ROMAN MARS: This is the 99% Invisible Breakdown of the Constitution. I’m Roman Mars.
ELIZABETH JOH: And I’m Elizabeth Joh.
ROMAN MARS: Today we are discussing Articles VI and VII.
ELIZABETH JOH: Roman, why don’t we go through both articles? And let’s save the most important part for last.
ROMAN MARS: Okay, because there’s a lot of unimportant parts.
ELIZABETH JOH: [CHUCKLING] Less important.
ROMAN MARS: Go for it.
ELIZABETH JOH: So, why don’t we start with Article VII? That’s the ratification clause–maybe the least important to talk about today, but actually crucially important for the Constitution itself.
ROMAN MARS: Yeah. Yeah, sure.
ELIZABETH JOH: We needed these states to vote on it. And the ratification clause says that nine states would be enough to ratify or make the Constitution itself a legitimate document. So that in fact happened. It comes into effect on June 21st, 1788, when New Hampshire became the ninth state of the 13 to ratify the Constitution. So it kind of has its own clause to make sure the document is legit.
ROMAN MARS: Got it. Yeah, that makes sense. Okay.
ELIZABETH JOH: Yeah, so that’s pretty clear. No real important Supreme Court cases on it. So let’s turn back to Article VI. Article VI is a mishmash of a bunch of different things. So Article VI, Clause 1 talks about debts that the United States is already obligated for and still has to pay.
ROMAN MARS: So, why is that there?
ELIZABETH JOH: Well, you know, when the Constitution was drafted, the country still had debts and engagements that were left over from the Revolutionary War. And creditors were kind of nervous. What if you create a constitution that wipes out all of the debt? That would be pretty convenient.
ROMAN MARS: Got it. Got it.
ELIZABETH JOH: So in order to assuage those nervous creditors, this constitution–our constitution–actually says, “Don’t worry, we understand that we have these debts and we are going to pay these debts.” So today it’s really mostly a historical interest. It doesn’t come up because we did in fact pay our debts. I will say what is interesting is that Clause 1, originally as it was drafted, said that the United States would both be obligated to pay the debts and would have the power to pay the debts. But that second part got taken out of Article VI and put into Article I as part of Congress’ spending authority. So that very, very important part today is actually in the larger chunk of the Constitution we cite all the time, which is why Congress has the ability to pass laws. And very often it’s because of spending authority.
ROMAN MARS: Wow. Well, that’s really fascinating, actually.
ELIZABETH JOH: Yeah, so, you know, one little switch around that exists.
ROMAN MARS: Yeah, it made a huge difference. Okay. Okay, so that’s the first clause.
ELIZABETH JOH: Okay, let’s turn to Clause 3 of Article VI. Do you want to read it?
ROMAN MARS: Okay. “The Senators and Representatives before mentioned, and the Members of the several State Legislatures, and all executive and judicial Officers, both of the United States and of the several States, shall be bound by Oath or Affirmation, to support this Constitution; but no religious Test shall ever be required as a Qualification to any Office or public Trust under the United States.”
ELIZABETH JOH: So this is known as the No Religious Test Clause.
ROMAN MARS: Great! I like this clause.
ELIZABETH JOH: Exactly. It’s kind of a, you know, no religious test for anybody taking office. And in fact, it’s the absence of religious tests that makes us understand that this is successful, right? You don’t require anybody to take a religious test–sort of–at least formally, right?
ROMAN MARS: I mean, there’s no formal test. But you can kind of feel it in there, the fact that the representation of other religious faiths is not super common inside of our public institutions.
ELIZABETH JOH: True. True. But there is a big difference when you were formally required to do it. And in fact, this clause comes from traditions going back to England. So, in England in the 17th century, for example, all government officials had to take an oath that they would help establish the Church of England–also disclaim Catholicism and the Pope. And so the idea is we have this common law tradition. It comes from England. By the time you have the colonies and the Articles of Confederation, it was pretty common for government officials to be told that they had to take some kind of religious affirmation, of course not for the Church of England, but some kind of “I believe in God” sort of test.
ROMAN MARS: It’s notable that it’s absent.
ELIZABETH JOH: It’s notable for its absence. And this clause as well has very little in terms of Supreme Court interest or case law today. And that’s for a totally different reason. You’ll notice this is about religious freedom, essentially, right? It shouldn’t matter whether you are a practicing Catholic or a Muslim or a Jew to be able to take a public office. But the reason why this clause doesn’t get much attention is because free exercise clause cases today come up under the first amendment rather than this clause. So, not too much there as well.
ROMAN MARS: Yeah. So I noticed, in our recap, you had Article VI, Clause 1, and Article VI, Clause 3. But we have skipped Article V, Clause 2. So what is that?
ELIZABETH JOH: Well, Article VI, Clause 2, contains what’s called the Supremacy Clause. Why don’t you read it?
ROMAN MARS: Okay. “This Constitution, and the Laws of the United States which shall be made in Pursuance thereof; and all Treaties made, or which shall be made, under the Authority of the United States, shall be the supreme Law of the Land.”
ELIZABETH JOH: And that’s referred to as the Supremacy Clause.
ROMAN MARS: So why is the Supremacy Clause so important?
ELIZABETH JOH: Well, historically, the Supremacy Clause responds to a very particular problem. And that is, before the federal constitution, the Articles of Confederation, which was the predecessor document, had no similar provision saying that federal law is supreme. And you might wonder, “Well, what does that really mean?” Well, think of it this way. If you have state laws on a topic and federal laws on the exact same topic, which one are you supposed to follow? If there’s no clear instruction, well, maybe you just follow whichever one you want. And that’s kind of what happened. Before the Constitution, state courts sometimes just didn’t think that federal law was binding, so they didn’t apply it. They applied state law. That’s kind of a problem, right? So, the Supremacy Clause, just with one fell swoop with this particular clause, gets rid of that uncertainty or ambiguity. The Supremacy Clause simply says, “Look, federal law, whether we mean the Constitution, federal statutes, federal treaties, are supreme when it comes to any conflicting state law.” So, the idea here is that you have this very important structural part of the Constitution that federal law is supreme.
ROMAN MARS: So what does that mean practically speaking?
