What this covers

Russ Roberts and Dean Ball explore a policy clash that spirals into constitutional crisis. In summer 2024, the Department of War contracted with Anthropic to use Claude for classified intelligence work, with contractual restrictions on domestic mass surveillance and autonomous lethal weapons. After Emil Michael's confirmation as a department official, the government demanded renegotiation, insisting on "all lawful use"—a term Ball argues is abnormal in defense procurement and effectively lets the administration define what is lawful until courts intervene. When Anthropic held two red lines, the Department of War issued a supply chain risk designation, treating a domestic company like a foreign adversary and threatening to bar contractors like Microsoft from using the service. This opening dispute becomes the frame for a wider argument: that constitutional constraints have eroded across decades, executive power has expanded unchecked, and norms defining what is off-limits have collapsed.

The conversation widens to examine whether institutions can hold when their foundational premises crack. Ball argues that the Common Fiction of the Republic depends on collective belief in shared rules; openly declaring it is dying risks collapsing it through run-on-the-bank dynamics. Roberts traces this decay back 60 to 90 years, showing how courts now serve as the only real check on discretionary executive action—a fragile arrangement because the executive ultimately enforces court rulings. The pair probe Legal Term vs Vernacular Meaning, showing how AI has made "domestic mass surveillance" legal in name but not in substance: combining location, satellite, and web data was once too labor-intensive to count as surveillance; AI scales expert-like attention infinitely, yet law has not caught up. They also examine whether Cognition vs Moral Judgment matters—whether machines answering factual questions convincingly creates illusion of moral clarity where only trade-offs exist. Throughout, they weigh whether the Public-Private Distinction can survive when government uses coercion to blur it, and whether democracies can remain republics if no power actually constrains the majority.

Sharpest takeaway

Ball and Roberts argue that the Department of War's coercive attempt to force Anthropic to drop its red lines on Claude exposes a deeper decay: the erosion of constitutional constraints and the blurring of public-private power, arriving precisely as a transformative AI technology demands governance the Republic is least equipped to provide.

  • The government's supply-chain-risk designation against a domestic AI firm treats a private company like a foreign adversary, usurping private property and possibly punishing political views.
  • The legal definition of 'domestic mass surveillance' no longer matches reality because AI can scale analyst-like attention infinitely on commercially available data, and the law hasn't been updated.
  • Executive power has been expanding and norms eroding for decades, so AI's most powerful capabilities arrive when institutions are weakest, leaving only the courts as a fragile check.

The argument · threads6 threads · 30 claims
0.76

The Department of War used supply-chain-risk designation and threats to force Anthropic to abandon safety restrictions on surveillance and autonomous weapons.

5 pointscentrality 5/5
  • When Anthropic refused to drop its two red lines, the Department of War threatened to designate Anthropic a supply chain risk, which would cancel all DOW contracts and bar DOW contractors like Microsoft from using Anthropic services in fulfilling DOW contracts; the designation was then actually issued.

    “we will designate your company Anthropic a supply chain risk. Which will mean that, a). All of your Department of War contracts are canceled; but more importantly so are all of your contracts with any Department of War contractors.”

  • After Senate confirmation in fall 2025, Department of War official Emil Michael reviewed the contract, decided the Department could not live with the surveillance and autonomous-weapons restrictions, and demanded renegotiation for 'all lawful use only'; Anthropic dropped many restrictions but held those two as red lines.

    “He comes in, he reviews the contract, he sees these usage restrictions, and makes the decision to... 'We have to have all lawful use only.'”

  • The supply-chain-risk designation is typically used only against companies from foreign adversaries to guard against adversary manipulation of American military systems, so applying it to Anthropic effectively treats a domestic company like an enemy of the state, and the broader threatened version would have been equivalent to sanctions.

    “this supply-chain-risk designation is only used--typically is only used--against companies from foreign adversaries. This is about adversary manipulation of American military systems... it's really treating Anthropic like enemies of the state, essentially.”

  • In summer 2024 the Department of Defense contracted with Anthropic for classified use of Claude (intelligence analysis, targeting recommendations), agreeing to usage restrictions including bans on domestic mass surveillance and on AI in autonomous lethal weapons that could identify, track, and kill targets without human intervention.

    “in that contract, the Biden Administration agreed to usage restrictions... two in particular were on domestic mass surveillance and the use of AI in autonomous lethal weapons.”

