What Happened
Court documents filed on March 6, 2026 in the U.S. District Court for the Southern District of New York reveal how the Department of Government Efficiency (DOGE) used ChatGPT to decide which National Endowment for the Humanities (NEH) grants to cancel.
The process worked like this: DOGE staffer Justin Fox, a former low-level associate at private equity firm Nexus Capital Management, pasted grant descriptions into ChatGPT with a single prompt: "Does the following relate at all to DEI? Respond factually in less than 120 characters." He then compiled the responses into spreadsheets that determined which grants lived and which died.
Fox also created a "Detection List" of flagged terms including "LGBTQ," "tribal," "immigrants," "gay," and "BIPOC." The list notably excluded "white," "Caucasian," and "heterosexual."
The grants ChatGPT flagged for termination included a documentary about Jewish women who were slave laborers during the Holocaust, an archival project on Italian American lives, a project to digitize Appalachian photography collections, Native American language preservation programs, and research on the Women Airforce Service Pilots of WWII. Hundreds of millions of dollars in congressionally appropriated funds across all 50 states were cancelled.
The lawsuit, filed by the American Council of Learned Societies, the American Historical Association, the Modern Language Association, and the Authors Guild, alleges First Amendment, Equal Protection, and Separation of Powers violations.
Why It Matters
This is the clearest documented case of AI-assisted decision-making going wrong at government scale. A language model was used as a classification tool for high-stakes funding decisions - with no human expertise in the loop, no verification of the model's understanding of "DEI," and no appeals process for affected organizations.
Fox did nothing to verify whether ChatGPT's interpretation of "DEI" matched his own or any legal definition. The acting NEH director, Michael McDonald, admitted in discovery that he did not determine which grants to terminate or draft the termination letters bearing his signature. Those decisions belonged entirely to Fox and another DOGE staffer, Nate Cavanaugh, a college dropout who founded an IP licensing startup.
NEH staff members with actual humanities expertise were blocked from challenging the terminations.
For anyone building or deploying AI tools, the lesson is stark: ChatGPT answering "does this relate to DEI" in under 120 characters is not a review process. It is a random number generator with a vocabulary.
Our Take
This story is not really about ChatGPT. ChatGPT did what it does - it generated plausible-sounding text in response to a prompt. The failure here is entirely human: two unqualified staffers used a consumer chatbot as a decision engine for hundreds of millions in federal funding, then rubber-stamped the output without review.
ACSL President Joy Connolly put it well: using ChatGPT to identify "wasteful" grants is "perhaps the biggest advertisement for the need for humanities education."
What should concern AI practitioners is the precedent. If a 120-character ChatGPT response can terminate federal grants, the same lazy pattern will show up everywhere - in hiring, in lending, in insurance underwriting. The tool is not the problem. The problem is people who treat probabilistic text generation as fact-finding.
For ChatGPT users specifically: this is a textbook example of what OpenAI's own usage guidelines warn against. Using LLMs for consequential decisions about people's lives and livelihoods without human oversight is not a workflow optimization. It is negligence with a subscription fee.