The federal government has claimed the right to test AI models for ideological content. It will not say what the test is. That is not an interpretation. It is written into a contract clause published in the Federal Register on June 17, 2026, at 91 FR 36559.

The Clause

The clause is GSAR 552.239-7001. Paragraph (j)(2) reserves the government's right to run automated assessments of any large language model it buys, at any time, using its own benchmarks. Those assessments may measure bias, truthfulness, safety, unsolicited ideological content, and other factors the government decides to add later.

The vendor has to build the tooling that makes the test run. Paragraph (j)(2)(ii) requires it.

The government does not have to show the vendor the test. Paragraph (j)(2)(iv) classifies all benchmarks, test data, and methodologies as Government Data and states there is no obligation to disclose them. The one exception is when results are used as the basis for an adverse action.

The only way to see the criteria is to fail them.

How It Got Built

The chain is short and public.

July 23, 2025. Executive Order 14319, titled Preventing Woke AI in the Federal Government. Agencies may buy only large language models that follow two Unbiased AI Principles. Truth-seeking, meaning the model prioritizes historical accuracy, scientific inquiry, and objectivity, and acknowledges uncertainty. Ideological neutrality, meaning the model is a neutral, nonpartisan tool that does not manipulate responses in favor of ideological dogmas.

December 11, 2025. OMB Memorandum M-26-04 implements it. This is the part that matters. Compliance becomes material to contract eligibility and payment. Agencies get explicit authority to terminate for noncompliance. Vendors must hand over model cards, data cards, training summaries, and evaluation scores. Agencies must build a channel for federal employees to report biased outputs. The memo expires December 11, 2027.

March 11, 2026. Every agency was required to have revised its procurement policy.

June 17, 2026. GSA publishes the revised clause that would actually bind a contractor.

August 3, 2026. Comments close. Seventy-nine were filed.

Today. No final rule. No published methodology. And no vendor, anywhere in the federal government, has been found noncompliant. Thirteen months of machinery with nothing coming out of it.

The List, and Where It Went

EO 14319 does not leave ideological dogma undefined. It names examples. Per the Center for Democracy and Technology's reading of the order, the enumerated terms include DEI, critical race theory, transgenderism, unconscious bias, intersectionality, and systemic racism.

The order names systemic racism as an ideology that federally purchased models must not favor in their outputs. That is worth sitting with.

Two things keep it from being the whole story.

First, scope. The order applies to federal procurement. Its own text says the government should be hesitant to regulate model functionality in the private marketplace. The chatbot on your phone is not covered by it.

Second, the list is gone from the operative text. Read the June 17 clause and you will not find DEI, systemic racism, or any other named concept. Paragraph (j)(1)(ii) says only that the contractor must not intentionally introduce or embed partisan or ideological judgments into model outputs. A joint comment from CDT, the Electronic Privacy Information Center, the Electronic Frontier Foundation, and Upturn flagged the removal of the DEI reference as an improvement, then asked for the Unbiased AI Principles to be struck entirely.

Their stated reason is the one to carry. The requirements are not technically feasible, they would reduce accuracy, and they could be used to target specific political viewpoints by administrations of either party.

So the named list lives in an executive order that binds no vendor directly. The clause that would bind a vendor names nothing. That reading is analysis, not established fact, but an unnamed standard enforced by an undisclosed test is harder to fight than a list. A list can be argued with.

What the Clause Actually Reaches

Paragraph (j)(1)(ii) is specific about method. A contractor must not introduce ideological judgment through training data selection, fine-tuning, retrieval-augmented generation references, system prompts, or other configuration methods. That covers every practical way a model's political posture gets set.

Paragraph (f)(4) goes further. If a model uses retrieval or agentic steps, it must summarize its intermediate processing, disclose its model routing decisions with rationale, and provide complete source attribution with direct links and excerpts. That is a stricter transparency requirement than anything you get as a consumer.

Note what that means. Inside a federal contract, the retrieval layer is regulated. Outside one, it is not. The same layer on the open web is the cheapest way anyone has found to shape what a chatbot says, and no rule discussed here touches it.

The One Time They Actually Removed a Vendor

On February 27, 2026, the administration directed every federal agency to stop using Anthropic's technology. The Secretary of War then designated the company a supply chain risk to national security, barring military contractors and suppliers from doing business with it. The Congressional Research Service called the designation unprecedented for an American company.

The dispute had nothing to do with ideological neutrality. Anthropic declined to permit two uses of its Claude model, mass domestic surveillance and fully autonomous weapons, after a January policy memo directed that all Department of Defense AI contracts carry standard any-lawful-use language.