  • OpenAI reached an agreement in principle to supply classified models without Anthropic's contractual red-line protections, instead relying on technical safeguards—training and controlling the deployed system so the model could reason in real-time about whether a request constitutes domestic mass surveillance and refuse.

    “OpenAI is essentially hanging its hat on the notion of technical safeguards. So, instead of putting these safeguards into the contract... that system could, for example, reason in real-time about whether or not what it's being asked to do is domestic mass surveillance and say no to the government.”

0.64

Government coercion against a domestic company for political reasons violates property rights and First Amendment protections against state power.

2 pointscentrality 5/5
  • By telling Anthropic that if it does not do business on terms the government unilaterally sets it will set out to destroy the company—including by jawboning customers to stop doing business with Anthropic—the government engages in a usurpation of private property, and if motivated by Anthropic's liberal political donations it also amounts to political interference and a serious abridgment of First Amendment rights.

    “the government is saying here that if you don't do business on the terms we unilaterally set, we'll set out to destroy your company. Which is a kind of usurpation of private property.”

  • American companies have been able to do business abroad partly because, unlike Chinese tech firms widely seen as assets of the military and state, US firms like Microsoft are trusted as stable institutions separate from the government; coercive actions like the Anthropic designation erode that public-private distinction, making everything political and undermining the trust that gives institutions like Microsoft credibility.

    “everyone knows that Chinese technology companies are assets of the military and are viewed that way by the government. And, that's not the case in the United States. And, that has aided American companies in doing business abroad because there is a trust.”

0.63

The republic is eroding as court authority weakens and core institutional principles decay through gradual, nearly imperceptible phase transitions.

5 pointscentrality 4/5
  • Republics work because everyone believes in the common fiction of the Republic, which creates run-on-the-bank dynamics: publicly declaring the Republic is dying can itself contribute to its collapse, which is why Ball hesitated to publish his piece.

    “The reason that republics work is that we all believe in the common fiction of the Republic. And that's always been true.”

  • If you model classical-liberal institutions as a grand game (per Huizinga's Homo Ludens) where one institution—the courts—enforces the written rules, then it remains incentive-compatible for smart long-range actors to pretend the rules govern their actions so they have a better case in court; but this breaks down as court authority erodes and rulings are increasingly ignored.

    “if you are a smart long-range actor who wants to win in court, it's incentive-compatible for you to pretend like those rules of the game actually do govern your actions.”

  • The Founding Fathers would be appalled rather than reassured to learn that mass voting is the one element that persisted, because pure democracy is tyranny of the masses; a republic is not an omnipotent executive that swings wildly between two omnipotent executives based on a democratic vote, so observed elections alone do not constitute a republic.

    “the Founding Fathers, if you told them that the one thing that persisted was the ability of the masses to vote... they'd be appalled!”

  • Institutions can be understood biologically—made of human beings and exhibiting fractal patterns of living things—so the slow erosion of the Republic's bedrock principles over decades resembles the gradual, fits-and-starts process of a body dying, where the actual moment of death is an arbitrary line in a long process rather than a single event.

    “I have come to a quite biological conception of institutions. I think institutions are made up of human beings and I think that nature is filled with fractals.”

  • The Left's focus on the erosion of democracy per se is misplaced because functioning elections are the thing an administration is least likely to mess with—precisely because voting and observing who takes power is so easily verifiable—so the persistence of elections is cold comfort and not evidence the Republic is healthy.

    “It's really easy to observe. Did I go to my polling place and vote, and did the person who won get into power? And so, it's very, very hard to erode that particular thing.”

0.57

AI's foundational nature will transform institutions and check-and-balance systems, making some government role in safety necessary but private restraint preferable.

7 pointscentrality 3/5
  • Combining commercially available data sets (smartphone location, satellite, web usage) on individuals has long been legal and not classified as surveillance because it was time-intensive per person; advanced AI removes that binding constraint by scaling expert-like attention to effectively millions of analysts, so the legal term 'domestic mass surveillance' no longer matches the vernacular reality and the law has not been updated.

    “The binding constraint, though, on the use of this data is simply that it's time-intensive to actually analyze for any individual person... once you have advanced AI systems which can scale human expert-like attention infinitely, essentially, it is all of a sudden as though the intelligence community has, instead of thousands of analysts, millions and tens of millions of analysts.”

  • Nature is filled with phase transitions—like decentralized cells activating until a heart suddenly beats, or a son progressively awakening to the world—with no single defining moment, and change from AI will be like this too, arriving through phase transitions in both capability progression and adoption rather than continuous gradual change.

