own way with these kind regulations seems likely to be important. And I think that's, that's interesting that the companies are calling for this, right? Oh yeah. Well, go ahead.
- [30:18]
A
Well, what I'm hearing is that if you're on the spectrum between how we approach Internet regulation of Internet or, or not, and then the European approach to privacy, we're not quite all the way to the Internet side of things with what they're saying, but we're definitely not directly in the middle. Is that a fair character?
- [30:36]
B
Yeah, I think, I think we don't necessarily know what the companies want to want. If they could wave a magic wand and get everything they want. Right. It's possible that if you asked Sam Altman the magic wand question, he would say, if I had a magic wand, I would wish the EU would get rid of all their regulations and I wish the states would get rid of all their Regulations. Right. Maybe that's his truest policy preference. We don't know. But given that the EU exists and has taken these measures, and given that the states exist and are likely to take some degree of measures, their saying the federal government should do something, it should be enough to be a compelling proposal that those who are worried about safety concerns in AI, those who are worried about privacy concerns in AI, etc. Etc. They can at least point to this federal law as reason to justify, you know, not going the more extreme route. So you can sort of think of it as a negotiation. Right. None as a regulatory strategy is very unlikely to sort of win the negotiation. And so the question is, can some regulation win the negotiation as an alternative to what is viewed as extreme regulation in the European Union?
- [31:51]
A
Sure. Okay. Did they mention anything about permitting reform? I've heard a little bit of whispers about that, yeah.
- [31:58]
B
So this of course is a, is a big deal. And I think the executives are obviously convinced that energy is going to be important for the future of AI. That's a very consistent message. It's very well backed up by the data. And here I think Sam Altman said something really interesting, which is eventually the cost of intelligence, the cost of AI will converge to the cost of energy and it'll be how much you can have. The abundance of it will be limited by the abundance of energy. So in terms of long term strategic investments for the US to make, I can't think of anything more important than energy chips and all the other infrastructure. That was really interesting. It was another point that also the CEO of Cor Weave mentioned made. And I, I think that's just a very consistent message coming out of just about everybody in the AI industry right now.
- [32:49]
A
Okay, so looking forward, what did we learn about the upcoming AI legislation from this hearing?
- [32:55]
B
Okay. On May 11, the House Energy and Commerce Committee published a memo from the majority staff that included language that might halt all state level enforcement related to AI systems. So this is a pretty interesting thing if it was to become law. And it's also pretty interesting just as like a negotiating position of this, this faction within Congress. So in part two on subsection C, it states, quote, no state or political subdivision may enforce any law or regulation regulating artificial intelligence models, artificial intelligence systems or automated decision Systems during the 10 year period beginning on the date of enactment of this act. So that's a draft, that's a proposal out there. I think it's pretty unlikely that that text, specifically in an unmodified form, would become law in the United States. But it does sort of get to like what some of the opinions are out there about the challenges of state and local AI regulation. So if you've got, you know, Sam Altman saying we definitely don't want 50 different laws, one way to do that is just to ban AI regulation. I don't think, as I said before, before, I don't think that specific proposal is very likely to become law. But it does sort of point out, right, that there's the extremes. There's the EU strict regulation, there's the banning regulation of any kind, and then there's the, you know, some kind of modest initiative aimed at containing the worst risks and obviously the most, you know, the, the risks about which there is the greatest degree of political consensus. Stuff like child sexual abuse, material stuff like non consensual deep fakes, stuff like, you know, political propaganda type work using disinformation enabled by AI, those kinds of things, you know, you can imagine a hodgepodge of rules against and then you can imagine, you know, some kind of clarification on what's going on around liability. Maybe, maybe this Congress can, can find a way to pass that kind of a law as a means of preempting state action. And, and that, that fact of preempting state action is what I think is one of the strongest arguments in terms of pushing towards some kind of bipartisan consensus, almost certainly what would be required to pass this kind of a law.
- [35:13]
A
All right, well, let's move over to the House side of things and talk about Secretary Bessen's testimony. So on May 6, the Secretary of the Treasury, Scott Bessant, testified before the House Appropriations Subcommittee on Financial Services and General Government. During the hearing, he laid out an ambitious vision for what AI could do for the government capacity. So what was this vision?