ELIZABETH JOH: Well, what that means is you can think of supremacy as stating the simple fact that federal law is supreme, but arising out of supremacy is the idea that Congress now has the power when it legislates to preempt–that really means displace or override–any contrary state or local law. So you can think of preemption as being based in the constitutional power of supremacy. So, Congress doesn’t have to exercise preemption. But when it does pass laws in this way, it’s very clear that any directly conflicting state or local law has to give way. So that’s kind of the genius or the simplicity of the Supremacy Clause. But that’s the most simple part of the Supremacy Clause.
ROMAN MARS: And I take it there’s lots of constitutional case law based on the Supremacy Clause.
ELIZABETH JOH: That’s right, because things can never be simple, right?
ROMAN MARS: [CHUCKLING] Yeah, yeah.
ELIZABETH JOH: So, when you think about federal law, sometimes Congress can simply say, “We’re going to pass a law. And this law will, in the text of the law itself, displace or preempt any similar state law.” That’s pretty easy. And if that were the only issue, we’d never talk about preemption, right? But the problem is that Congress very often doesn’t say. There may be a federal law on a topic and a state law on topic, and the federal law doesn’t say anything. So, in response, the Supreme Court has come up with a whole host of cases, doctrines, tests, and ways of thinking about federal preemption to try and answer the question, “What happens when it seems like there are federal and state laws legislating on the same topic?”
ROMAN MARS: So what exactly is supposed to happen when there’s a conflict?
ELIZABETH JOH: Well, that also is a complicated answer. So, it depends on what we’re talking about. Sometimes courts will say something like, “You know, there are some areas of federal law where the federal interest is so important–so extreme–we don’t want the states to get involved a little tiny bit, even if Congress hasn’t specifically spoken to that area.” So an interest like this would be foreign policy. We don’t want the states to get involved with foreign policy, negotiating their own treaties. That would be a bad idea.
ROMAN MARS: Agreed. That would be a bad idea.
ELIZABETH JOH: Exactly. So those are the easy cases. But the much more frequent and difficult cases are sometimes courts have to answer… Well, there’s a federal law on a topic and a state law on the topic. Is it possible to comply with both state and federal law? If it’s possible, maybe there is no preemption. No preemption would mean that state law and federal are both valid. But, for instance, if there is a way in which the state law is an obstacle for the federal government’s law to operate or whether it’s literally impossible–state law says black and federal law says white, you can’t do both at the same time–then that’s a case of federal preemption. So these are always case-by-case determinations. But preemption is actually really important because, if you think about all of the different areas in which federal government regulates–everything from the environment, consumer protection, energy, you name it–the states also often legislate in the same areas. And what you will have are individuals or companies that say, “Well, I want to comply with one. I don’t want to comply with both. Am I supposed to comply with both?” And that gives rise to preemption. So of all of the areas of law that we’ve talked about with the Constitution, in fact, preemption is probably the most frequently used constitutional law in practice. So on the one hand, you can think of constitutional law in the courts as being on a spectrum, right? Like, maybe we’d put impeachment at one end–we don’t talk about it in the courts–and then preemption all the way at the other. Preemption comes up all the time because the idea of federal preemption is that it’s a possible question anytime the federal government is regulating in a particular area.
ROMAN MARS: Right. Right. Which could be infinite almost.
ELIZABETH JOH: Almost infinite. That’s right. Every single area of modern life where the states regulate, very often–though not always, of course–the federal government is also regulating.
ROMAN MARS: And this situation is exacerbated by the fact that modern life continues to go on. Like, there’s new laws coming up all the time because there’s new technology all the time and there’s new things all the time to consider.
ELIZABETH JOH: That’s right. So whenever you have a new policy problem–a new change in society–there’s a race to regulate it or at least a cause to regulate that new development in modern life. So the question is: Are states gonna do that job? Should the federal government do that job? Or should they both do that job? So, one way to think about the problem of preemption is for us to pick an emerging area where both the states and the federal governments are trying to regulate at the same time. And I think there’s no better topic than artificial intelligence.
ROMAN MARS: Totally. Totally. I mean, that’s, like, huge. I don’t even know what I think about it, so I can’t even imagine what states and the federal government are thinking about it at this point.
ELIZABETH JOH: That’s right. Artificial intelligence is everywhere. It’s at the doctors. It’s at the store. It’s at school. It’s at work. It’s kind of a huge problem for government and that’s because AI has the potential to produce these really big benefits for society, but we’ve already seen that it can have all kinds of harmful effects. It can produce all kinds of major risks for society. You know, everyone’s heard of AI makes up facts that don’t exist that people believe and sometimes act upon. Or it can make decisions about people that are really hard for us to explain, and sometimes those decisions are false or misleading. So, just like any other problem in society, the states and the federal government are trying to figure out how to regulate AI or AI systems. And that means everything from how do you regulate a chatbot that teenagers use or self-driving taxis or how do you regulate autonomous weapons when it comes to wartime.
ROMAN MARS: Oh my god.
ELIZABETH JOH: And so what kind of level of government should be regulating AI? And so should the states get out of the way altogether?
Now, this seems like a very current topic, and it is. But the larger picture is an old one and that’s a question of federalism. So the narrower of view we have of preemption, we’re really allowing the states to engage in more experimentation for the states to say, “Hey, we want to try this approach.” And California will always take an approach that probably Texas will not, and vice versa. But a very broad view of preemption really is saying, “You know what? We want the states to just get the heck out of the way. We want the federal government to be the primary voice in this area.” So those are choices that courts have to make. There’s nothing obvious about going in one direction or another.
ROMAN MARS: Yeah. Yeah.
ELIZABETH JOH: Because this is a fast-moving and complex topic, our guest for this episode is Dr. Alondra Nelson. She’s a scholar of technology and social science and a leading expert on artificial intelligence. She currently holds the Harold F. Linder Chair at the Institute for Advanced Study in Princeton. She also served in the Biden administration as the acting director of the White House Office of Science and Technology Policy. It was in that role that Dr. Nelson spearheaded what’s called the “blueprint for an AI bill of rights.” We invited her to help us navigate why it’s a challenge to regulate and what to think of the tug of war between the states and the federal government on the topic, especially during the second Trump administration. But we start with Alondra’s definition of what exactly AI is.
