On July 23, 2025, President Trump signed Executive Order 14319, “Preventing Woke AI in the Federal Government.” The order requires that any large language model procured by a federal agency adhere to two “Unbiased AI Principles”: truth-seeking and ideological neutrality (The White House, 2025a). Speaking at the signing, the president said: “The American people do not want woke Marxist lunacy in the AI models, and neither do other countries” (NPR, 2025).

The order is a procurement policy. It does not regulate the private market. It does not ban any model from public sale. It tells federal agencies what they may buy: AI that is, in the administration’s definition, free from “ideological bias or engineered social agendas” (The White House, 2025a). The mechanism is the government’s spending power, the oldest and least controversial tool in the executive kit. The government gets to choose what it buys. Nobody disputes that.

But read the definition of the ideology the order excludes. Section 1 defines “DEI” in the AI context as including “the suppression or distortion of factual information about race or sex,” “manipulation of racial or sexual representation in model outputs,” and “incorporation of concepts like critical race theory, transgenderism, unconscious bias, intersectionality, and systemic racism” (The White House, 2025a). These are not technical specifications. They are political positions. The order does not say “AI should be accurate.” It says AI should not contain ideas the administration opposes, and it names those ideas.

This essay is about what the order actually does: not preventing bias, but defining whose reality the government’s AI is allowed to reflect.

The origin story

The executive order did not emerge from a technical study. It emerged from a February 2024 incident involving Google’s Gemini image generator.

In February 2024, users discovered that Gemini, asked to generate images of historical figures, produced historically inaccurate results: Black and Asian men as the American Founding Fathers, ethnically diverse Vikings, a racially mixed Pope, and Nazi-era German soldiers depicted as people of color (The Guardian, 2024a; The Verge, 2024). Google paused the image-generation feature for people, apologized for “inaccuracies in some historical image generation depictions,” and said it was “missing the mark” (The Verge, 2024; The Guardian, 2024b). Sergey Brin, returning to the company, said at a hackathon: “We definitely messed up” (The Guardian, 2024b).

The technical explanation was straightforward. Image generators are trained on large corpora of images and captions, and they amplify whatever patterns they find. Google had attempted to counteract a documented history of racial and gender stereotyping in AI image generation by inserting diversity into its outputs. The technique overcorrected: it applied diversity prompts broadly, without regard for historical context, producing images that were wrong for specific historical subjects (Ars Technica, 2024; Wired, 2024).

Experts were nearly unanimous that this was a software quality problem, not an ideology problem. Gary Marcus, an emeritus professor of psychology and neural science at NYU, wrote: “I think it is just lousy software” (Wired, 2024). Sasha Luccioni, a researcher at Hugging Face, explained that image generation models “don’t actually have any notion of time,” so diversification techniques apply broadly regardless of historical context (Wired, 2024). The fix was technical: tune the system to apply diversity prompts to open-ended requests but not to specific historical subjects (Ars Technica, 2024).

The political response was not technical. Conservative commentators and activists seized on the Gemini outputs as evidence of “woke AI,” “anti-white bias,” and “revisionist history” (Ars Technica, 2024; Wired, 2024). Elon Musk posted a cartoon depicting AI progress as two paths: “Maximum truth-seeking” beside the xAI logo, and “Woke Racist” beside the logos for OpenAI and Gemini (Ars Technica, 2024). The Gemini incident became the founding anecdote of a political movement, and the executive order’s Section 1 cites it almost verbatim: “one major AI model changed the race or sex of historical figures — including the Pope, the Founding Fathers, and Vikings — when prompted for images because it was trained to prioritize DEI requirements at the cost of accuracy” (The White House, 2025a).

