The court found that cancelling more than 140 humanities grants totaling over $100 million was unconstitutional. Judge Colleen McMahon highlighted that DOGE used ChatGPT to screen grants for connections to "DEI" without defining the term or considering project substance. The AI reviewed brief spreadsheet descriptions and produced terse determinations, sometimes misclassifying projects such as an environmental study on plastics. The judge warned that generative AI's "hallucinatory propensities" and lack of reasoning undermined those decisions.
Judge Rules DOGE's Use of ChatGPT Helped Cut $100M in Humanities Grants — Ruling Finds Process Unconstitutional
DEI Cropped - Credit: J Studios via Getty Images
A federal judge in New York has ruled that the Trump administration's cancellation of more than 140 humanities grants, totaling over $100 million, was unconstitutional. The lawsuit was brought by the Authors Guild and several individual recipients of National Endowment for the Humanities grants.
Judge Highlights Role of AI in Decision-Making
In her filing, U.S. District Judge Colleen McMahon emphasized that the Department of Government Efficiency (DOGE) relied on OpenAI's ChatGPT during parts of its grant-screening process. Testimony from former DOGE staffer Justin Fox described how the team used the AI to identify items "to do with DEI" and to "highlight why [a] grant may relate to DEI." According to the filing, staff fed ChatGPT a spreadsheet of short grant descriptions and prompted the model: "Does the following relate at all to DEI? Respond factually in less than 120 characters."
Key Problems Identified by the Court
The court found several serious problems with this approach. DOGE did not provide the AI with a definition of "DEI," and staff did not instruct the model to consider a project's purpose, methodology, or scholarly substance. Fox testified that he did not know how the chatbot interpreted "DEI," and that the AI's determinations were based only on the terse spreadsheet entries.
"The record reflects that these ChatGPT determinations were generated without any additional context beyond the cursory spreadsheet descriptions themselves," the filing said.
Judge McMahon also warned about the well-documented tendency of generative AI systems to produce misleading or fabricated explanations. The ruling noted the tool's "hallucinatory propensities" and observed that it would "hardly be surprising" if ChatGPT produced rationales that simply mirrored what users appeared to want. The filing added: "The utter lack of reasoning behind so many of its 'rationales' certainly suggest[s] as much."
Example: A Misclassified Environmental Study
One illustrative example cited in the filing is a terminated grant titled "How the City Became Plastic," a project that examined how the plastics industry lobbied municipal officials and how microplastics entered local water systems. Despite its focus on industrial policy and environmental contamination, the project was flagged as DEI-related by the AI-driven screening process.
Broader Implications
The ruling underscores the legal and practical risks of relying on generative AI tools for consequential public-administration decisions without clear definitions, human oversight, or consideration of substantive evidence. The opinion also noted that, while AI is likely to remain part of government workflows, its use should be limited and carefully governed. As of March 2026, the filing observed, Senate aides have been permitted to use AI chatbots such as Google Gemini, ChatGPT, or Microsoft Copilot for limited "routine work," like summarizing information.
What This Means: The decision is a reminder that automated tools can speed administrative work but cannot substitute for reasoned, well-documented decision-making when civil liberties, public funding, or scholarly work are at stake.
Help us improve.