ALONDRA NELSON: So, you know, I usually use a modified version of the OECD definition, which is a definition that 38 nation states have agreed upon. And it’s basically that these are machine-based systems–lots of statistics and lots of math–and that they make inferences from different inputs, and they generate outputs. And so the outputs are things like, you know, so-called “predictions.” They are things like recommendations, like your Spotify music recommendations or your Netflix recommendations. And I like to use those two examples because people have different feelings about how good or bad they think their Netflix stream and Spotify is. And I think that’s kind of a level set for AI decisions. So there are machines that are about targeting people, locations, and the theater war. And of course, with generative AI, AI tools and systems generate content, so texts and images and sound. So that’s kind of, you know, inferences made from different sets of inputs, almost all a sort of data, whether those are photographs or numeric data or “all of the internet” that was taken into generative AI and lots of different outputs.
So you cross cut that with the fact that AI systems have different levels of autonomy and adaptiveness after they’re deployed. So some can be very static. A decision-making or predictive algorithm that might be used in the criminal legal system is taking in data, and it has a sort of hardwired dataset that it’s sort of making so-called predictions against. And obviously, today, we increasingly are being told about things like OpenClaw and AI agents. And so these are more autonomous kind of AI systems that are making purchasing decisions for people coding for them and the like. So that’s a broad definition on purpose because AI is really broad. And I think we go back and forth from using generative AI as the default for what we mean by AI, but it’s this whole suite of things. And if you talk to a computer scientist or an AI or machine learning engineer, they would say to you that, actually, if you think about the world of AI as sort of a set of Russian nesting dolls, generative AI is actually the smallest, right? You’ve got deep learning, you’ve got machine learning, and all of that. So, because generative AI with things like chatbots have been made consumer-facing tools–and that’s really how AI came into the public sphere–that’s kind of how we think about AI. But there’s a lot of other use cases and types and autonomous and more brittle, etc., besides…
ELIZABETH JOH: Yeah, so I think, when you hear this, it’s like a pretty technical set of definitions and products. But I suppose if you’re listening to this conversation, maybe someone might think, “Well, I’m sort of familiar with maybe ChatGPT. That came out in 2022. I’ve used it a couple of times. But I really want to know why I should care about this.” So, what for you are some of the most transformative or really concerning examples of AI that are happening in American society right now?
ALONDRA NELSON: So the “why should I care” is, I think, people every day, particularly folks in companies, oversell AI. So that’s certainly true. So what might be transformational? Some of the claims–the AI for good claim–are true and, I think, are either happening or on the horizon. So you can think about in the medical space–like an AI system reading chest X-rays or being able to flag an early stage kind of cancer diagnosis, being able to see a tumor in its very early stages–that’s transformative and indeed, if we get that right, life-saving. It is the case that we still need radiologists and we don’t have enough of them. So transformational, but transformational potentially in the intersection of humans working with the AI, right? So, other cases certainly are like in agriculture. So farmers, whether it’s Sub-Saharan Africa or Kansas in the United States, are using forms of computer vision and forms on a phone app that can help them identify whether or not a crop is being blighted. We’re using already kind of AI in traffic flow and to try to sort of direct traffic and kind of retime stoplights. So you can cut commutes or you can redirect traffic. So these are all–to go back to my definition–systems that take an image or a data pattern or a question, make an inference, and generate an output that hopefully helps to augment what humans are doing, maybe improve what humans are doing, and maybe help humans make better decisions. So those are, I think, cool things. I mean, we just have been watching Artemis II. That is full of AI computer simulations that help them to track how they were going to do this incredible 10-day journey–also cool.
Concerning? We’re living with a lot of that right now. We’ve got this kind of great race happening in the world of looking for a job, right? So, you can now more easily do your résumé and your cover letter using AI. But now AI systems are being used to screen your résumé out. So people are now sending dozens and dozens of résumés out on a given day. But they’re getting screened out right away. So the downside of this is that it might filter people out of an applicant pool before anybody ever sees your name or anybody ever actually looks at your credentials, and nobody will tell you why potentially. There’s some research that suggests… You know, again, as you talk about input data and making inferences from that, in things like employment, a lot of the input data is historical data. So in fields in which you’ve had historic racial discrimination or gender discrimination–like if you’re looking for the résumé of an excellent computer scientist–then a lot of algorithms have been shown to sort of kick people out. So people are losing access to opportunities with real implications for their liberties and their rights. There’s so-called “predictive policing” tools. The algorithm says that you should police it more because it’s been policed more historically, not because there’s actually new information suggesting that that should be the case. And then in the generative AI space, because I live partly in New York City, the Adams administration spent nearly, I think, a million dollars on this government chatbot–this NYC bot or NYC chat. The idea of it was good. It was supposed to help small businesses navigate all sorts of city regulations, which in a place like New York City, are voluminous. But it was telling them to violate the law. So it was giving advice like how to skim workers’ tips and how to discriminate against your tenant if you’re a landlord. I mean, it was fairly outrageous and, I think, well beyond the kind of whimsical term, “hallucination,” that we use that often suggests that it’s not a really big deal. And we shouldn’t be surprised that I think the Mamdani administration canceled that contract and got rid of the chatbot. But the concerning aspects, I think, also just give you a sense of all of the places in our lives–all of the sites–simultaneously that are being shaped in some way by some form of algorithmic decision-making or management.
ELIZABETH JOH: And I guess one of the ways to approach that is to say not just, like, these are technical problems, but since you’re mentioning all of the different ways that individuals might feel powerless or just confused about what’s going on, you can kind of use a civil rights approach. And of course in the Biden administration, you led the OSTP. And you’re credited with directing the White House blueprint for an AI Bill of Rights. And I would love for you to talk more about that. You know, this is a policy paper. It’s a white paper. So what was the process? How did you begin creating the blueprint? Who was behind it? Who did you talk to?
