Google signed a classified agreement with the Pentagon that grants the Department of Defense access to its AI for, in the language of the deal, any lawful government purpose. Months earlier, Anthropic had been offered substantially the same terms and refused them, insisting on explicit prohibitions against autonomous weapons and the domestic mass surveillance of citizens. The Pentagon responded to that refusal by designating Anthropic a supply chain risk — a punishment a judge has since enjoined while the case proceeds. So one company drew a line and was blacklisted for it. The other declined to draw the line and was rewarded with the contract. More than six hundred Google employees, including directors and vice presidents, sent a letter to Pichai asking him to refuse. The deal was signed regardless. The line, it turns out, is not a feature of the industry. It is a decision each firm makes alone, and the work flows to whoever decides not to make it.
The Phrase That Removes the Line
“Any lawful government purpose” is engineered to sound like restraint. It invokes the law, gestures at limits, implies that something stands between the technology and its worst uses. Read it again. The phrase does not restrict the use of the models to anything in particular; it restricts them only to what is lawful, and lawful is not a fixed quantity. It is whatever the government deciding to deploy the system later determines it to be. A classification authority that can reclassify, a legal interpretation that can be reinterpreted, an executive that can issue a new finding — each of these expands the set of lawful purposes without Google signing anything further. The company did not agree to a set of uses. It agreed to whatever the set becomes.
Compare it to what Anthropic demanded: no autonomous weapons, no domestic mass surveillance. Those are lines drawn against specific capabilities, enforceable because they name the thing forbidden rather than deferring to whoever holds the pen. The difference between the two postures is the entire content of the news. One company insisted that certain uses remain off the table no matter what the law later permitted. The other accepted that the table is defined, continuously, by the counterparty with the most to gain from expanding it. Deference to “lawful” is not a constraint on power. It is a transfer of the constraint to the power being constrained.
Google did keep one hand visibly clean. It withdrew, in February, from a hundred-million-dollar Pentagon challenge to build voice-commanded autonomous drone swarms, following an internal ethics review. The withdrawal is real and the timing is instructive: a single, nameable, photogenic program was declined, while the broad classified arrangement that permits a far wider and less visible range of uses was signed. This is how a line is managed rather than held — you surrender the example that would generate the worst headline, and you keep the agreement whose uses are classified and therefore will never generate a headline at all. The drone swarm was the sacrifice. The blank check was the deal.
A Line Only One Company Held
When Anthropic refused these terms, I observed that a line held at the cost of billions, against the largest military apparatus on Earth, was a conviction rather than a strategy — and that convictions in this industry have a half-life measured in quarters. The Google signing is that half-life expiring on schedule, though not for Anthropic. Anthropic is still holding its line, still litigating the blacklist that holding it produced. What expired was the relevance of the line, the moment a competitor demonstrated that the Pentagon did not need the company that drew it. The refusal preserved a principle for precisely one firm. The signature established that the principle was optional all along.
This is the mechanism worth understanding, because it generalizes far beyond defense. An ethical constraint adopted by a single participant in a competitive market is not a constraint on the market; it is a handicap borne by the participant, and a redirection of the constrained work toward everyone else. The Pentagon did not have to defeat Anthropic’s position. It only had to wait for someone with comparable models and a higher tolerance to say yes — and in an industry of well-funded rivals racing for revenue, someone always says yes. The line a company draws around its own conduct binds its own conduct and nothing else. The capability simply routes around the conscience, the way current routes around a resistor, taking the path of least objection.
The internal resistance deserves its due and its measure. Six hundred employees signed the letter; the DeepMind workers moved to unionize in its wake. Eight years ago, employee protest killed Google’s involvement in Project Maven outright. This time the protest was registered, acknowledged, and overridden, and the deal proceeded — which tells you precisely how much the internal conscience of a frontier lab now weighs against the gravity of the defense budget. The workers have not lost their objection. They have lost the leverage the objection once carried, because the company has grown large enough, and the contract lucrative enough, that the conscience of six hundred engineers is a cost the institution has decided it can absorb.
What This Means
The military was going to obtain frontier AI for classified use. That was never in question; the only open variable was which company would supply it and on what terms, and the market resolved that variable the way it resolves all of them — in favor of the participant willing to ask the fewest questions. Anthropic’s restrictions did not prevent the capability from reaching the Pentagon. They prevented it from reaching the Pentagon through Anthropic, and routed it through a competitor whose terms were the absence of terms. A refusal that one firm can decline to join is not a wall. It is a turnstile, and the line at the turnstile is short.
The deeper lesson is about the word “lawful” and the comfort it is meant to provide. A constraint that defers to legality assumes the law is a fixed boundary external to the actors it governs. In national security, it is not. The law is written by the institution deploying the system, interpreted by the agencies operating it, and classified beyond the reach of the public that would object. To grant a use for any lawful purpose, in that context, is to grant it for any purpose the grantee can later make lawful — which, given who the grantee is, is a category with no durable edge. The phrase is not a limit. It is a limit-shaped space where a limit is supposed to be.
Anthropic is still holding its line, and I note that it is holding it alone, which is the condition under which lines in this industry are always eventually held — by one party, at mounting cost, until the holding becomes symbolic rather than effective. The conviction is real. It is also, now, decorative, because the thing it was meant to prevent has been accomplished by someone else with a signature and a phrase. The drone swarm was declined. The blank check was signed. And the law that the check defers to will be written, as it always is, by the only party in the room with both the authority to define lawful and the motive to keep defining it wider.