Constitution Breakdown #9: Alondra Nelson

This episode of 99% Invisible's Constitution Breakdown series tackles Articles VI and VII of the U.S. Constitution, which contain the Supremacy Clause, the No Religious Test Clause, and the ratification process. Hosts Roman Mars and Elizabeth Joh walk through these often-overlooked articles before diving into a deep conversation with Dr. Alondra Nelson about one of the most pressing federalism questions of our time: how to regulate artificial intelligence. The episode moves from the 18th-century mechanics of federal supremacy to the 21st-century reality of states racing to regulate AI in the absence of federal action, creating a rich exploration of how constitutional structures shape contemporary governance.

1:21Article VII: The Ratification Clause

Article VII, the ratification clause, is the least legally consequential today but was absolutely essential for the Constitution's existence. It specified that nine states would be sufficient to ratify the document and make it legitimate. This actually happened on June 21, 1788, when New Hampshire became the ninth of the original 13 states to ratify. Joh notes that there are no important Supreme Court cases on this clause—it served its purpose and has largely faded from legal relevance. The clause was a practical necessity: it established the ground rules for how the Constitution would become binding, solving the chicken-and-egg problem of needing a constitution to create a government but needing a government to adopt a constitution.

2:10Article VI, Clause 1: The Debts Clause

The first clause of Article VI addresses debts the United States already owed, particularly those from the Revolutionary War. Creditors were nervous that a new Constitution might wipe out existing obligations, so the clause explicitly assured them that the new government would honor all pre-existing debts and engagements. Mars points out an interesting drafting detail: the original version gave Congress both the obligation to pay debts and the power to pay them, but the second part was moved to Article I as part of Congress's spending authority. This seemingly minor rearrangement had enormous consequences—that spending power in Article I is now one of the most frequently cited justifications for federal legislation. Today the Debts Clause is mostly of historical interest, since the United States did in fact pay its debts, but the story illustrates how constitutional drafting choices can have ripple effects far beyond what the framers likely anticipated.

3:35Article VI, Clause 3: The No Religious Test Clause

This clause requires all federal and state officers to take an oath supporting the Constitution, but explicitly states that "no religious test shall ever be required as a qualification to any office or public trust under the United States." Joh explains that this provision is notable primarily for its absence—it breaks from English common law tradition, where government officials had to swear allegiance to the Church of England and disclaim Catholicism. Under the Articles of Confederation, many colonies still required some form of religious affirmation from officials. The clause formally prohibits religious tests, though Mars notes that in practice, representation of non-Christian faiths in public institutions remains uncommon. Interestingly, this clause generates very little Supreme Court case law today because religious freedom questions are almost always litigated under the First Amendment's Free Exercise Clause rather than this provision. The clause represents a foundational commitment to religious pluralism in public office, even if its practical legal significance has been superseded.

5:57Article VI, Clause 2: The Supremacy Clause

This is the heavy hitter of Article VI. The Supremacy Clause declares that the Constitution, federal laws made pursuant to it, and federal treaties "shall be the supreme law of the land." Joh explains that this responded to a specific problem under the Articles of Confederation: there was no equivalent provision, so state courts sometimes simply refused to apply federal law, creating chaos. The Supremacy Clause eliminated that ambiguity with one stroke—federal law wins when it conflicts with state law.

But the simplicity ends there. The clause gives rise to the doctrine of preemption, which is probably the most frequently used constitutional law in practice. Preemption means Congress has the power to displace or override contrary state or local law. Sometimes Congress explicitly states that a federal law preempts state law—those are the easy cases. More often, Congress says nothing about preemption, and courts must figure out whether state and federal laws can coexist. The Supreme Court has developed various tests: sometimes an area of federal interest is so important (like foreign policy) that states cannot touch it at all; other times courts ask whether it's possible to comply with both laws simultaneously. If it is, both may stand. But if state law creates an obstacle to federal objectives, or if literal compliance is impossible, federal law preempts state law.

Joh emphasizes that preemption comes up constantly because the federal government and states regulate in overlapping areas—environment, consumer protection, energy, and countless others. Every time a new policy problem emerges, there's a race to regulate, and the question of which level of government should take the lead is always present. This sets up the episode's central tension: the old constitutional question of federalism meets the new technological frontier of artificial intelligence.

12:39Defining AI and Why It Matters

Dr. Alondra Nelson begins by defining AI using a modified version of the OECD definition that 38 nations have agreed upon. AI systems are machine-based systems that use statistics and math to make inferences from inputs and generate outputs—predictions, recommendations (like Spotify or Netflix), or content. She emphasizes that AI is not just generative AI like ChatGPT, which came out in 2022 and became the public face of the technology. Generative AI is actually the smallest Russian doll within a broader set that includes machine learning, deep learning, and other approaches. AI systems vary in autonomy and adaptiveness: some are static, like predictive algorithms used in criminal justice that work from fixed datasets; others are increasingly autonomous, like AI agents that make purchasing decisions or write code.

