The 120-Character Axe: How DOGE Used ChatGPT to Gut the National Endowment for the Humanities

A single AI prompt. No expert review. No deliberation. Just a chatbot and a binary answer. That is now how the U.S. federal government makes decisions about which scholars get funded and which projects get killed.

AI is now making federal funding decisions. The implications extend far beyond the humanities. (Unsplash)

The prompt took fewer characters than a tweet. "Does the following relate at all to D.E.I.? Respond factually in less than 120 characters. Begin with 'Yes' or 'No.'"

That was it. That was the entire decision framework that DOGE operatives fed into ChatGPT when they arrived at the National Endowment for the Humanities with a mandate to cancel grants the Trump administration considered contrary to its anti-DEI agenda. According to reporting by The New York Times in early March, the DOGE team did not review funded projects carefully. They pulled short summaries off the internet and fed them to an AI chatbot. The chatbot responded. Grants died.

The results, by the Times's own description, were "sweeping, and sometimes bizarre." Researchers who had spent years building projects, many with no connection whatsoever to diversity, equity, or inclusion in any substantive sense, found themselves cut because a language model skimming a web summary said "Yes."

This is the most concrete example yet of what AI-powered government actually looks like in practice. Not the sci-fi version with robot bureaucrats and intelligent algorithms carefully weighing competing values. The real version: a consumer chatbot used to launder political decisions through a veneer of technological objectivity. A speed tool. A plausible-deniability machine.

What Happened at the NEH

The National Endowment for the Humanities has operated since 1965, funding academic research, library projects, public education, and cultural preservation. Its annual budget runs around $207 million. It is not a large federal agency. But it funds thousands of projects across every state, touching universities, museums, local libraries, and independent scholars.

DOGE's entry into the NEH followed the same pattern seen at other agencies: a small team of operatives arriving with broad authority and a mandate to cut anything that could be labeled DEI-adjacent. Unlike the Pentagon or the intelligence community, the NEH had no hardened bureaucratic resistance. It was a soft target with a meaningful symbolic value - culture, learning, the humanities as a concept.

What made the NEH action different from other DOGE operations was the reported methodology. Previous DOGE cuts at agencies like USAID and the Department of Education involved lists, categories, and at least some human decision-making at the top. At the NEH, according to the Times, that human layer was replaced with a ChatGPT session.

"Instead of looking closely at funded projects, they pulled short summaries off the internet and fed them into the A.I. chatbot. The prompt was simple: 'Does the following relate at all to D.E.I.? Respond factually in less than 120 characters. Begin with Yes or No.' The results were sweeping, and sometimes bizarre." - The New York Times, March 7, 2026

The phrase "sometimes bizarre" is doing a lot of work in that sentence. Language models are pattern-matching engines. They do not understand context in the way humans do. A project studying the historical experiences of migrant farmworkers might be correctly flagged as touching on marginalized communities - or it might be flagged because the summary used the word "diverse" in a demographic sense. A project studying the rhetoric of civil rights movements might be flagged because it mentions race. A project about ancient Roman governance with no DEI component whatsoever might be spared because its summary used academic jargon that the model did not associate with trigger terms.

The model does not know what any of these projects actually are. It knows what their summaries say, as scraped from the web, processed through whatever training data shaped its understanding of language. That is the system that is now making federal funding decisions.

The Second-Order Effects Nobody Is Talking About

The automation of government decision-making creates cascading effects beyond any single agency cut. (Unsplash)

The immediate story is about the NEH and the scholars who lost funding. But the structural implications go much further, and they are moving faster than the commentary can catch up.

First: the precedent. If DOGE successfully used ChatGPT to cancel NEH grants and nobody stopped them - no court injunction on the methodology, no congressional hearing that forced them to justify the AI's decision-making - then the approach becomes replicable. Every future administration, not just this one, now has a template. You can automate ideological filtering of federal programs. You just need a prompt and a chatbot subscription.

Second: legal accountability. When a human bureaucrat makes a decision to cancel a grant, there is theoretically a chain of accountability. The bureaucrat can be deposed. The decision memo can be subpoenaed. The reasoning can be interrogated. When a language model makes the decision, the accountability chain collapses. Who is responsible for a ChatGPT output? OpenAI? The DOGE operative who wrote the prompt? The White House that issued the mandate? The answer, in practice, is probably nobody - or at least nobody who can be held legally responsible in a way that would result in grants being reinstated.

Third: the OpenAI dimension. This is a company that has simultaneously signed a major Pentagon contract enabling AI deployment in military contexts, watched its head of robotics resign in protest over concerns about lethal autonomous weapons, and is now - through its consumer product - being used to make federal cultural funding decisions. OpenAI almost certainly did not sanction or even know about this specific use. But that is precisely the problem: a general-purpose AI product that anyone can use for anything is now embedded in federal decision-making with no oversight mechanism.

