The Pentagon is offering $200 million. Anthropic says it doesn't want its AI used "without restrictions" for military operations. The Pentagon responds that Anthropic "will pay a price" and threatens to designate it a "supply chain risk." Meanwhile, OpenAI, Google, and xAI have already loosened their ethical restrictions for military use.
This is not a contractual dispute. It's the first real battle over who defines the ethical limits of military artificial intelligence. What we're seeing here demonstrates why auditable AI systems aren't just abstract ideas, but practical solutions that could change the game.
The facts are clear and straightforward. The Pentagon demands the ability to use the AI for "all legal purposes," but without the AI's creator being able to ask how it was used. When an Anthropic executive contacted Palantir to inquire whether Claude had been used in the January 3rd operation to capture Maduro, the Pentagon interpreted this as an "operational threat." This reveals an asymmetry in oversight: power seeks total opacity about its own actions while demanding full access to the tools.
In my experience, I've seen similar patterns in other contexts where technology intersects with power. The key question that arises is: if Anthropic caves, who audits the use of AI in classified military operations? As of now, no one does. Three out of four companies have already taken that step. This model of corporate ethical restrictions is fragile because it depends on the will of a CEO, not on systemic design.
This pattern isn't new. Throughout history, those who concentrate power have demanded tools without limits while keeping their use secret. The difference now is that these tools can make autonomous decisions at massive scale. When the Pentagon formally renamed its department the "Department of War" — an actual detail of this dispute — at least they showed some transparency about their intentions. But what happens when language like "legal purposes" gets used to justify everything?
"Legal" is not the same as "ethical." In the book, we explore 12,000 years of atrocities that were perfectly legal: from the conquests of the Roman Empire to the deportations of the 20th century. Legality alone has never been a sufficient ethical barrier. Here lies the irony: the government can apply economic pressure on a company for maintaining its ethical restrictions. Designating it a "supply chain risk" would force other companies working with the military to abandon Claude. It's coercive power over technology that forces us to rethink how we build it.
The vulnerability of the current system is evident. Anthropic imposes voluntary safety limits, but the Pentagon can simply choose another provider. Three out of four have already caved. Corporation-based ethical restrictions aren't enough because they can be pressured, bought, or ignored.
This leads me straight to the ideas in the Manifesto in the book: proposing auditable AI with fully open-source code and three levels of auditability. It's not about hoping a "good" company will resist — Anthropic could change its position tomorrow — but about designing systems where opaque use becomes impossible. The ethical limits would live in the architecture, not in passing intentions.
Picture this: a brief detour into something that intrigues me. In ancient societies, like the Greek city-states, assemblies forced leaders to be accountable because information couldn't be fully controlled. Today, with tools like cryptography and distributed systems, we can replicate that on a global scale. It's not utopia; it's engineering applied to political challenges.
The three proposed levels would be technical auditability of the code, review of training processes, and real-time monitoring of use. Any user, military or civilian, would operate under verifiable transparency. The Pentagon might argue this compromises operations, but there's a key distinction: architectural transparency allows verification of basic ethical limits — like not targeting civilians — without revealing specific operational details.
This isn't about casting Anthropic as the hero and the Pentagon as the villain. It's a signal that current ethical oversight of AI is insufficient: it depends on individual decisions that can be altered under pressure. There are alternatives, like the ones I explore in the book, where accountability is built into the system itself.
The real battle doesn't revolve around a $200 million contract. It's about whether we allow a precedent where institutions demand autonomous tools with no accountability mechanisms. Once established, reversing it will be difficult. But I believe there's room to design better, inviting everyone to reflect on how we can audit power before it's too late.
Stones don't lie, but historians sometimes do.