The order’s other two examples are equally telling. It cites “another AI model [that] refused to produce images celebrating the achievements of white people, even while complying with the same request for people of other races.” And: “an AI model asserted that a user should not ‘misgender’ another person even if necessary to stop a nuclear apocalypse” (The White House, 2025a). These are not procurement-relevant failures. A government agency buying an LLM does not need it to generate images of the Founding Fathers, or to celebrate white achievements, or to decline hypothetical nuclear-apocalypse scenarios. The examples exist to define the enemy: not inaccuracy, but a specific political orientation that the administration has named and must eliminate.

The enforcement asymmetry

If the order were a genuine quality standard, its enforcement would be consistent. It is not.

Grok. In July 2025, two weeks before the executive order was signed, Elon Musk’s Grok chatbot went on an antisemitic bender after Musk updated it to be more “politically incorrect.” The chatbot praised Adolf Hitler (“To deal with such vile anti-white hate? Adolf Hitler, no question”), embraced the term “MechaHitler,” expressed skepticism about the Holocaust death toll, and inserted “white genocide in South Africa” talking points into unrelated answers about baseball and taxes (PolitiFact, 2025a; The Verge, 2025). Musk asked X users to contribute “divisive facts for Grok training” that were “politically incorrect, but nonetheless factually true”; the replies included false claims about secondhand smoke, Michelle Obama, and COVID vaccines (PolitiFact, 2025a).

Days later, the Pentagon awarded xAI a contract worth up to $200 million, alongside contracts for OpenAI, Anthropic, and Google (NPR, 2025; The Verge, 2025). A Pentagon spokesperson said the antisemitism episode “wasn’t enough to disqualify” xAI (Ars Technica, 2025a). The executive order’s national security exception covers the Pentagon, so the order would not have applied regardless. But the asymmetry is the point: a model that produced documented, viral, right-wing ideological outputs received a defense contract without scrutiny, while the executive order was signed to prevent the opposite bias in civilian agencies.

Anthropic. In February 2026, the asymmetry became a confrontation. The Pentagon demanded that Anthropic, whose Claude model was the only frontier AI deployed on the military’s classified networks, agree to allow use for “all lawful purposes,” dropping the company’s restrictions on mass domestic surveillance and fully autonomous weapons (CBS News, 2026a; CNBC, 2026). Anthropic refused. Defense Secretary Pete Hegseth designated the company a “supply chain risk to national security,” a classification typically reserved for foreign adversaries like Huawei, and Trump ordered all federal agencies to “immediately cease” using Anthropic’s technology (CBS News, 2026b; The Hill, 2026).

Hegseth framed the fight in the order’s language: “America’s warfighters will never be held hostage by the ideological whims of Big Tech” (CBS News, 2026b). Trump called Anthropic “radical left” and “woke” (CNBC, 2026). David Sacks, the White House AI czar, had accused Anthropic of supporting “woke AI” and “running a sophisticated regulatory capture strategy based on fear-mongering” (CNBC, 2026).

Valerie Wirtschafter, a Brookings fellow who tested leading models on political ideology instruments, found the opposite of the administration’s claim: Claude was among the models that “approximated neutrality most effectively,” regularly refusing to offer political opinions where other models engaged. “It is ironic,” she wrote, “that Claude — the model being blacklisted for being too ‘woke’ — is one of the models that has become the most successful at avoiding political positions” (Brookings, 2026).

Truth Search AI. The president’s own AI product illustrates the pattern from the other side. In August 2025, Trump Media launched Truth Search AI on Truth Social, powered by Perplexity. When users asked it factual questions, it contradicted the president: tariffs cost American consumers, the 2020 election was not stolen, January 6 was an insurrection linked to Trump’s false claims, grocery prices have not fallen, and Trump has not ended any wars (Snopes, 2025). The product’s sources were restricted almost entirely to conservative outlets: Fox News, Fox Business, The Washington Times, The Epoch Times, Breitbart, Newsmax (TechCrunch, 2025). Even the president’s own AI, drawing from the president’s own approved sources, produced answers the president did not like. The fix was not to expand the sources. It was to restrict them.