Fourth: what this reveals about the actual state of AI deployment. The AI industry has spent years telling governments and enterprises that AI should augment human judgment, not replace it. The NEH case shows that this aspiration collapses the moment someone with authority and a mandate decides to skip the augmentation step. There is no technical barrier to using ChatGPT as the sole decision-maker. There is only policy, and policy can be overridden by anyone with enough power and enough urgency.

OpenAI's Government Entanglement Gets Deeper

The NEH story lands at the worst possible moment for OpenAI's internal culture. The company is already navigating one of its most significant personnel crises: Caitlin Kalinowski, its head of robotics, resigned in the first week of March over the company's Pentagon contract.

Kalinowski's resignation statement, posted publicly on X, was direct. She said the Pentagon deal did not do enough to protect Americans from warrantless surveillance, and that granting AI "lethal autonomy without human authorization" was a line that "deserved more deliberation" than the company had given it.

The resignation matters because Kalinowski was not a junior employee. She was the person leading OpenAI's push into physical robotics - arguably the most consequential frontier in AI after large language models. Her departure signals something about internal temperature at a company that is making very large bets on government relationships.

Meanwhile, the QuitGPT protest movement claimed in early March that 1.5 million people had taken some form of action against OpenAI's Pentagon deal - signing a boycott, sharing on social media, or both. A physical protest outside OpenAI's offices in San Francisco took place on March 3rd, focused on AI-powered mass surveillance and autonomous weapons. These are not fringe concerns. They are being voiced by researchers, civil society organizations, and now by OpenAI's own former robotics lead.

What ties the NEH story to the Pentagon story is a single thread: OpenAI is becoming infrastructure for government power, and the company is not in control of how that power is applied. The Pentagon contract is a deliberate business decision. The NEH ChatGPT session was not. Both are now part of what OpenAI's technology does in the world.

The Architecture of Automated Ideology

To understand why this matters beyond the immediate political controversy, you have to understand what large language models actually do when given a classification task.

ChatGPT - and models like it - work by predicting plausible next tokens given a prompt. When you ask it "Does this relate to DEI? Answer Yes or No," it is not performing a careful conceptual analysis. It is pattern-matching against its training distribution. The word "diversity" in almost any context will push it toward "Yes." So will words like "equity," "inclusion," "marginalized," "underrepresented," "community," and hundreds of related terms.

This creates a systematic bias in the filtering process that has nothing to do with whether a project is actually DEI-focused in the policy sense DOGE intended. A project studying historical immigration patterns might use the word "diverse" once in a summary sentence. A project about medieval manuscript preservation might quote a scholar discussing the "equity" of access to historical archives. These are not DEI programs. But a language model skimming a web summary is not equipped to make that distinction reliably.

Language model researchers call this "surface form competition" - the model responds to lexical patterns rather than semantic substance. It is a known limitation of current AI systems. It is precisely why AI researchers have consistently argued these systems should not be used as sole decision-makers in high-stakes contexts. The NEH case is a real-world demonstration of exactly what researchers warned against, deployed at federal scale, with no apparent technical review.

There is also a selection bias problem with the data source. DOGE operatives pulled project summaries "off the internet" rather than using NEH's own detailed project documentation. Web summaries of academic projects are often written for general audiences, are frequently abbreviated, and may not accurately represent the scope or methodology of funded work. Feeding these compressed, potentially inaccurate descriptions to a language model compounds every error: garbage in, garbage out, but the garbage is now a federal funding decision.

A Timeline of AI Entering the Government Apparatus

Codex Security and the Dual-Use Paradox

OpenAI's Codex Security launch positions the company deeper inside government and enterprise infrastructure just as scrutiny of its government ties intensifies. (Unsplash)

The day after the NEH story broke, OpenAI launched Codex Security - a new AI agent designed to identify and fix security vulnerabilities in software applications. It launched as a research preview, with the Codex Open Source Fund offering six-month ChatGPT Pro subscriptions with access to Codex Security to open source developers.

In isolation, Codex Security is a genuinely useful tool. Security researchers spend enormous time hunting for vulnerabilities manually. An AI agent that can scan code, identify potential attack vectors, and suggest patches could meaningfully accelerate the defensive security posture of projects that do not have dedicated security teams - most open source projects, countless startups, even mid-size enterprises.

But Codex Security does not exist in isolation. It exists inside a company that is simultaneously being used by DOGE to make federal funding decisions, has signed a Pentagon contract that its own head of robotics found ethically untenable, and is expanding aggressively into government infrastructure. An AI agent that is exceptionally good at finding security vulnerabilities has obvious dual-use potential - the same capabilities that let it find and fix bugs let it find bugs to exploit.