Nelson then explains why people should care. On the transformative side, AI can read chest X-rays and flag early-stage cancers, potentially saving lives while addressing shortages of radiologists. Farmers use computer vision apps to identify crop blight. AI helps optimize traffic flow and commutes. NASA's Artemis 2 mission uses AI simulations. But the concerning examples are equally vivid. In employment, people now use AI to write resumes and cover letters, but AI systems also screen those resumes out—often for reasons nobody can explain, and potentially replicating historic patterns of racial or gender discrimination because the training data reflects past bias. Predictive policing tools direct police to areas that were heavily policed historically, creating a feedback loop rather than reflecting actual crime patterns. And in New York City, the Adams administration spent nearly a million dollars on a government chatbot meant to help small businesses navigate city regulations—but it told users to violate the law, including how to skim workers' tips and discriminate against tenants. The city canceled the contract.

23:34The Blueprint for an AI Bill of Rights

Nelson explains the origins of the White House Blueprint for an AI Bill of Rights, which she spearheaded as acting director of the Office of Science and Technology Policy. The Biden administration came into office during a pandemic, a racial reckoning, and an economic crisis—and they could see algorithmic harms brewing. Nelson drew inspiration from historical precedents: the Obama administration's Patients' Bill of Rights accompanying the Affordable Care Act, and Ralph Nader's Consumer Bill of Rights. The framing deliberately paralleled the original Bill of Rights, which was created to guard against the powerful new government technology the Constitution had just created. The analogy was that powerful AI technologies and powerful companies pushing them needed equivalent guardrails.

The team spent a year doing public engagement—publishing an op-ed in Wired with a direct White House email address, holding focus groups and "office hours" where anyone could talk to the team, including high school students, rabbis, and technology company lobbyists. The five principles distilled from these conversations were: (1) AI systems should be safe and effective; (2) people should have protections from algorithmic discrimination; (3) there should be data privacy; (4) people should have notice and explanation when AI systems make consequential decisions about their lives; and (5) there should be a human alternative or fallback, allowing people to opt out of automated systems for important matters like health insurance, jobs, and housing.

Nelson emphasizes that safe and effective doesn't mean error-free—large language models will always hallucinate. But it does mean companies should test for obvious and historically predictable use cases before releasing products. As her friend Damon from the Lawyers' Committee for Civil Rights puts it, "There's more laws around your toaster than around the chatbot that you might have used this morning." The Blueprint became the springboard for President Biden's 2023 executive order on AI, which was reportedly the longest executive order in history at 101 or 102 pages. Its philosophy was that new technology doesn't require a new social contract—if intentional discrimination is illegal without AI, it's illegal with AI too. The order directed agencies to use existing tools and levers to accelerate beneficial uses and mitigate harms.

36:27The Federalism Tug-of-War Over AI Regulation

The conversation turns to the current regulatory landscape. Biden's executive order could only direct executive branch agencies—it couldn't preempt state law because Congress has that power. When Trump began his second term, he rescinded Biden's order and replaced it with one focused on accelerating AI development rather than safety and ethics. Trump's order called on Congress to use its preemption power to override state AI laws, but Congress has not responded. This has left a vacuum that states are filling.

California has been the leader, passing laws requiring AI developers to disclose what data they use to train models, and requiring police departments to disclose if they use generative AI in writing reports. Colorado passed an algorithmic discrimination law. Texas has weighed in on harms including discrimination but requires proof of intent rather than allowing claims based on unintentional harm. Nelson notes that this patchwork is actually normal—insurance is regulated mostly by states, consumer protection varies, and companies already navigate different employment and privacy regimes across states. The "compliance burden" argument from industry lobbyists is overstated, she argues, especially given that companies are also dealing with Trump's unpredictable tariff policies, immigration restrictions that make it harder to hire foreign AI talent, and government intervention in business (the U.S. taxpayer is a shareholder in Nvidia and Intel).

Nelson sees the state patchwork as laboratories of democracy that test different approaches. Some will work, some will fail, and the accumulation of state action creates upward pressure for federal norms. She draws a parallel to social media: the federal government has utterly failed to legislate in that space for a generation, and the only governance has come through lawsuits using the Big Tobacco playbook—arguing that companies knew their products were harmful and sold them anyway. States can't afford to wait that long again, especially when they're hearing directly from constituents about harms to their children.

48:47Advanced AI, AGI, and the Need for Federal Action

Mars asks about artificial general intelligence (AGI)—the hypothetical all-purpose AI that could develop biological weapons or take over defense systems. Nelson says she doesn't care about the name, preferring "advanced AI," but takes the concerns seriously. She gives a concrete example: DOGE (the Department of Government Efficiency) in the early Trump administration was reportedly breaking the Privacy Act of 1974 by sharing data across agencies. Powerful AI systems enable the interoperability of that data and the discovery of dangerous associations that could create a surveillance panopticon. That's not even AGI—it's just powerful systems with access to data.