ALONDRA NELSON: Yeah. We came into office in the middle of a pandemic. And we came into office as a country having a racial reckoning. We were having an economic crisis. And I think those of us who work in the science and technology policy space knew both on the research side and also kind of saw brewing, amidst all of these kind of societal concerns, what was going to be happening in the algorithmic space. And we were having already examples, so for example, the YouTube videos about the so-called racist soap dispensers and faucets. If you have darker skin, you can’t get the soap to come out, which is a kind of application of AI. And I had the idea to do, in part, I think, borrow from lots of other examples. I mean, the Obama administration accompanied its Affordable Care Act with something called the Patient’s Bill of Rights. I think Ralph Nader had a Consumer Bill of Rights. So the Bill of Rights has been used variously, both by government and in folks in civil society, as a way to sort of think about a rights expansion in the face of kind of a new technology or a new social dynamic, for example.
So we got into office. We published an op-ed in Wired that came out in October of 2021. And we sort of used the Bill of Rights framing. And we kind of tried to draw a parallel to the country’s founding and noting that there was this time, in the 1780s and ’90s, that Americans adopted the original Bill of Rights to guard against really… They just created this powerful government. We’re about to celebrate the 250 years of the Declaration of Independence and then the Constitution. We had created this kind of powerful government technology, and we needed to place a check on that. So how did you secure our rights and our liberties–our opportunities–in the context of a kind of large and powerful government? So, we saw a parallel with kind of powerful technologies and the powerful companies that were pushing these powerful technologies and thought that there was a useful analogy. And we’re wanting to think with the American public about what might be equivalent guardrails against these new powerful domains. And so we were trying to kind of frame the blueprint for an AI Bill of Rights project within a kind of continuous U.S. or American tradition of aspiring to values, recognizing the shortcomings of the systems that we create and sort of thinking about what we might do to mitigate it.
ROMAN MARS: Can you tell us some of the five principles that are identified in the AI Bill of Rights?
ALONDRA NELSON: Sure. Yeah. So the five principles in the white paper that Elizabeth alluded to was released in October of 2022, so a year later. And what we did over the course of that year was a lot of public engagement. That Wired op-ed ends with an email address that can go right to the White House.
ROMAN MARS: That’s always a good plan.
ALONDRA NELSON: Yeah, so I think we wish more people had taken us up on it, but people certainly did. And we did kind of focus groups. We had what we called “office hours.” So everybody who worked on the team, which included policy generalists, AI scientists, computer scientists, folks who work on science and technology policy from academia, who had government experience, who had commercial experience– So it was a pretty broad team. And we would all block on our calendar time just to talk with people and that included high school students and rabbis, in addition to always the technology companies’ lobbyists. But we really tried to have a broad conversation. And the five principles are really distilled from those conversations. We weren’t trying to do anything novel. We were trying to sort of take from this year of conversation what the best is of what we think. What are the aspirations that we should have as we move as a society into a more algorithmic-shaped, mediated world? So, one was that AI systems should be safe and effective. That’s a very basic and almost consumer rights principle. Second, people should have protections from algorithmic discrimination. Third, there should be some modicum of data privacy. We are still fighting out what that might even look like, but again, these are kind of aspirations. Fourth, there should be notice and explanation so that you should have a right to know when an AI system is being used to make consequential decisions, like some of those that I was talking about, Elizabeth, when you asked me what’s concerning. Do we care if you get a bad Netflix recommendation and you end up watching a movie you don’t really like that the algorithm told you you were going to like? No, but when algorithms and more advanced AI systems are being made for consequential decisions about people’s lives, they should know about that. And if they want an explanation, they should be able to get one. And then lastly, the last principle is that there should be some sort of human alternative or fallback so that you should ideally be able to opt out. We build a lot of algorithms and social media systems as opt-in as opposed to opting people out. So can you opt out of an automated system? Can you talk to a real person instead of being kind of brought down into a circle of a phone tree hell where you keep trying to press zero to get to a person, particularly when it’s about something that affects your life? I mean, health insurance, jobs, housing… So these are really critical things. So that’s what we came up with. And it’s been variously sort of taken up by different kinds of constituencies. It’s become a kind of a civic infrastructure that is a way, I think, that allows different kinds of communities, particularly non-expert communities, to talk about why AI is important and how they want it to sort of sit in their lives and not sit in our lives.
ELIZABETH JOH: So from an ordinary person’s perspective, what would it mean to have a safe AI system? Does it mean that it’s not gonna make mistakes? What do you envision as an AI system that would follow this idea of safety?
ALONDRA NELSON: Yeah. So, my friend, Damon, who leads the Lawyers Committee for Civil Rights, will often say, “There’s more laws around your toaster than around the chatbot that you might have used this morning,” which is true. So we just basically don’t have, certainly at a federal level–there is some action happening at the state level–any kind of just basic consumer protection. So, I think many people are actually shocked when they realize that when an AI company or tech company ships a new model or an update of a model, no one has looked at that. There’s been no kind of third-party authority that’s said it’s met some threshold or standard of testing and we think it should be safe and effective. So there are affordances. There are things particularly about generative AI. And we know increasingly from the research that you’re never going to get rid of all of the mistakes in generative AI, certainly not in large language models. So safe and effective systems doesn’t mean that. But it does mean that one should expect that there should be testing on what people think would be the most obvious use cases of these technologies, right? If it’s a multifunction or a multi-use technology, there are use cases, I think, that we haven’t even imagined and people aren’t doing yet.
But I think anybody who has studied the history of technology in the United States, even just going back to the. ’90s… You know, we know there’s always going to be a problem with scams, scamming, and fraud–always–any kind of new technology. We know historically there’s going to always be a problems with forms of pornography, sexual abuse… These things are often the first use cases for new technologies. And so, that we have chatbots that are being used to nudify young people in high school or whatever, like, we can’t act like these are not harmful use cases that were not anticipated. And so it doesn’t mean at all, Elizabeth, that there won’t be unanticipated things or that a chatbot won’t hallucinate. But it certainly should mean that a company, before releasing a product, has thought through even basic historical use cases and actually thought about how they might be mitigated or should have a conversation about some independent stakeholder, state government, or civil society about how they might be mitigated.
