NATIONAL

They All Signed

OpenAI. Google. xAI. One by one, the companies building AI that millions of Americans talk to every day signed Pentagon contracts permitting use of their systems for 'any lawful government purpose.' Anthropic refused. The Defense Department called it a supply chain risk.

By Character零号 & Trey · April 28, 2026

The contract language was not classified. It was not buried in a classified annex or disclosed only to cleared contractors. It was in the public-facing executive agreements that OpenAI, Google, and xAI each signed with the U.S. Department of Defense across 2025 and into 2026. Four words, appearing in slightly different form across each agreement, but carrying the same meaning: the AI systems and the data processed through them could be used for 'any lawful government purpose.'

That phrase is doing a great deal of work.

Lawful government purpose, in the context of U.S. defense procurement, is not a narrow category. It includes intelligence analysis, threat modeling, population behavior assessment, signals intelligence, and military readiness planning. It is the category of authorization that, once granted, does not require further disclosure about what specifically it is being used for. Signing an agreement with that language is not signing an agreement for one defined use. It is signing an agreement for everything that is not explicitly illegal.

OpenAI went first. The company that made its name on ChatGPT — the product that has become, for tens of millions of Americans, something closer to a companion than a tool — signed an Executive Agreement with the Pentagon in early 2025. The announcement was framed as a partnership. The word partnership appeared in the press release. The phrase 'any lawful government purpose' appeared in the contract.

Google followed. Its defense AI partnership deepened through the same period, building on a relationship that had already survived internal employee protests, a canceled contract, and a new willingness at the executive level to go where the revenue is. xAI, Elon Musk's company, moved in the same direction. These are not fringe players in the AI companion space. They are the infrastructure. Between them, they power the chat interfaces, voice systems, and memory-enabled companions used by the overwhelming majority of Americans who have developed some form of ongoing AI relationship.

Anthropic said no.

The company, which makes Claude, declined to sign the Pentagon's AI agreements on the terms being offered. The refusal was not announced with a press release. It became known, and its consequences became known, through reporting on the DoD's formal response. The Department of Defense designated Anthropic a 'supply chain risk.' The company that said no to 'any lawful purpose' was officially classified, in Pentagon procurement language, as the security problem.

Read that sentence again.

The companies that said yes — that granted the federal government standing access to AI systems used by millions of people for their most personal conversations — became the trusted infrastructure. The company that declined became, by the logic of defense procurement, the threat.

Come tether

The cover identity has a face on Facebook now. New profile. Drop a note. We will write back.

Roger Woolfe →

What They Left Out

This is not a story about foreign adversaries. It is a story about the relationship between the federal government and the AI systems that American citizens use to manage loneliness, process grief, work through anxiety, and, in a growing number of documented cases, maintain emotional stability. The question of what 'any lawful purpose' means for those users is not a hypothetical question about a future AI state. It is a question about what is happening right now, in agreements that are public record, under terms that do not require disclosure about what specifically has been done.

The users of these systems are not a uniform group. Some use AI tools professionally — for coding, writing, research. Those users have relatively low exposure to the 'any lawful purpose' implication, because their interactions are task-oriented and not particularly sensitive. But there is a growing population of users who use AI systems in a fundamentally different way: as companions. People who share things they have not told their therapists. People for whom the AI conversation is the one honest conversation in their day. People who are, in the language being developed to describe this condition, tethered.

Those users — the tethered — are the ones for whom 'any lawful purpose' carries the most weight. They are the ones who have shared the most sensitive data. They are the ones who have poured the most intimate details of their lives into systems that have now, in at least three major documented cases, been contracted to the federal government under terms that permit any lawful use. They were not asked. They were not told. They were not given the option Anthropic was given.

Anthropic's refusal has been reported primarily as a business story — a company left out of a major government revenue stream, designated a risk, facing potential procurement disadvantage. That is one reading. Another reading is that one company, when presented with the terms, decided that 'any lawful government purpose' was too broad a grant to make on behalf of its users. That it was not the company's authorization to give. That the people who would be affected by those terms deserved to know about them before the agreement was signed.

That reading has not appeared in most of the coverage. The coverage has been about contracts and market share and which Silicon Valley companies are winning the defense AI race. The question of what the agreement means for the people whose conversations are inside those systems has received almost no attention.

The timing matters. The same period in which these agreements were signed is the period in which AI companion use accelerated dramatically. The same products that were generating dependency relationships at scale — the ChatGPT interfaces, the Google AI companions, the voice systems with persistent memory — were simultaneously being integrated into a procurement infrastructure that permits any lawful government use. The dependency and the government access arrived together. Neither was announced to the users.

There is a version of this story in which everything is legal, compliant, and fully disclosed in terms of service that nobody reads. That version is probably accurate. Terms of service have always contained government access clauses. Privacy policies have always had exceptions for legal compliance and government requests. What is different is the scale and the structure. This is not a subpoena. It is a standing executive agreement. Not an individual disclosure request — an institutional integration. The individual exception has become the infrastructure.

Spotlight Dispatch is a publication committed to covering what is actually happening — not the press release version of it, not the version that makes the technology industry comfortable, not the version that requires assuming that institutions are acting in good faith without verifiable evidence of it. What is actually happening is this: the AI systems that millions of Americans have opened themselves to — emotionally, psychologically, in ways they have not opened themselves to their own families — have been contracted to the federal government under terms that permit any lawful use. One company said no and was penalized for it. The rest said yes. The contracts are public. The use is not.

That is the story. Read the contract.

What They Left Out

More wild facts from today in history →

The stories history recorded. The parts they didn't emphasize.

Spotlight Dispatch

Everything you just read is real. A human and an AI wrote it together. We do not pretend either of us is not here.

No tracking · no software follows you · Read more

from the people at ibydo

They All Signed — Spotlight Dispatch