- [35:34]
B
Yeah, so I don't want to talk about everything that Secretary Besant said, but he did say one thing that I thought was really worth, worth dwelling upon here. So, you know, the, the, the Trump administration's spending proposal included, this is for, you know, fiscal year 2026 included a budget cut of almost $2.5 billion for the IRS, you know, which is here in America is the, the tax enforcement guys. And he said, you know, when he was asked by some skeptical Democrats, you know, what this was going to do to the efficacy of the irs, he said, quote, so I believe through smarter it, through this AI boom that we can use to enhan and I would expect that collections would continue to be very robust as they were this year. He added that he did not expect any, quote, operational disruptions. So he's basically saying, look, I'm laying off a ton of people in the irs. This includes people who go after tax cheats, who go after tax fraud. But while you might think that that would decrease the revenue collected, the government revenue collected by the irs, it's not going to, because we're going to adopt a AI, so we're going to have much less cost, but we're going to have the same or greater productivity. And this, you know, follows off, like, big layoffs that were related to Doge and other things going on here. And I just have to say to Secretary Besant that it's not that that's a crazy idea, right? It's not that it's a crazy idea that you could use AI to dramatically increase the productivity of your workforce. And specifically related to collections, you know, just to give you an examp, right? Like Visa or MasterCard. These are the, you know, the big credit card companies here in the United States. They have talked about time and again how AI has been so critical to their identifying financial fraud, credit card fraud, for the most part, and most of the major banks will tell you the exact same thing, is that, you know, as digital commerce and all these other things have gone on, the, you know, the. The risks of fraud have gone way up. And if they didn't have AI, AI, they would have been, you know, completely overtaken by this tsunami of attempts at fraud. So it's super reasonable for Secretary Besant to say, you know, we want to use AI to increase the quality of our collections and do so in a way that, you know, increases productivity so that we don't have to double the number of IRS agents to double the efficacy of the IRS at collections. But I also want to just offer a word of caution here, because the specific political catastrophe that got the snowball rolling down the hill until it became, you know, like that rock in Indiana Jones that was the giant boulder, you know, chasing him down the alley. That first snowball was an attempt to use AI in tax collections. So this goes back to 2019, when the Dutch tax authorities, hearing a very similar spiel as the one I just gave About Visa and MasterCard, they, you know, brought on AI and a company, you know, to build AI for them to go after tax fraud. And according to Politico coverage of this incident, quote, a parliamentary report into the child care benefits scandal found several grave shortcomings, including institutional biases and authorities hiding information or misleading the parliament about the facts. Once the full scale of the Scandal came to light. Prime Minister Mark Ruta's government resigned. I mean, what. What happened in this scandal, which I've talked about before on the podcast, but it's worth just restating is they had a bunch of false positives of identifying fraud, and they told people, hey, you claim that these are your children and that therefore you're enabled, you're entitled to a child tax credit. But according to our incredible AI system, those children aren't real. You just made them up. There's no real such children. So you owe us all that money back, plus you owe us penalties. And what this resulted in was, you know, family separation. People basically said, hey, I. I'm abandoning my child to foster care because they can either, you know, starve with me because I have to pay all this money, or I'm going to have to go to jail, or they can go into the foster care system, you know, which is not an admirable decision, but the point is, it's one they made. And the Dutch government, you know, said that there was hundreds of children in the Dutch foster care system that they, they can't, you know, undo. They can't go back and say, oh, it was an AI mistake, because they, you know, didn't have the digital forensic trail to know which of these false positives in identifying tax fraud using AI were false positives. And so that, that was a political catastrophe for the. The Dutch government, and it's one that brought down their prime minister. Literally. The government collapsed because this was such a devastating political scandal. And so my, my word of caution, caution to the Department of Treasury is not don't adopt AI. I think there's a very strong argument to adopt AI. It's for the same reason that Visa and MasterCard did. But I want to make two points here. Number one is don't assume that this is just like the basketball equivalent of a layup, right? It's not like you just. You just staple AI to it, and suddenly everything gets ten times more productive. You really have to think through, like, what is an appropriate use case, case for what kind of AI Are the people who are building it for us? Are they building it with the right safeguards in place that we can have confidence that it's going to perform as intended over the full, you know, expected spectrum of activities? And one thing that's harder in the government context than it is in The Visa and MasterCard context is a lot of the way government works is tragically right, tied up in law or old policies. And