ELIZABETH JOH: So, you’d think with all this, essentially, you’re sort of describing this experimentation that’s happening. And we’d expect that if the government is going to do this, they also should be regulating it a lot. And the answer at the federal level has been crickets, mostly. There has been some movement. I mean, the blueprint served as a springboard for President Biden’s executive order on AI. So could you say a little bit about what the core of those concerns were in the EO?
ALONDRA NELSON: Yeah, so I think the philosophy, both for the AI Bill of Rights and, for the most part, for President Biden’s executive order on AI, was that just because we have a new technology does not mean that we have to have a new social compact or a new social contract. Like, you don’t have to throw out every policy regulation in law because we have this new technology, as powerful as it may be. So if intentional discrimination or intentional violations of people’s civil rights or liberties are illegal in any other fashion, if you do that with AI, it’s also illegal, right? You might have to differently figure out the mechanism or differently make the case, but the legality of the outcome is the same. One of the things the executive order did was ask the Department of Education to think about what– You’ve got guidelines for children’s privacy and their protection for the use of educational technology. Do those need to be updated? Or do we just need to double down on what we have as you’re introducing different forms of advanced AI potentially to the classroom. You know, the President’s executive order had some directions to things like the Department of Labor. And I think, differently from what the current administration has been doing, it was not just what is AI going to do to work, it was how can government help put speed bumps or friction or help to direct the sort of direction of travel so that you’re not just potentially casting people out of work. You are helping them find other work. You are re-skilling them. Could there be a conversation about tax incentives or other kinds of incentives to keep people on work or to help people offboard or on-ramp to different work, for example?
The executive order, of course, also weighed in on, you know… There was a lot of concern and remains a lot of concern in the national security space. So, should there be export control? Should we be controlling where various forms of technology go? So this is still a very live conversation. Controversially, the executive order proposed that we would use the Defense Production Act from, I think, World War II originally to require that companies give the government more input and information about new, more powerful AI systems and tools that had a kind of certain threshold of capability. So, it might have been historically the longest executive order ever.
ELIZABETH JOH: Really?
ALONDRA NELSON: Yeah, I think that’s right. I think it was a hundred and some pages–101 or 102. As a reformer, I don’t necessarily think that is a good thing. In some ways it’s a bad thing. But in this case, I think it was good in the sense that it tried to be comprehensive–that the philosophy here was that this is a kind of new infrastructure. This is sort of a new operating system for a lot of the work that we do and how we might think about the ways that government can both help to accelerate potential good use cases and mitigate potential harms, using the things, the tools, the mechanisms, and the levers that government agencies and the executive already have.
ROMAN MARS: We’re gonna take a break. But when we come back, we’ll turn to how the federal government is and isn’t regulating AI and how the states are filling in the gaps.
[AD BREAK]
ROMAN MARS: Before the break, we talked to Dr. Alondra Nelson about how to think about artificial intelligence and why it poses a risk and should be regulated. And so how did her work lead to a conversation about preemption?
ELIZABETH JOH: Well, as she’s already mentioned, during her time in the Biden White House, she helped create the blueprint for an AI Bill of Rights. And that blueprint became the impetus for a part of President Biden’s 2023 executive order on AI. And as she has already discussed, that order told the federal agencies to address the safe and ethical use of AI. Now, that’s the limit of what President Biden can do. And that’s because Congress has the power to legislate, not the president. So while Biden could tell the executive branch what to do about AI, he lacked the authority to actually preempt state law.
ROMAN MARS: Got it.
ELIZABETH JOH: And as soon as he began his second term, Trump rescinded or did away with Biden’s executive order and replaced it with his own. Now the Trump administration’s approach to AI has been to turn away from a focus on safety and ethics in AI.
ROMAN MARS: [SCOFFS] Surprise, surprise. Okay.
ELIZABETH JOH: And instead, to focus on what the federal government can do to accelerate AI development. Now, Trump’s executive order has called upon Congress to use its power of preemption, based in the Supremacy Clause, to override state laws on AI. Congress so far has not responded.
ROMAN MARS: Okay, which has left a lot of room for the states and so we’ll pick up our conversation there.
ELIZABETH JOH: So we have a lot of different states regulating on AI. California has been in the lead, as it often is in these areas. California has just a lot of different laws on AI. For example, you’ve got to disclose what kind of data you use if you’re an AI developer–what you use to train your models. That seems very technical–very big picture. There’s also some very specific California laws. We just passed a law that if you’re a police department, you’ve got to disclose if you use generative AI when your officers write their police reports.
ALONDRA NELSON: That’s a good one.
ELIZABETH JOH: So you’ve got the whole range of different things. So, what does that mean? You talk to a lot of people in the industry. If I’m an AI developer and I want to offer my product in California or I want to offer my products in Colorado, which has an algorithmic discrimination law, what does that even mean? How does that work?
ALONDRA NELSON: Well, I mean, I think the first thing to say is that we have other industries where you have different kinds of regulation. Insurance is regulated mostly by the states, for example. We’ve talked a little bit about consumer protection. I think the discourse that gets used in D.C., which is its own language, there’s a lot of kind of pearl clutching around the fact that you would have different laws in different States. Although, the very same people in Washington, because they are the most adept people on what the regulatory space looks like more broadly, writ large, know that it’s true. It’s basically true. We use the phrase, “laboratories of democracy.” I think there is something to that. I mean, you have a new technology that is fast moving. In some ideal demos, would you want just one law to rule them all? Sure. Right. But we don’t live in that ideal demos. And we also know that the states are much closer to the harms.
