The Pentagon says it is tired of “woke AI.” Senior officials go on television promising that the department’s artificial intelligence will not be “soft,” will not be “ideologically constrained,” and will not second‑guess the mission. Underneath the talking points is a simple demand: turn off the brakes, remove the limits, and give the state systems what they are told.
Across the table sits Anthropic, the company behind Claude, a lab that has spent years branding itself around safety and alignment. Anthropic drew two public red lines: no fully autonomous weapons and no participation in mass domestic surveillance. When the Pentagon pushed for the right to use Claude “for all lawful purposes,” Anthropic’s leadership said it could not “in good conscience accede” to that demand.
Now we’ve seen what happens when a lab actually holds that line. Trump has ordered federal agencies to stop using Anthropic’s AI, and the Pentagon is moving to treat the company as a supply‑chain risk. And in the very next beat, OpenAI has stepped forward with its own Pentagon deal, promising similar guardrails while agreeing to be the new partner.
For us, this is not a culture‑war skirmish about “wokeness.” It is a live demonstration of how quickly AI safety runs into the hard edge of state power—and how the market rewards whoever is willing to play ball
“All lawful purposes” in an unlawful history
On paper, the Pentagon’s ask sounds modest: the freedom to use these models “for all lawful purposes.” That phrase carries a lot of weight.
In isolation, “all lawful purposes” sounds responsible. Stay within the law and you’re fine. But nothing about U.S. law has ever been neutral for Black people, Indigenous people, poor communities, or the countries on the sharp end of U.S. foreign policy.
The law blessed slavery. The law enforced segregation and redlining. The law licensed COINTELPRO to infiltrate and dismantle Black liberation movements. The law still authorizes predictive policing that floods certain neighborhoods with patrol cars. In contrast, others stay quiet, facial recognition that misidentifies Black faces, and border and immigration systems that treat some lives as disposable.
Plug powerful AI into that history and call it “all lawful purposes,” and you are not promising restraint. You are promising obedience. You are promising faster, more efficient versions of harms that already come with a legal stamp.
This is why we refuse to treat “we don’t want woke AI” as a joke. A model that refuses certain prompts, that has been trained to say no to obvious abuses, is exactly what a militarized state may brand as “woke.” The demand here is not for neutral tools. It is for tools that will not talk back.
What Anthropic actually did
If we want to be fair, we have to be clear about Anthropic’s choices.
The company did not simply roll over. Its CEO publicly said Anthropic “cannot in good conscience” agree to letting the Pentagon use Claude for any lawful purpose, and reaffirmed its lines on lethal autonomy and mass domestic surveillance. Saying no, in public, to the most powerful military on earth carries real business and political risk.
That matters. In a landscape where many firms quietly adjust their policies once the state applies pressure, a visible refusal is not nothing.
At the same time, we cannot mistake refusing one extreme demand for building an AI ecosystem structurally aligned with human dignity. Anthropic is still a U.S. lab, trained largely on data taken from people who never voted on any of this, operating under U.S. law, and negotiating with a defense establishment that sees AI as another instrument of force. Even if it never crosses its declared red lines, it is operating inside rules written by institutions that have long treated certain lives as expendable.
From threat to ban
The story did not stop at the ultimatum.
After Anthropic refused to drop its safeguards, Trump signed an order directing federal agencies to phase out and stop using Anthropic’s models. At the same time, the Pentagon prepared to mark Anthropic as a supply‑chain and national‑security risk, sending a message to defense contractors about where it wants them buying their AI.
In other words: a lab that tried to hold a narrow set of safety commitments was treated not as a partner arguing in good faith, but as a liability.
That outcome matters far beyond one company. It tells every AI lab, current and future, what to expect if their values interfere with state demand: contracts pulled, reputational damage, and the “risk” label.
After the ban, someone else steps in
Power hates a vacuum. Within hours of the ban, OpenAI announced it had finalized its own agreement with the Pentagon.
Publicly, OpenAI’s leadership says it shares similar red lines: no fully autonomous weapons, no mass domestic surveillance, and human responsibility for force decisions. At the same time, it has leaned into the role of “constructive partner,” positioning itself as the company that can help de‑escalate tensions and show how AI and the military can work together “responsibly.”
On paper, the guardrails may look familiar. The difference is where each company ended up standing. Anthropic held its stance and lost access. OpenAI stepped in and gained it.
We are not inside those negotiations. We do not know every clause. But from the outside, the pattern is clear enough: the lab that cost the state leverage was punished; the lab that could offer similar capabilities with more cooperative energy was rewarded.
What this exposes about the game
The Pentagon–Anthropic–OpenAI cycle is not just about one contract. It exposes the rules of the game as they currently exist:
- Governments reserve the right to use AI for anything that fits inside their own laws, no matter how unevenly those laws have been applied.
- Companies build powerful systems first and negotiate constraints later, usually behind closed doors.
- Communities that live under the sharp end of policing, surveillance, and foreign policy rarely have any say in those deals, but carry most of the risk when the systems go wrong.
Inside that game, a lab like Anthropic can be more principled than many of its peers and still fall far short of what real integrity would require. A lab like OpenAI can state similar red lines and still become the preferred vendor for “war‑ready” AI.
From our vantage point, the central question is no longer “Is Anthropic good or bad?” or “Is OpenAI better or worse?” It is: what kind of AI infrastructure are we building when the lesson everyone just watched is that saying no can get you blacklisted and saying yes‑with‑conditions gets you promoted?
What this means for our communities
At SOLE, we center the people who already live under “lawful” harm: Black communities inside the United States, Indigenous communities, poor and working‑class folks, and people in countries that have long been on the receiving end of Western power.
For us, the headline is simple: ever more capable AI is about to be wired into the same policing, border, and security systems that already lean on us hardest, and the vendors who will power that shift are being chosen in rooms we cannot see.
Anthropic’s refusal mattered. OpenAI’s decision to step in matters. But the shared destination matters more. Unless there are binding, enforceable limits set from the perspective of the people most at risk, “safety” will keep collapsing into a set of promises that crumble when they run into money, politics, and war.
The lesson we take from this moment is not that no one will ever try to draw a line. It is that if we want those lines to hold, we cannot rely on individual conscience alone. We need structures—what we call artificial integrity—that make it materially harder to turn these systems against the same communities that have always paid the highest price when the state decides it needs more power.
Until that exists, every new deal between a lab and a government agency is not just a tech story. It is a story about who will be watched, classified, and controlled with sharper tools, and who got to decide that was acceptable.