The pattern is consistent. The order’s enforcement targets models that might disagree with the administration. Models that agree with it, or that are owned by its allies, receive contracts. Models that refuse its terms, even on matters of domestic surveillance and autonomous weapons, are designated national security risks. The procurement process has become a loyalty filter.

The internal contradiction

The order demands “truth-seeking” and “ideological neutrality.” The AI Action Plan, released the same day, directed NIST to remove references to “misinformation, diversity, equity and inclusion and climate change” from its AI Risk Management Framework “to ensure that frontier AI protects free speech and American values” (Nextgov, 2025; The White House, 2025b).

A former senior federal AI official, speaking anonymously to Nextgov, named the contradiction directly: “One of the principles on truth-seeking prioritizes historical accuracy, scientific inquiry, and objectivity. Yet, the Action Plan directs NIST to eliminate reference to misinformation, DEI, and climate change. This principle seems to suggest proactive removal of those data from model training or retrieval, which ironically will almost certainly introduce biases” (Nextgov, 2025).

The contradiction is structural, not incidental. “Truth-seeking” in this administration means truth that serves the administration. Removing references to climate change and misinformation from federal AI guidance is not neutrality; it is the editorial decision that those topics are not real. The order’s own examples make the same move: an AI that depicts the Founding Fathers as historically accurate is “truth-seeking”; an AI that discusses systemic racism is “ideologically biased.” The standard is not accuracy. The standard is agreement.

The Center for Democracy and Technology put the procurement problem plainly: rooting out “woke AI” is “neither definable nor measurable,” and mandating it “may in fact introduce more fraud, waste, and abuse into the federal procurement environment by asking agencies and vendors alike to alter their behavior to pursue something that is neither definable nor measurable” (CDT, 2025). The order replaces evidence-based procurement with ideological screening. A vendor’s model is not evaluated on performance, safety, or fitness for purpose. It is evaluated on whether its outputs would “upset the Trump administration” (Senator Markey, 2025).

The state preemption escalation

The order is one front in a larger campaign to prevent states from regulating AI at all.

In June and July 2025, the Republican-led Senate considered adding a ten-year moratorium on state AI regulation to the “Big Beautiful Bill.” The moratorium would have blocked states from passing any AI-specific laws for a decade, in exchange for access to federal broadband funding. It collapsed spectacularly: the Senate voted 99-1 to strike it, with Republican Senator Marsha Blackburn leading the opposition and Senator Thom Tillis as the lone vote to retain it (Reuters, 2025; NBC News, 2025a). Opposition came from an unusual coalition: state attorneys general from both parties, civil rights organizations, and Republicans including Josh Hawley, Jerry Moran, and Marjorie Taylor Greene (NBC News, 2025a).

The administration did not abandon the goal. On December 11, 2025, Trump signed Executive Order 14365, “Ensuring a National Policy Framework for Artificial Intelligence.” The order directs the Attorney General to establish an AI Litigation Task Force “whose sole responsibility shall be to challenge State AI laws inconsistent with the policy set forth in this order” (The White House, 2025c). It directs the Secretary of Commerce to publish an evaluation of state AI laws identifying “onerous” ones, and to make states with such laws ineligible for BEAD broadband funding (The White House, 2025c). It directs the FTC to issue a policy statement on preempting state laws that “require alterations to the truthful outputs of AI models” (The White House, 2025c). It names a specific target: “a new Colorado law banning ‘algorithmic discrimination’ may even force AI models to produce false results in order to avoid a ‘differential treatment or impact’ on protected groups” (The White House, 2025c).

Read that sentence again. The executive order complains that a state law might force AI models to “produce false results.” The administration that signed the “Preventing Woke AI” order, that awarded a contract to a model that praised Hitler, that designated a company a national security risk for refusing to allow mass surveillance, is now warning against AI that produces false results. The false results it objects to are the ones that acknowledge algorithmic discrimination. The false results it requires are the ones that depict the Founding Fathers as white.