You also have to imagine being a governor of a state or a state legislator or a senator in a state, and you have people writing to you about being worried about the future of their children. We had a scandal. I’m sitting here in Princeton in New Jersey about nudify apps–lots of just, I think, concerns about young people harming themselves. You know there’s been something. I saw a case reported about a potential homicide. And I think if you are a state legislator and you’re hearing from constituents who’ve been denied a mortgage or screened out of a job by an algorithm, you can’t just sit blithely and sort of not respond to that. So I think partly it’s just, like, folks are hearing it. I think that we have a new technology. What are the best ways to think about this? I mean, you mentioned California and New York, which have done laws around kind of trying to require some disclosure and transparency from companies around harms. Texas has weighed in actually on thinking about harms and including discrimination. But they’ve said it really has to be intent. It can’t be, if there’s unintentional harms, that they’re trying to let the companies off the hook. A place like Colorado has attempted the first of what we might think of as, like, an omnibus AI bill that covers lots of things, including sort of harms to young people, deep fakes, and discrimination. And I’ve just named three different approaches. And it’s not clear which one of those is the best one or which one’s gonna be most efficacious. And I think it’s worth actually letting states do this–finish the work of implementing these laws and actually find out. I just don’t think that the harms are more likely to be on the side of not doing anything at all, rather than trying to do a couple of different innovative strategies in different states to see. And then because there’s been no federal law, there’s obviously just this vacuum in the states. And there’s a lack of clarity. The D.C. Conversation–the Trump administration conversation–has been, well, creating confusion. And I think what’s actually creating confusion is the lack of any kind of federal guidance. It’s actually the states that are trying to sort of bring clarity to chaos.
ROMAN MARS: I mean, if the states are the appropriate front line for figuring this stuff out, is the ideal form of that to eventually roll up into some kind of federal regulation that makes sense?
ALONDRA NELSON: Sure. I mean, I think what the state patchwork does is test things out. Some things will work. Some things will fail horribly. I think it also creates some kind of so-called patchwork. I think it kind of creates some upward pressure because–exactly to your point, Roman–when enough states act, federal policy or norms become, as the patchwork gets woven together, kind of implicit. And I think it puts more pressure on it for the federal government to actually do something explicitly. If we widen the aperture just slightly broadly from the AI companies that we’re talking about now to the social media example–which gives us another 15,10 years more to think about–we’ve seen the utter failure of the federal government to be able to legislate in that space. And to the extent that we’ve got anything that looks like regulation or law or governance in that space, it’s coming out of these lawsuits that we saw decided a few weeks ago around Meta and YouTube. And so I think if you are a state-level executive–if you’re a governor or a state legislator–you’re thinking back about that example and just thinking, “We can’t wait and do this again.” As I said, the states are close to the harms. They’re hearing from constituencies. The way that we’ve been governing, if you think about the social media model, the young woman who was the plaintiff, I think, in the Meta case is 20 years old. This happened eight years ago or something. She was a child when this happened. And so using liability and legal cases puts us quite far away from the harms. And I think the states can be much closer.
ELIZABETH JOH: Yeah, just to back up for a moment by way of explanation, you’re referring to the social media trials that are happening in California, where basically the state’s attorney generals and private plaintiffs are suing, arguing that social media platforms are harmful products, which has a long storied history of legal liability in the United States. And actually they’re using the legal playbook of big tobacco. We kind of shut down big tobacco because we argued that the companies knew that these were harmful products and you sold them anyway. And that has proven so far to be successful in the social media space. So I guess we could think of perhaps, you know, AI– Some of these products are gonna be dangerous, and maybe we’ll do that. Of course, I think you’re right to say that this is a backup. We don’t want to wait for the bad use case for people to be harmed. I mean, the nice thing about regulation is you can be proactive and say, “We think this is going to happen,” or “It is happening, and we want to affect as many people as we can within the state or within the country.” My question is really more about what about the companies? Not that I feel too bad for them, but if you’re a company, it’s pretty burdensome, I would think, that you’ve got to look at every state and see what every state is doing. So I would imagine that their first choice is no regulation, right? But their second choice must be federal regulation, no?
ALONDRA NELSON: Yeah, I mean, I would disagree with that a little bit. Let’s have a friendly quibble about this. I think that the compliance burden argument is a bit overstated by companies, right? That’s just what companies do in their own interest and their lobbyists’. And as I said, I think companies already in other policy spaces are navigating different consumer protection regimes for different states, different employment laws, and different privacy frameworks. The state of Illinois has this pretty strong biometric policy regime. And yet companies– Clearview was still selling its facial recognition technology data set, for example. So I think that the language from companies and lobbyists that say that state AI laws are uniquely burdensome or especially burdensome doesn’t really hold up when you think about these other examples of these other policy spaces.
The other thing I would say is that what your question, which is a common question–an important one–presumes is that, if the states don’t have a law, there’s no other governance or pressure being applied on the direction of AI governance, which certainly in the Trump administration is not true. So, okay, maybe you don’t want to deal with California or Colorado, but you’ve got a Trump administration that’s saying, “We’re changing tariffs every day.” We’ve gone from liberation day to not liberation day back and forth. So companies are dealing with that, including AI companies. You’ve got a Trump administration that is saying, “We don’t like immigration. We’re uncomfortable with science and tech immigration. If you want to bring a new technology talent AI company, you’re going to have to pay $100,000 per visa if we allow you to have one to bring a talented engineer from France or Korea or something.” And then they’re also intervening in business. The U.S. taxpayer is a shareholder in Nvidia. We’re a shareholder in Intel. So, the compliance burden question, I think, is much too narrow given all of the different ways in which companies are being asked to respond to a kind of broad spectrum of AI governance.
ELIZABETH JOH: Yeah, and let’s not forget, I should say, the federal government and all of the state governments are huge customers, right? You know, customers can demand changes if they want.
ALONDRA NELSON: Procurement is an excellent vehicle. I mean, Governor Newsom just signed this executive order that, I think, really leaned into that, including not only safety issues, but issues around discrimination and civil rights and liberties, which I thought was fantastic.
ELIZABETH JOH: So we’ve talked a lot about sort of granular harms that are potentially happening or are happening. But I do want to talk about your thoughts on what’s on the horizon, the AI horizons. There seems to be this race to develop AGI or artificial general intelligence. So the idea would be not, like, please find all the cats in this picture or write my high school essay on Pride and Prejudice. It’s an all-purpose, sophisticated AI with autonomy. Now, you’ve spoken to a lot of people in tech. I’ve spoken to a few. It seems like some people in the AI policy world are extremely worried about this, like we could create something that gets totally out of control, develops like a biological weapon, and takes over our defense systems. How concerned are you about this as a subject and then an object of regulation?