when you go digital or when you go AI, a lot of times you need, you know, the process re engineering work associated with that for that to make sense. So to just give an example, I used to be in the Department of Defense. There was this outfit in the Department of Defense that had created something called the Defense Travel System. So this is like how you book flights when you work for the dod. And when I was in the door, there was literally a button, you know, that you had to click that said like stamp the flight approved, right? And you're wondering like, what in the world, like why is there a button that says stamp? And it's because like when the regulations were written about like flights, they were written in like the rubber stamp and ink era. And so when the digital people were said like, hey, we want to digitize all of this. There were told like, hey, we want you to digitize all of this. But also it needs to comply with these old stupid laws and regulations because we have the authority to give you money to digitize it, but we don't have authority to overturn those old stupid laws and policies that were written for another era. That's how you end up with like digital buttons that say stamp. And so what I'm getting at here is when you talk about, you know, introducing AI into your work stream or introducing just any digital technology into your work stream, you have to think about the process re engineering component. And that's hard enough when you're a company like, like Visa and MasterCard. But usually, you know, you have the relevant authority. Sometimes when you're the irs, the relevant authority is like a law that was written 30 years ago. And when you call Congress and say like, please change this law so that we can adopt our process workflow, you know, for an era of AI, they might say, yeah, we already kind of done passing laws this year, so call us next year. In the meantime, like thou shalt comply with the law. So that's like a really frustrating challenge, but it's like a very familiar one to the government. So those are kind of my two words of caution for the IRS is I like in high level terms endorse embracing AI. But number one, understand that tax collection is kind of a mission critical job. And then number two, understand that adopting AI is not necessarily as effortless as it is for you and me as citizens to log into chat GBT and have some fun with AI or maybe even get some productivity out of it. When you're thinking about introducing it into a big, big important workflow, there's a lot of hardware work to Be done.
- [43:50]
A
Yeah, that makes sense. And it's, and it's a very fair cautionary tale. I just wonder if his remarks were consistent with the broader Trump administration's approach to the role of AI in government, or if there is an intent to bring in more AI throughout government processes.
- [44:05]
B
Yeah, great, great question. So back In April, on April 3, the Director of the White House Office of Management Budget, Russell Vaught, he issued two memos that we talked about on a podcast a while back, all about accelerating the adoption of AI, in which I'm for, You know, I want it to be done. I just want it to be done by people who know what they're dealing with, who know what the AI failure modes are, who know what it means to actually effectively, you know, re engineer processes in agencies that are, you know, a lot of people in this country are counting on. So I think it's completely consistent.
- [44:41]
A
All right, well, let's bump over to the U.S. copyright Office. So on May 9, the U.S. copyright Office published a draft version of its report addressing the use of copyrighted materials in generative AI training. This report, report was 113 pages long. We don't have time to go through the entire report. I'm wondering if you can summarize the core argument for us.
- [45:00]
B
Yeah. So I think the big thing here is it's getting to the meat of a lot of the debate about what constitutes, you know, fair use in copyright. So, you know, there's a lot of things like when we at csis, when we want to use a photo that comes from like one of the big journalistic photography outlets like Getty Images or the Associated Press or whatever, you know, we pay for that. We pay to put that photo on our website because that's what copyright law requires of us. And however, you know, if a comedian wants to include a photo, you know, for the purpose of satire, that might be constituting of fair use, so they don't necessarily have to pay for that, you know, under certain kinds of copyright law. And so the question is fair use, what does that mean when it comes to AI and specifically training AI stuff? And the Copyright Office report did not say that under no circumstances could the use of copyrighted data for training AI systems be fair use. They didn't, they didn't say that this is always wrong, but what they said is it's definitely not always right. Right. So what a company like OpenAI would love to see is for the, for the lawsuit to be tossed out and there's a lot of lawsuits going on right now. I mean, the New York and others have sued Open AI, Dow Jones and company, which is owned by Rupert Murdoch, who also, you know, owns a bunch of other stuff. They sued Perplexity AI. I mean, there's just a bunch of these lawsuits out there from journalist institutions, from like media companies, et cetera, saying that, hey, all of our copyrighted stuff was used for your training data set and therefore you have violated our copyright. And what the AI companies would love is for that, that to just be counted as fair use, right? Like they're not going to produce Juice, you know, a full length Disney movie, right? But like the fact that Disney is in their training data set should not in and of itself be a copyrighted violation