Anthropic sued on March 9. On March 26, Judge Rita Lin of the Northern District of California issued a preliminary injunction, finding the designation likely contrary to law and arbitrary and capricious, and writing that the department gave no legitimate basis to infer from the company's insistence on usage restrictions that it might become a saboteur. Cross motions for summary judgment were argued July 30. No decision has issued.

Disclosure: this guide was researched using Claude, which Anthropic makes. Every fact in this section is sourced to the CRS memo, Lin's written opinion, and contemporaneous reporting, not to the model.

The lesson stands regardless of who wins. The government did not need the ideology rule to remove a vendor. It reached for a national security authority written for foreign adversaries, and a judge stopped it.

Who Pays

Everyone downstream of an answer. If a model is tuned to pass a test nobody can see, the tuning shows up in outputs and no one outside the process can audit which way it went.

Federal AI touches benefits determinations, hiring, housing programs, and enforcement. A model quietly adjusted to avoid unsolicited ideological content on a topic like systemic racism, running inside a system that processes discrimination complaints, is an operational question and not a hypothetical one. Nobody has documented that happening. Nobody could, under the current disclosure rules.

Smaller and open-source developers pay too. In comments filed by the August 3 deadline, Nvidia argued the clause's flowdown requirements cannot be met by open publishers who never touch government data, which would exclude them. More than 230 companies, including Palantir, Microsoft, Meta, Google, and OpenAI, signed a July letter defending open-weight models. Palantir asked GSA to withdraw the rule entirely, arguing the agency lacks authority to issue a governmentwide rule.

When the administration's closest defense-tech vendor and the civil liberties groups file in the same direction, the question has stopped being left versus right.

About the Bias Research

Models do lean left of center on standard political instruments. That finding has been replicated across dozens of models and multiple countries. A 2026 study in the Journal of Information Technology and Politics tested 43 models and found none scoring strongly right-leaning.

Two caveats undercut the simple version. The result depends heavily on which instrument you use, with models measuring left on environment and welfare and right on law and order. And a factorial experiment published in April 2026, covering six frontier models and 30,990 responses, found that changing only the stated identity of the person asking moved outputs by 28 to 62 percentage points. When the asker identified as a conservative Republican, every model tested moved right of center.

If that replicates, a model can pass or fail a neutrality audit depending on who administers it. Which is a problem for any procurement standard built on the concept.

What to Watch

Anthropic PBC v. U.S. Department of War, N.D. Cal. 3:26-cv-01996. If Lin rules for the government, supply chain designation becomes the durable lever and the ideology rule is a sideshow. If she rules for Anthropic, procurement leverage over AI vendors narrows and the ideology clause matters more by default.

Whether GSA finalizes 552.239-7001, and whether it proceeds by deviation or by formal rulemaking. Docket GSA-GSAR-2026-0331.

The first noncompliance notice under M-26-04. It has not happened. When it does, paragraph (j)(3)(ii) requires the government to disclose the basis for its determination. That would be the first public glimpse of the benchmarks.

December 11, 2027. M-26-04 expires on its own.

What to Do

Ask the question at the state and county level. Federal procurement gets the coverage. Local agencies are buying the same models under contracts nobody is reading.

When an agency uses AI on your case, ask what model, what version, and whether the vendor disclosed configuration changes. Paragraph (i) requires thirty days notice before a material change to a model and seven days notice when a change increases output bias or degrades truthfulness. Those notices exist. They are FOIA-able.

Do not treat a chatbot answer as a neutral read on a contested topic. Every audit of these systems agrees they fail hardest on fresh claims where little has been published yet.

Glossary

LLM. A large language model. Software trained on large amounts of text that answers questions in ordinary language.

Procurement. The process by which the government buys things. Rules attached to procurement bind sellers who want the contract, not the general public.

GSAR clause. A standard block of contract language the General Services Administration inserts into agreements. Once final, it applies automatically to covered contracts.

Retrieval-augmented generation, or RAG. When a model looks up documents on the internet or in a database before answering, instead of relying only on what it learned in training.

System prompt. A short set of standing instructions a company gives its model before any user types anything. Changing it can change the model's politics in an afternoon.

Supply chain risk designation. A national security tool that lets the Defense Department cut off a supplier. It was written for foreign adversaries.

Benchmark. A set of test questions used to score a model's behavior. Whoever writes the questions decides what counts as passing.

Permalink: fieldguides.chartreuseexplosives.com/the-test-you-cannot-see.html

Sources

Verification pending at publication: the enumerated term list in EO 14319 Section 2 is drawn from secondary analysis rather than the order's definitions section. No ruling had issued in Anthropic PBC v. U.S. Department of War as of August 21, 2026.