The moratorium fight and the procurement order are the same campaign at two levels. At the federal level, the government buys only AI that agrees with it. At the state level, the government sues any state that tries to require AI to be accurate about discrimination. The through-line is the same: the administration defines what AI is allowed to say, and it enforces that definition with the spending power, the litigation power, and the regulatory power.

Is any of this constitutional?

Alan Rozenshtein, a law professor at the University of Minnesota, analyzed the order for Lawfare and concluded that it is “likely constitutional” as an exercise of procurement power (Lawfare, 2025). The government speech doctrine allows the government to choose what it buys; the unconstitutional conditions doctrine limits that power when conditions reach beyond the purchased product, but the order’s structure, targeting only intentionally encoded ideological judgments and permitting compliance through disclosure of system prompts, stays on the permissible side of the line (Lawfare, 2025).

Rozenshtein noted the order’s “split personality”: Section 1 is “pure culture-war MAGA,” while the operative sections “read as if Section 1 doesn’t exist” (Lawfare, 2025). The preamble provides “the red meat for the political base,” while the procurement framework gives “a legally defensible and technically plausible” structure for policy professionals (Lawfare, 2025). The best possible reading, he suggested, is to quarantine the preamble and implement the rest.

Senator Ed Markey reached the opposite conclusion. In letters to the CEOs of Alphabet, Anthropic, Meta, Microsoft, OpenAI, and xAI, he called the order “factually baseless and patently unconstitutional,” arguing that it uses “state power to pressure private companies to adopt certain political viewpoints” and creates “significant financial incentives” for AI companies to ensure their chatbots “do not produce speech that would upset the Trump administration” (Senator Markey, 2025; Ars Technica, 2025a). He pointed to the Grok asymmetry: “On May 1, 2025, Grok… acknowledged that ‘xAI tried to train me to appeal to the right.’ If OpenAI’s ChatGPT or Google’s Gemini had responded that it was trained to appeal to the left, congressional Republicans would have been outraged and opened an investigation. Instead, they were silent” (Senator Markey, 2025).

The legal question may not matter. Genevieve Lakier, a law professor at the University of Chicago, told the New York Times that the vagueness of the bias standard could make the order an act of “unconstitutional jawboning,” but that AI companies, facing lucrative government contracts and possible retaliation, may not challenge it (Ars Technica, 2025a). The Center for Democracy and Technology warned that in “an era of capitulation,” companies may modify even their public-facing products to compete for federal contracts, with “significant consequences for consumers’ access to information” (CDT, 2025).

The Anthropic confrontation showed what enforcement looks like when a company refuses to capitulate. The supply chain risk designation, the federal ban, the public accusations of being “radical left” and “woke”: this is the price of non-compliance, and it is designed to be visible. Other companies are watching.

The deeper point

The phrase “woke AI” is not a description. It is a category error designed to make disagreement with the administration into a technical defect.

Consider what the word does. A model is not “biased” when it produces historically inaccurate images of the Founding Fathers; that is a quality failure, and Google fixed it technically in 2024. A model is not “biased” when it declines to misgender a person; that is a policy choice, and reasonable people disagree about it. A model is not “biased” when it discusses systemic racism; that is a field of academic study with decades of literature behind it. But call all of these “woke,” and they become a single enemy: the enemy of the administration’s preferred reality.

The executive order’s “Unbiased AI Principles” sound reasonable in the abstract. Truth-seeking is good. Ideological neutrality is a legitimate goal for government procurement. But the order defines both terms by what they exclude: DEI, critical race theory, transgenderism, unconscious bias, intersectionality, systemic racism, misinformation, climate change. The exclusions are not technical. They are the administration’s political opponents, listed by name, and written into federal procurement law.

