Court voids DOGE's ChatGPT-driven cancellation of NEH grants

Federal agencies have started experimenting with LLMs, and a May 2025 court ruling has now drawn one of the first hard lines around where that experimentation ends. Writing in Tech Policy Press, Jordan Ascher walks through what the Department of Government Efficiency actually did at the National Endowment for the Humanities: it asked ChatGPT, "Does the following relate at all to DEI?", fed in grant descriptions, and used the answers to cancel more than 1,400 grants.

The classifications the model returned tell you why the case became a test. A whaling museum grant was flagged because it sought to "create an inclusive and impactful experience, which is aligned with DEI principles." A documentary on the Colfax Massacre was flagged for exploring "a historical event that significantly impacted Black civil rights." A Holocaust documentary was cited for focusing on "Jewish cultures" and "voices of the females." An academic study got the analytic tag "#DEI" and nothing more.

The court threw the terminations out. Terminating funding based on "disfavored ideas" violated the First Amendment, and canceling grants because they referenced a particular race, gender, or other protected class "defied the Constitution's promise of equal protection." The move that matters most for anyone building AI into government work is what the court did with the attribution question. DOGE argued ChatGPT was just providing "context." The court disagreed. Because DOGE "selected the AI tool, formulated the prompt, and defined the operative viewpoint-based criterion," the model's classifications "are the Government's own for constitutional purposes."

Read strategically, that is the sentence. Every prompt, every model choice, every threshold an agency sets is, from now on, the agency's own act. The honest caveat is that this is one court's ruling on a specific and vivid factual record, and Ascher does not tell you whether it will be appealed, how other agencies are adjusting their own LLM pilots, or what specific procedural safeguards the court would consider sufficient beyond meaningful human review.

Ascher's forward-looking read is that LLMs will still have real applications in federal administration, but only where "accountable agency officials can meaningfully review and validate their work." For general counsels, CIOs, and vendors selling AI into government, the useful takeaway is that a human sign-off can no longer be a rubber stamp bolted on at the end. If the prompt and the tool are the government's, the output is too.