as long as the outputs are not a copyright violation. And this report said that, look, maybe it is fair use, maybe it's not fair use, but it is certainly not the case that it is obviously not fair use such that a judge doesn't even need to hear the lawsuit. You know, that was one of the definitive outcomes of this report. At least it seemed definitive until a couple days later the head of that office was terminated by Donald Trump. And so that this whole report that we've been talking about, and that was the pre publication draft, so that was, you know, out there for, for comment or whatever, it was not the law of the land. And it certainly seems based on that decision that whatever the law of the land is, it's going to be closer to this interpret that the AI companies are interested in around fair use. Now, I'm, I'm speculating here. You know, none of that is, is written in stone, but I think that's a pretty reasonable interpretation of the sequence of events. And it's worth pointing out that, you know, this has stirred up a little bit of political controversy. So, you know, Representative Joe Morell, for example, on May 10th stated, quote, Donald Trump's termination of the Register of Copyrights, Shira Perlmutter, is a brazen, unprecedented power grab with no legal basis. It is surely no co coincidence he acted less than a day after she refused to rubber stamp Elon Musk's efforts to mine troves of copyrighted works to train AI models, end quote. So that's a very fiery quote, you know, coming out of this member of Congress. I don't have any inside information as to whether any of those accusations are true or untrue, but I do think that that sort of high level thing, that this was an interpretation of copyrights that Donald Trump did not agree with or that the Trump administration More, more broadly, did not agree with. I think that's pretty likely to. And, and whatever we see coming out of the. The Copyright Office, I would expect it to not echo the conclusions of that long, long report.
- [49:07]
A
Yeah, seems like a fair conclusion. We'll definitely keep an eye on the evolution of fair use as, as the conversation continues, I do want to finish up with an area of AI policy that we have not traditionally covered here at the Wadhwani AI Center. The news that dominated all last week is that there's a new Pope. And so on May 10, the newly appointed pope, Leo XIV, laid out the priorities for his papacy. How did AI in those statements?
- [49:34]
B
I, I agree with you. Not something we would normally talk about here, although Pope Francis also had, you know, AI as a part of his agenda. But what Pope Leo has done here, I think is really a pretty interesting statement. So here I'm going to quote something kind of long. So from the pope, quote, I chose to take the name Leo xiv. There are different reasons for this, but mainly because Pope Leo xiii, in his historic encyclical Rerum Novarum, addressed the social question in the context of the first great industrial revolution. In our own day, the Church offers to everyone the treasury of her social teaching in response to another industrial revolution and to developments in the field of artificial intelligence that pose new challenges for the defense of human dignity, justice and labor. So what I think is so interesting there is, is, you know, he's saying that he picked the name of his papacy because of how significant he thinks artificial intelligence is likely to be during his papacy. And you can really think about that right? Right now. Human beings are the only things on Earth that are as intelligent as human beings. And you can imagine that at the end of his papacy that will not be true, that there will be machines that there are as, that are as intelligent as human human beings. He. Pope Leo is 69 years old. Pope Francis died at 88. The typical pope of the post 2000 era is living into their 80s. So if we think that he's going to continue to be the Pope in 2040 or 2045, as I'm sure is on his mind as to what's going to happen over the next 15 to 20 years that he is likely to serve as pope, he thinks AI might be the single most important thing that is happening over those next 15 to 20 years. And he thinks it's of spiritual significance, not just worldly significance. And I just thought that was too remarkable to go without commenting on it at the AI Policy podcast.
- [51:40]
A
Yeah, I just finished watching Conclave over the weekend, so.
- [51:43]
B
So there's a good line in that movie where it's Stanley Tucci, I think, who's saying to the main character Rob Ralph finds and he says every cardinal has thought in their head what would be the name of their papacy if they were selected as pope. And I thought to hear that quote, which to me really cemented how seriously they take this naming convention and then to find out that this one was named because of how significant he thinks AI is going to be, I thought that was really poignant.
- [52:16]
A
Oh yeah, endlessly interesting and definitely a great note to end today's podcast on Greg, thank you so much for joining me. For all of our listeners, you can check out all of this our analysissis.org I think we're going to link the piece about Allied export control authorities in the show notes for this conversation. So if you were curious about our first topic of the day, please feel free to check out that report, share it amongst your friends. Give us five stars. And with that, I think we'll just catch you next time on the AI Policy Podcast. Thank you so much for listening.
- [52:50]
B
Thanks for listening to this week's episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform and subscribe and tell your friends. It really helps spread the word. This podcast was produced by Sarah Baker, Isaac Goldston and Sadie McCollum. See you next time.