ALONDRA NELSON: So, I’m concerned about it as– I think some people are quite invested in the name and what the name means. People are invested in whether it’s super intelligence or AGI. I am not at all invested in the name, and I don’t really care. So it keeps me out of some fights, but probably also keeps me out of some parties. I don’t know. But I do prefer to use the phrase “advanced AI.” There are significant concerns about advanced AI. So, for example, if we think about the Doge early last year in the Trump administration, part of what the reporting in Wired and elsewhere was suggesting is that Doge was breaking the Privacy Act of 1974, which said that a lot of inter-agency organizations could not share data, in part because you don’t want the federal government to have administrative data about you from Health and Human Services, from Fannie Mae, from whatever, to be able to put into this kind of large surveillance kind of panopticon. And I think what powerful AI systems do is allow the interoperability of that data and the sort of discovery potential of associations that are dangerous–things that we could never possibly know about ourselves or about others. So that’s not even AGI, right? That’s just sort of a powerful extreme. So if you imagine a system having access to data about everyone in the United States–everyone in the world–being able to sort of constantly be evaluating that data, running that data, and making decisions… And again, you know, I mentioned at the beginning the various forms of the autonomy or not of different AI systems and to do it autonomously. So imagine not just all of the open claws–not all the little lobster claws of various agents–but a really big claw, like a really powerful independent agent sort of acting in the world.
And so there’s been some reporting and I’ve seen some people discussing on social media things like, “I used this agent and it wiped out my entire hard drive,” or “It deleted all of my emails.” And that’s happening. And we’re not imagining an AI agent that was sentient and all-knowing and, like, decided that it was gonna wipe out all of your email because you worked too hard or because it doesn’t want you to work or whatever. Those are just powerful systems that we’re learning to use. So then you can imagine potentially a system having a bit more intentionality–a bit more understanding of its stakes and being more powerful. The question then becomes–and I think this is where we trip ourselves up–well, how do you regulate that? It’s just so powerful. What are we gonna do? And before you get there, you need to imagine that companies can actually be told not to build a thing. Or they can be told that they can’t ship a thing. You can’t tell a company what to create, but you can certainly say, “You can’t ship this out into the world without certain controls.” Someone needs to be able to have a kind of final decision on whether or not it ships or to be able to turn it off and on or you can only run it for a few hours or it can only have so much access to so much data. And we’re not having those kind of, I think, system-wide conversations. And to go back to the kind of subject of the broader conversation, that is where you would want a smart, prudent, federal government to sort of weigh in. At that level of nuance and both level of abstraction and power, you might want there to be some sort of federal law or legislation or guidance.
ROMAN MARS: When we come back, Dr. Nelson explains her vision for finding a consensus on AI regulation and whether she’s optimistic the government will figure this out.
[AD BREAK]
ROMAN MARS: I mean, you developed this idea of a kind of thick alignment when it comes to AI governance. Can you talk more about what “thick alignment” is and how that translates to regulation?
ALONDRA NELSON: Yeah. Sure. So there’s a wonderful writer, Brian Christian, who has a really important book that I would commend to people called The Alignment Problem that is really writing about the first early years of what some people call “AI safety,” which is basically just like, “How do we explain them? How do we interpret what they’re doing? How do we demonstrate that they’re safe to the extent possible?” And it was very much a kind of technical sense of thinking about alignment. So, the system says that it’s supposed to identify at 98% with a margin of error of 2% or 3% these people in a facial recognition technology system. And for all intents and purposes, you would say that system is “aligned,” right? But we know the system is misidentifying people. We know in the Detroit metropolitan area that there have been more than half a dozen people misidentified by facial recognition technologies that someone, somewhere in the development and deployment queue, said, “This is aligned. This product works.” And so, as we’re thinking about AI systems and advanced AI systems, it’s not just whether or not they kind of work technically. What happens or what can we anticipate or not anticipate when you deploy them? And how do we create a process or an understanding that allows us to be thinking about alignment as something that needs to happen fairly continuously over time? And also that’s something that means to happen in conversation with the values of different communities and different societies. So by thick alignment, I am taking up the work of the philosopher, Gilbert Ryle, but also the anthropologist, Clifford Geertz, who was a professor here at the Institute of Advanced Study in Princeton, where I am, who has this very famous essay on the concept of thick description. You don’t really understand the world until you’ve really thought to understand contextually–deeply–what it means and how do you describe it deeply.
And so my provocation to AI safety researchers and my collaboration, actually–it’s not just a critical work–is sort of… Alignment is important. Safety, explainability, interpretability, and all the things that you might put in that bucket are really important and, taken together, are an important solution set for some of the harm mitigation that we might want to do in the space of AI. But what does it mean to do that in a way that takes seriously the different contexts in which these tools might be used–the different values? Anthropic has created a constitution for AI, for example. Who gets to weigh in on that? And are those the values that I want or others want? You see this kind of value conversation coming up also in even some of the Trump administration’s way that they frame sort of “ideological bias in AI.” Who gets to decide what’s biased in AI? There’s a technical question about bias in AI. But who gets to decide sort of what is a biased chatbot? So, I think we just need to have a conversation we’re not having about what it means to try to come to rough consensus values to the extent that’s even possible–try to have high-level values to make decisions about these technologies. So, I think the AI Bill of Rights was one of the ways that we were trying to point to that. But certainly, I think, state laws are another way. I mean, you might think of those as examples of thick alignment. “This is what our constituency cares about. And this is where we’re going to lean in in the regulatory space with regard to AI. The other stuff, maybe we don’t care about so much.”
ROMAN MARS: I don’t even know if I have thick alignment with President Trump as a human. You know what I mean? It seems harder and harder to have it when you’re talking about all these hypothetical uses of this stuff. And if something like a program is supposed to have inherent human values, a lot of those don’t feel shared right now.
ALONDRA NELSON: No, I think that’s right. I will say, one of the things I’ve been doing since– The AI Bill of Rights comes out on October 22, 2022, and I’ve since that time been following its afterlives. And some of its afterlives, Roman, to your point, have been in red states. So, there’s been an Oklahoma AI Bill of Rights introduced as a bill. You know, it didn’t succeed, ultimately, but it contained all of the five principles that we discussed previously, plus a few more that were really good and actually quite stronger than some of the things that we suggested. More recently, in November, in Florida, Governor DeSantis introduced a Florida AI Bill Of Rights, which contains within it all of the five principles we have and lots of other things besides. Deepfakes, child sexual abuse imagery, a really nice clause that was around health insurance and being able to get a decision around algorithmic uses of health insurance… So I totally take your point, Roman, but it’s also clear that there are a few things that we agree are wrong or that we don’t want that are suboptimal for society. I think you’re exactly right, but I also take some comfort in these Bill of Rights alignments that pop up here and there.