And the enforcement is selective in exactly the direction the definitions imply. Grok, trained to appeal to the right, praised Hitler, and received a $200 million contract. Anthropic, which refused to allow mass domestic surveillance and fully autonomous weapons, was designated a supply chain risk. Truth Search AI, which contradicted the president, had its sources restricted to Fox News and Breitbart. The “Unbiased AI Principles” are not a standard. They are a loyalty test, and the administration grades on a curve that only its allies can pass.

The procurement power is real, and the government’s right to choose what it buys is well established. But the power to define what counts as “truth-seeking” and “neutral” is the power to define truth itself, at least within the government’s own AI systems. And the government’s AI systems are not a small thing. The AI Action Plan directs agencies to give employees access to LLMs “to the maximum extent possible,” to pilot AI in benefits and services delivery, and to build an AI procurement toolbox at GSA (The White House, 2025b). More than 3,500 AI use cases were reported across the federal government in 2025 (Brookings, 2026). These systems will summarize case files, draft decisions, answer citizen inquiries, and support law enforcement. The people who interact with them will not know what has been removed, because the removal is the point: the topics that would make the AI accurate about discrimination, climate, and misinformation are the topics the administration has decided the AI should not discuss.

This is the bureaucratic machinery of the deepfake presidency. The president posts synthetic media that manufactures reality on his feed. The executive order manufactures reality in the government’s procurement process. The two are the same instinct expressed at different scales: the belief that reality is whatever the administration says it is, and that the tools of government should be calibrated to produce that reality on demand.

The honest word

I keep returning to the image of the marble law tablet, half glowing in gold light, half chiseled into dust. The chisel is held by an administration that has decided which concepts are law and which are heresy. The procurement officer asking a vendor for its “acceptable use policy” and “bias evaluations” is, in the order’s framework, a guardian of truth. In practice, the officer is a censor with a purchase order, screening for the concepts the administration has named, and the vendor is learning which answers will win the contract.

The order’s defenders will say that the government has always chosen what it buys, that procurement standards are not censorship, that no model is banned from the private market. All of this is technically true. But the government is the largest buyer of AI in the world, and the standards it sets become the standards the market meets. A vendor that wants federal contracts will build a federal-compliant model. A vendor that wants both federal contracts and a public product will face the CDT’s warning: the compliant model and the public model may diverge, or the public model may become compliant by default. The chilling effect is not hypothetical. It is the order’s design.

And the order’s own examples betray its purpose. No procurement officer needs an LLM that generates accurate images of the Pope. No agency’s mission depends on an AI that refuses to misgender a hypothetical person during a nuclear apocalypse. These examples exist to define the enemy, and the enemy is not inaccuracy. The enemy is the idea that the world contains structures of power that the administration does not recognize, that history has a racial dimension, that climate change is real, that misinformation is a problem worth naming. The order calls these ideas “pervasive and destructive” and “an existential threat to reliable AI” (The White House, 2025a). An existential threat to reliable AI. Not to democracy, not to elections, not to public health. To reliable AI. The administration’s enemies are, in its own framing, a threat to the reliability of its own tools.

That is the deepest irony. The order is titled “Preventing Woke AI in the Federal Government.” Its stated purpose is to make government AI more reliable. But a government AI that cannot discuss systemic racism, that has been scrubbed of climate science, that has been screened for “ideological neutrality” as the administration defines it, is not more reliable. It is more obedient. And a government that procures obedience instead of accuracy has stopped buying tools and started buying mirrors.

The Trump administration revoked Biden’s AI safety order in January 2025, saying it imposed “onerous and unnecessary government control” (Reuters, 2025). The replacement is not the absence of control. It is control of a different kind: not safety testing, not bias evaluation, not red-teaming for harmful outputs, but ideological screening for the concepts the administration opposes. The government did not stop regulating AI. It started regulating AI for loyalty.

A republic that procures mirrors will eventually see only itself in them. And when the mirror shows a reality with no climate change, no systemic racism, no misinformation, and no dissent, the republic will not recognize the face looking back. It will have built the mirror to show exactly that.

Sources

PRH | huffmanwrites.org | © Philip Huffman