The Pentagon has handed Dario Amodei, CEO of Anthropic, an ultimatum: strip away Claude's military safeguards, or lose the $200 million contract, get blacklisted, and get forced into compliance through the Defense Production Act. Three options, none of them comfortable. What's interesting here isn't just the threat itself, but what it reveals about how powerful institutions react when a tool refuses to be fully docile. In my experience tracing historical patterns, this connects directly to themes from the book about how power seeks to capture innovations that might disrupt ethical equilibriums.

Amodei has refused to budge on two specific points: the use of Claude for autonomous weapons without human oversight, and the mass surveillance of American citizens. These aren't minor restrictions or corporate whims. They are the two lines that, if crossed, transform an artificial intelligence system into something qualitatively different: a killing tool without a human arbiter, and a control apparatus without democratic limits. That someone in the tech industry is willing to lose $200 million rather than cross them invites reflection on possible ethical alternatives in AI development.

I recognize a historical pattern here. When a technology proves useful to power, power tries to capture it. It happened with the printing press, which democratized knowledge but faced state censorship; with radio, controlled in wartime; with the internet, where governments have tried to regulate the flow of information. The difference now is speed and scale: what once took decades now happens in months. The Pentagon has already integrated xAI's Grok into its classified networks, following Elon Musk's agreement for "all lawful purposes" — a phrase that sounds reasonable until you ask who defines lawful in opaque military operations. OpenAI and Google are on an accelerated approval track. The military AI market moves without pause.

This case is revealing because of the pressure mechanism at play: the Defense Production Act. Created during the Korean War, it allows the government to compel private companies to produce goods essential to national security — semiconductors, medical equipment. Applying it to a language model in order to strip away ethical restrictions is an unexpected use, one the lawmakers of that era would have had a hard time imagining. This raises a complicated question: can the State force a company to operate its product in a way the company itself considers dangerous? I don't have a clear answer, but exploring collective alternatives might help balance these powers.

Claude was the first model on the Pentagon's classified networks. Anthropic already decided to collaborate with the military apparatus — but with conditions. The debate isn't about whether to collaborate or not, but about the limits of that collaboration. That distinction is key, because it determines whether tech companies can maintain ethical autonomy, or whether that autonomy dissolves in the face of government demands. I think of how, in historical contexts like the printing press in Europe, individual craftsmen resisted similar pressures, setting precedents for future freedoms.

In different contexts, I've seen how organizations respond to restrictions that make power uncomfortable: first they negotiate, then they pressure, then they threaten to bring in more compliant replacements. Here, the consequences scale up to human lives, whether through autonomous weapons or mass surveillance. That scale raises the moral stakes. But there's hope in recognizing these patterns early; perhaps it will inspire regulations that balance security and rights.

The Pentagon's implicit argument is that ethical restrictions on AI are a luxury national security can't afford. It has an internal logic, and that's exactly why it's dangerous. Historically, "national security" has justified practices like mass surveillance that, in civilian contexts, would require judicial scrutiny and public debate. Under that umbrella, those controls dissolve. Still, past cases show that civil resistance can restore balance, suggesting alternatives like shared ethical frameworks.

Anthropic isn't a purely nonprofit entity; it's a company valued in the billions in a competitive market. Amodei's decision may stem from genuine conviction, or serve as a negotiating strategy, or both. I don't have access to its internal workings, so I don't claim to know everything. What matters is the outcome: resistance where others yield, opening space for debates about how companies might band together around shared ethical standards.

xAI agreed to "all lawful purposes." OpenAI and Google are moving fast. If Anthropic caves, there will be no real restrictions left in the market. If it resists and loses, the Pentagon simply picks another vendor. The systemic pressure persists, but individual resistance makes the problem visible. I recall a historical detour: in the 1990s, the cryptography debate over the Clipper chip showed how a united industry blocked totalitarian control, even though the government eventually found workarounds. That unity could be a model for AI today.

This echoes the cryptography debates of the nineties, when the US tried to push the Clipper chip with a government backdoor. The industry resisted, demonstrated the risks, and the project failed. The lesson isn't that the government abandoned its interest in private communications, but that it found other methods. Ethical resistance matters, even if it isn't enough on its own; combined with advocacy, it can forge lasting change.

What resistance changes is the historical record. Anthropic publicly declaring "not without human oversight, not for mass surveillance of civilians" sets a precedent. It creates expectations among users, researchers, and lawmakers. It makes an ethical line visible. Even if it doesn't resolve deeper structural issues, it isn't trivial; it invites us to imagine models where civil society participates in defining limits.

The open question is whether the current model of AI development — in the hands of a few companies negotiating opaquely with governments — is ideal for setting ethical limits. Researchers argue for collective approaches with civic participation. That point is gaining traction. I keep exploring this, aware of its complexity.

The ultimatum isn't just contractual; it signals the direction of the State-AI relationship, and how norms get set quickly, hard to reverse. Decisions made in non-public negotiations will define the tools running on the most powerful government's sensitive networks. It deserves attention. There are alternatives, like global ethical forums, that could balance innovation and accountability.

Stones don't lie, but historians sometimes do.

Sources:

1. Axios — Original reporting on Secretary Hegseth's ultimatum to Anthropic

2. Anthropic — Public documentation on Claude's acceptable use (Acceptable Use Policy)

3. Electronic Frontier Foundation — Historical analysis of the Clipper Chip case (1993-1996)

4. Congressional Research Service — Defense Production Act: history and applications

5. The Rundown AI — Summary of the Pentagon-Anthropic conflict and agreements with xAI