The real question, Nelson argues, is whether we can tell companies not to build or ship certain things. You can't tell a company what to create, but you can say they can't ship it without controls—someone needs to be able to turn it off, limit its compute or data access, or make a final decision on deployment. These are the kinds of system-wide conversations we're not having, and this is precisely where a smart federal government should weigh in.

53:33Thick Alignment and Bipartisan Concerns

Nelson introduces the concept of "thick alignment," drawing from philosopher Gilbert Ryle and anthropologist Clifford Geertz's idea of "thick description." Technical AI alignment asks whether a system works within its specified parameters—whether it identifies faces with 98% accuracy, for example. But we know that facial recognition systems have misidentified more than half a dozen people in the Detroit metropolitan area alone, despite being technically "aligned." Thick alignment means thinking about alignment continuously over time and in conversation with the values of different communities. Who gets to decide what values an AI system embodies? Anthropic created a constitution for its AI, but whose values does it represent?

Nelson finds encouragement in the fact that the AI Bill of Rights has had "afterlives" in unexpected places. An Oklahoma AI Bill of Rights was introduced (though it didn't pass) containing all five principles plus stronger provisions. In November, Florida Governor DeSantis introduced a Florida AI Bill of Rights containing all five principles plus additional protections around deepfakes, child sexual abuse imagery, and health insurance decisions. Despite deep political polarization, there are things people across the political spectrum agree are wrong.

1:02:28Optimism and the Growing Public Voice

Nelson is not optimistic about federal legislation—Congress hasn't passed a major technology law since the Communications Decency Act of 1996, a generation ago. But she is encouraged by the growing public empowerment around AI governance. When she started working in this space, people would say you couldn't possibly understand AI without a PhD. The public has proven that wrong. People are pushing back against data centers in their communities, concerned about water and energy use. They're worried about chatbots encouraging suicidal ideation in young people. Maine has temporarily banned data centers. Communities are discovering that local politicians signed agreements under nondisclosure agreements, and they're demanding transparency.

Mars notes a fascinating dynamic: the biggest proponents of AI are also the biggest fear-mongers, simultaneously claiming their tools will destroy humanity and that they should be allowed to do whatever they want. Both narratives feed into their sense of power. Nelson agrees, noting that this makes AI regulation politically complicated in a way that tobacco regulation never was—the political valence is heterogeneous and cuts across partisan lines. But the growing bipartisan dissatisfaction with AI, across issues from discrimination to children's safety to fraud to healthcare to jobs, creates what she calls "political encouragement, if not optimism."

Conclusion

This episode matters because it connects the oldest structural questions of American constitutional law to the newest technological challenges. The Supremacy Clause, written in 1787 to solve the problem of state courts ignoring federal law, now frames the debate over whether California or Colorado or the federal government should set rules for artificial intelligence. Dr. Alondra Nelson's expertise bridges the technical and the constitutional, showing how the Blueprint for an AI Bill of Rights drew on the same logic as the original Bill of Rights—creating guardrails against powerful new forces. The conversation leaves the listener with a clear sense that the federalism question is not abstract: it determines whether a chatbot can tell a landlord to discriminate, whether a facial recognition system can misidentify someone without recourse, and whether communities can push back against data centers that consume their water and energy. The episode's texture comes from Nelson's blend of technical precision and practical wisdom, her refusal to accept either industry alarmism or industry complacency, and her insistence that ordinary people have both the right and the capacity to shape how AI enters their lives.

Key takeaways

• The Supremacy Clause (Article VI, Clause 2) establishes that federal law is supreme over conflicting state law, giving rise to the preemption doctrine—the most frequently used constitutional law in practice.
• The No Religious Test Clause (Article VI, Clause 3) formally prohibits religious tests for federal office, though religious freedom cases are now litigated under the First Amendment rather than this clause.
• The Blueprint for an AI Bill of Rights, spearheaded by Dr. Alondra Nelson, established five principles: safe and effective systems, protection from algorithmic discrimination, data privacy, notice and explanation, and human alternatives/fallbacks.
• Biden's 2023 executive order on AI could only direct executive branch agencies; it could not preempt state law because that power belongs to Congress.
• Trump rescinded Biden's executive order and replaced it with one focused on accelerating AI development, calling on Congress to preempt state AI laws—but Congress has not acted.
• States like California, Colorado, and Texas are passing their own AI regulations, creating a patchwork that Nelson argues is normal and potentially productive as laboratories of democracy.
• The "compliance burden" argument from industry is overstated, as companies already navigate different state regimes for insurance, employment, privacy, and consumer protection.
• Nelson finds encouragement in growing bipartisan public pushback against AI harms, even as she remains pessimistic about federal legislative action in the near term.