Trump’s Latest AI Contract Clause Scraps Crucial Safeguards - Jacobin

The Trump administration is trying to turn one of the ugliest parts of its fight with Anthropic, the AI safety company behind Claude, into baseline procurement policy for the rest of government. In a nine-page draft clause from the General Services Administration, the agency says AI vendors must give the government an irrevocable license to use their systems for any lawful government purpose and must not refuse outputs based on the vendor's own discretionary policies.

That would already be a major escalation from the Pentagon-specific dispute now being litigated by Anthropic. But the stranger part is what sits around it. The same GSA draft clause says the procurement language overrides conflicting commercial terms, requires vendors to summarize retrieval, routing, and reasoning processes in some cases, and lets the government benchmark systems for “unsolicited ideological content” using methods it does not have to disclose. If you were looking for a clean procurement rule about data rights or vendor lock-in, this is not that. This is governance by everything bagel.

The practical significance is that the Anthropic fight stops being just about Anthropic once this language enters the General Services Administration's acquisition machinery. As FedScoop reported, the clause was floated through the Multiple Award Schedule refresh process rather than ordinary notice-and-comment rulemaking, then drew enough industry blowback that GSA extended the comment deadline to April 3 and pushed consideration from Refresh 31 to Refresh 32. For a clause this sweeping, lawyers quoted by FedScoop called the process unusual to the point of alarm.

The core legal move is blunt. Under the draft, if a contractor's policies, terms, conditions, or other commercial agreements conflict with the clause, the clause controls. That is a direct shot at the way frontier AI companies usually enforce guardrails: not by changing model weights every time, but through system policies, use restrictions, and product terms layered on top of the model. Dean Ball, a policy analyst at the Foundation for American Innovation who said in a public comment that he previously served as a senior White House AI adviser, argued in a Foundation for American Innovation post and a formal comment letter that this makes the clause “unworkable and legally unstable” because it pretends to preserve model weights while attacking the policy layer that actually keeps systems usable and safe.

That critique matters because Ball is not coming from the usual civil-liberties flank. He broadly supports stronger government AI procurement and still thinks this mechanism is unstable. Jessica Tillipman, an associate dean at George Washington University Law School, was even more caustic in Lawfare, calling the proposal “governance by sledgehammer.” Her point was not that government should ignore AI procurement risks. It was that GSA is trying to solve data rights, interoperability, ideology, safety, and supply-chain politics in one contract clause, which is a good way to produce litigation before it produces clarity.

The ideological language is where the draft starts to feel less like ordinary procurement hardening and more like an attempt to regulate model behavior through purchasing power. The clause's “unbiased AI principles” say systems must be truthful and neutral, then single out “ideological dogmas such as Diversity, Equity, Inclusion” as something outputs must not favor. The government also reserves the right to benchmark vendors for ideological content using undisclosed methods, according to the GSA text. Raised eyebrow here: a procurement term is being asked to do culture-war content moderation, model auditing, and commercial preemption all at once.

This is why the draft matters beyond process nerds and government contracts lawyers. The frontier model vendors selling into government — Anthropic, OpenAI, Google, Microsoft, and the integrators built on top of them — do not just ship raw weights. They ship layered systems with refusal logic, safety classifiers, retrieval pipelines, logging controls, and often globally shared product terms. A rule that says the government can demand lawful-use access while overriding conflicting vendor policies is a rule that reaches down into how those systems are actually operated.

It also makes the bridge to Anthropic's current lawsuit much more explicit. In its federal complaint, Anthropic alleged the Pentagon demanded contract terms that would force Claude to support all lawful uses, including uses the company says would erase its bans on mass surveillance of Americans and lethal autonomous warfare. The new GSA draft does not resolve that fight. It republishes its logic in a form other agencies could reuse.

That is the genuinely new fact pattern here, and it is why the early coverage from Jacobin and reporting cited there from The Lever landed. What had been a contested term in a politically charged Pentagon standoff is now visible as a proposed purchasing template for much broader government AI buying.

There are real procurement problems buried in this draft: data ownership, portability, incident reporting, and dependence on opaque vendors are not invented concerns. But the clause reads like an administration trying to solve all of them in one swing, while also taking a shot at the policy guardrails vendors use to keep their systems inside their own risk tolerance. That might prove politically attractive. It does not yet look technically clean, legally stable, or easy to implement.

What to watch next is simple. First, whether GSA materially rewrites the clause after comments close on April 3, a date confirmed by FedScoop. Second, whether other agencies start copying the language before any broader rulemaking. And third, whether courts looking at Anthropic's case treat this as evidence that the administration's position was never a one-off contracting dispute at all, but an attempt to normalize federal access to AI systems on terms the vendors themselves would not otherwise accept.