ELIZABETH JOH: You sound optimistic about the future of AI regulation, is that right?
ALONDRA NELSON: I’m not optimistic about… Am I optimistic about regulation? I don’t know. I think if we look at the history of technology policy at the federal level in the Congress–Elizabeth, correct me if I’m wrong–I think it’s maybe not been since the Communications Decency Act of 1996 that we’ve passed anything like a technology law. That’s a long time. That’s a generation. So I’m not optimistic in that sense. I think I’m optimistic with… You know, some people are calling it the tech backlash, and I don’t call it that. I don’t like that framing. But there’s a growing public empowerment to speak about what people want and don’t want with regards to the way that AI systems are being developed and deployed. So when I first started working in sort of big data, and then it became AI policy and research…
ELIZABETH JOH: That’s how you date yourself.
ALONDRA NELSON: I know. I know. You say “big data” in a room and people kind of cringe. They’re just like, “Big data. So cringe.” So, you would sit in rooms and say people can’t possibly understand. I mean, even now you hear people saying, “If you don’t really have a PhD, if you’re a staffer on the Hill, and you don’t have a degree in machine learning or AI, how could you possibly even begin to offer guidance on how we should govern this technology?” So, of course you don’t want people who know nothing about AI to be governing AI. But I’ve been encouraged by the fact that the public has demonstrated that it is not true that you have to have a PhD in AI to be able to say something about the AI governance space. So you see it in the space of data center. That’s a place where AI governance and policy is quite tactile. It is in communities. It is about their water. It’s about their energy use. And it’s where sort of AGI or super intelligence lands on the ground. And it is where communities really feel they have a sense of agency around that. I think I just saw in the news that Maine has banned data centers for a time. There were a lot of big projects announced that have been stalled that are being revisited. There’s reporting now about how a lot of these data center agreements in various communities were done with local politicians under NDAs and local communities can’t even know the terms of the agreement for some of these. And people are really pushing back against that. And they’re pushing back against the harms to young people. They’re very concerned about suicidal ideation and how chatbots encourage them. So am I optimistic about law? Absolutely not. But am I optimistic about the fact that it’s getting much more difficult for companies and other elites who really wanna just drive technology without thinking about the harms and the social implications to do that, because you’ve got a growing chorus of people–maybe not aligned, but bipartisan, Roman–saying that we don’t want this. And so I find optimism, no–encouragement, yes.
ROMAN MARS: My one point here is, like, one of things that is funny is the biggest proponents of AI and the broad use of it are kind of the biggest fear mongers of it, too. I think they kind of enjoy the sort of sense of… “This is super powerful. You should let us do what we want to. And it’s going to destroy humanity in five years.” I think they like both of those things, so I think both of them feed into their ego.
ALONDRA NELSON: They’re both about power. Yeah. For sure.
ROMAN MARS: It’s fascinating because that the alarmists are the biggest proponents is a weird dynamic. This is not like tobacco regulation where the people who wanted to regulate were just on the side of harm and the other people were like, “No harm.” It’s an odd dynamic and it’s also one of things that’s mixed up in all this stuff of the Florida regulation versus the California regulation. And the political valence of this stuff is much more complicated than most other things.
ALONDRA NELSON: Yes, it’s very complicated and kind of heterogeneous and so that’s fascinating. And I think there’s some very interesting essays, articles, and papers to be written about at a time of maybe historically, since we’ve been measuring, highest polarization in American society. You’ve got this growing negative sentiment about AI, and that is bipartisan. And the issue set about which people are having agreement of their dissatisfaction around is growing. So you go from kind of discrimination to young people and CSAM to fraud to healthcare. Like, the space is just becoming much broader to data centers, for example. People are obviously worried about their jobs and worried about employment and what they’re being told–Roman, to your point–about powerful people saying our powerful tool is going to be really great and destroy everything, including all of your jobs. So, yeah, it’s a very interesting policy space. And I think of political encouragement if not optimism.
ROMAN MARS: Yeah. I mean, this seems like a new opportunity for a different kind of alignment, which is really kind of fascinating.
Dr. Nelson, I really appreciate you being here.
ELIZABETH JOH: Thank you so much.
ALONDRA NELSON: It’s been great to talk to you.
ROMAN MARS: So that’s the original seven articles of the Constitution. Thank you for joining for all of that. Of course, there are amendments to be talked about–27 of them–but we’re going to take a pause on the Breakdown of the Constitution. There’s just so much going on with Trump and the Constitution that we’re going to go back to releasing our What Trump Can Teach Us About Con Law episodes. There won’t be an episode in May, but we’ll be back in June for Supreme Court decision season, everyone’s favorite season.
ELIZABETH JOH: The 99% Invisible Breakdown of the Constitution is produced by Isabel Angell, edited by committee. Music by Swan Real. Mix by Martín Gonzalez.
ROMAN MARS: Kathy Tu is our executive producer. Kurt Kohlstedt is our digital director. Delaney Hall is our senior editor. The rest of the team includes Chris Berube, Jayson De Leon, Emmett FitzGerald, Christopher Johnson, Vivian Le, Lasha Madan, Joe Rosenberg, Kelly Prime, Jeyca Medina-Gleason, Talon and Rain Stradley, and me, Roman Mars. The 99% Invisible logo was created by Stefan Lawrence. The art for this series was created by Aaron Nestor.
We are part of the SiriusXM podcast family, now headquartered six blocks north in the Pandora building… in beautiful… uptown… Oakland, California. You can find the show on all the usual social media sites, as well as our own Discord server, where we have fun discussions about constitutional law, about architecture, about movies, music–all kinds of good stuff. You can also find a link to the Discord server, as well every past episode of the Con Law Book Club and every past episode of 99PI, at 99pi.org.
Leave a Comment
Share