    “nature is filled with phase transitions... all these cells, these decentralized cells that begin to activate; and then enough of them activate, and all of a sudden you have a heart beating.”

  • Modern models like GPT 5.4 can, knowing enough about a user, advise on complex interpersonal and professional dilemmas in a way that reflects how the user sounds at their best, and Ball expects a top model like Claude Opus 4.6 would give better moral and ethical answers than a given official most of the time.

    “it actually said what I should have said. It was, like, 'No, you shouldn't say that, you should say this.' And, I was, like, 'Wow, that's really,'--like, because it knows enough about me to know what I want to sound like.”

  • Because the Republic is in poor health, Ball has trouble trusting an unfettered executive as the governing institution over AI, which shifts his weighing of private versus public power accumulation—since private corporations, unlike the state, do not hold the monopoly on legitimate violence—and he suspects we are in a new Founding moment requiring new checks and balances.

    “it affects my view of the accumulation of private power versus the accumulation of public power because the thing about private corporations is they don't have the monopoly on legitimate violence.”

  • Like the printing press, AI may be so foundational to the capabilities of organizations and institutions that it changes the institutional complex defining the technocratic nation state, challenging the status quo in which technocratic regulators control large swaths of the economy, so even classical-liberal political concepts may become outdated as a new institutional complex emerges.

    “like the printing press, may well be so foundational to the capability of organizations and institutions that it actually changes sort of the institutional complex that defines the technocratic nation state.”

  • Advanced AI lets anyone engage in object-level dangerous acts like autonomous cyber attacks without needing a military arsenal or intelligence apparatus, creating catastrophic misuse potential (including bioweapons) that implies at least some government role in mitigation.

    “I don't need to have a military arsenal to make use of these models, or an intelligence-gathering apparatus. Anyone can launch a cyber attack.”

  • While AI may eventually outperform the best humans at speech acts, strategy, and rhetoric—better strategy than Bismarck, better rhetoric than Lincoln—the human touch will persist in things definitionally based on relationships, trust, integrity, charisma, and persuasion, because politics is an inherently relational act and automating it is a category error.

    “the notion of automating politics doesn't really make sense to me. That seems like a category error.”

0.40

Private AI companies should hold safety guardrails through contractual terms, but Anthropic should have demanded statutory protections instead.

3 pointscentrality 3/5
  • 'All lawful use' is an abnormal and question-begging notion in defense procurement; a contract for a missile does not say it can be used for all lawful use, so the Department of War's position pretends contracts are something they are not, and 'lawful' is itself undefined—effectively the administration decides what is lawful until courts stop it.

    “I've talked to defense procurement and procurement law experts: This is an abnormal notion in contracting. It's sort of question-begging... a contract for a missile does not say, 'You can use the missile for all lawful use.'”

  • If the surveillance and autonomous-weapons protections matter as much to Anthropic's leadership as the red-line stance suggests, then Anthropic should have refused to sell anything to the government until Congress passed a statute writing those protections into law—and should have done so back in 2024—rather than relying on contractual terms that the government can later coerce away.

    “if they're that important, then you should have just said, 'We're not selling you anything until there's a law.' And, they should have said that in 2024.”

  • The strongest version of the opposing argument is that Anthropic is using its private power to unilaterally set what amounts to public national-security policy through contractual terms, and that a private party simply cannot make public policy through contracting regardless of market competition.

    “Anthropic is essentially using its private power to set what amounts to public policy unilaterally. And, there is some truth to that... A private party can't do public policy through contracting.”

0.22

Government choosing winners in AI on non-merit grounds undermines capital formation and America's ability to maintain technological dominance.

2 pointscentrality 2/5
  • Anthropic's Claude is currently the only large language model available for use on classified government systems, which gave the Department of War strong leverage but is not a permanent fact and could change as competitors enter.

    “it's worth noting Anthropic is the only LLM that is available to be used on classified systems.”

  • Government picking winners and losers in AI on non-merit grounds—even while Secretary Hegseth admits Claude is used 'because it's the best'—imposes regulatory risk that makes it harder for the capital-intensive AI industry, especially Anthropic, to raise needed capital, thereby diminishing America's ability to maintain its technological lead at a critical time.

    “this is an incredibly capital-intensive industry, and all of this regulatory risk is making it much harder for Anthropic in particular, and probably the industry in general, to raise the capital that they need. And so, yeah, you are diminishing America's ability to maintain its lead in this technology right at a critical time.”