This is the dual-use paradox that the AI industry has never satisfactorily resolved. The more powerful and capable AI security tools become, the more valuable they are to both defenders and attackers. OpenAI's response to this tension has historically been to focus on access controls and responsible use policies. Those policies are only as effective as the enforcement mechanisms behind them - and as the NEH case shows, once a capability exists in a consumer product, it can be applied in ways the company never anticipated or sanctioned.

The open source developers who receive Codex Security access as part of the fund are presumably operating in good faith. The DOGE operatives who used ChatGPT to screen NEH grants also had access through ordinary subscription terms. The same product. Wildly different applications. No technical mechanism to distinguish between them.

Microsoft, Anthropic, and the Race to Become Government AI

While OpenAI navigates its Pentagon entanglement, Microsoft has quietly moved to deepen its own AI government positioning. The company announced this week that it is integrating Anthropic's Claude into Microsoft Copilot through a feature called "Claude Cowork" - built specifically to handle "long-running, multi-step tasks" that standard AI assistants struggle to complete reliably.

The Cowork integration is currently in testing and will be available through Microsoft's Frontier program later in March. It represents a significant architectural evolution: rather than a chatbot that answers questions, this is an AI agent that can autonomously execute extended workflows across enterprise software environments. Think of it as the difference between asking someone a question and handing them a project with a week-long deadline.

The timing is notable. Anthropic has just come through one of the most difficult weeks in its history - the Pentagon "supply chain risk" designation, the Dario Amodei memo, defense contractors abandoning Claude, and an unprecedented surge in public sign-ups that broke daily records. The Microsoft integration represents Anthropic's other government-adjacent pathway: not the Defense Department, which is explicitly off limits for now, but the vast ecosystem of federal contractors, regulatory agencies, and state governments that run on Microsoft infrastructure.

For Microsoft, the strategy is portfolio diversification. Copilot runs on OpenAI's models by default. Adding Claude as a capable alternative for complex long-running tasks reduces dependency on a single AI provider and positions Microsoft as the neutral platform layer - the infrastructure company that runs whatever AI the customer chooses. That is a very different position than being committed to a single model, and it is almost certainly where enterprise software is heading: AI-agnostic platforms that plug in whichever model performs best for a given task.

The government implications of this are significant. Federal agencies running Microsoft 365 - which is most of them - will eventually have access to long-running AI agents through their existing enterprise agreements. The question of whether those agents should be used the way DOGE used ChatGPT at the NEH is a governance question, not a technical one. And governance is what the industry has been weakest on.

What Comes Next: The Governance Gap Widens

The NEH case is not going to be the last time a government entity uses a consumer AI product to make consequential decisions with inadequate review. It is going to be replicated, scaled, and systematized - not just in the United States but globally, as governments everywhere watch DOGE's playbook and adapt it to their own contexts.

The structural problem is that AI governance - the policies, regulations, and institutional mechanisms designed to ensure AI is deployed responsibly - is moving at legislative speed. Technology moves at deployment speed. Those are not the same pace, and the gap between them is where cases like the NEH happen.

The European Union's AI Act, which came into force in 2024, categorizes "AI systems used in the administration of justice and democratic processes" as high-risk, requiring conformity assessments and human oversight. The NEH grant screening would almost certainly fall into this category under EU standards. But the AI Act is European law. It does not apply to how the U.S. government uses consumer AI products in Washington.

The U.S. has no equivalent framework. The Biden-era AI executive order, which contained some provisions for responsible AI use in government, was revoked in January 2025. The Trump administration's replacement focused on removing barriers to AI deployment rather than creating new guardrails. The result is a legal vacuum in which DOGE operatives can use ChatGPT to cancel humanities grants without violating any specific rule, because no specific rule exists.

OpenAI's own usage policies prohibit using its products to "make decisions that could significantly affect the life chances of individuals" without appropriate human review. The NEH case raises legitimate questions about whether that policy was violated. But OpenAI would have to know about the specific use case to enforce it, and enforcement of consumer product usage policies at the level of individual government sessions is not something any company is currently equipped to do.

The deeper issue is one of incentive alignment. OpenAI benefits commercially from being used broadly, including by government. Restricting government use of its consumer products would reduce revenue and market penetration. The company is also actively courting government relationships through its Pentagon contract and Codex Security launch. It cannot simultaneously pursue aggressive government business development and robustly restrict government misuse of its consumer products. Those two goals are in tension, and commercial incentives favor the former.

For the scholars and institutions that lost NEH funding because a chatbot said "Yes" to a 120-character prompt, that tension is not abstract. It translated into terminated projects, lost jobs, and research that will never be completed. The speed of AI deployment in government has real victims. The governance framework that should prevent those outcomes does not yet exist.

What DOGE demonstrated at the NEH is that the age of AI-powered government is already here. The question is not whether AI will be used in government decision-making. It already is. The question is whether anyone will be accountable for what it decides - and right now, the answer is no.