Federal Court Rules Government's LLM-Driven Grant Cancellation Unconstitutional, Citing First Amendment Violation

A federal court has drawn a definitive line on how the United States government can deploy artificial intelligence to police its own funding, striking down the mass cancellation of over 1,400 federal grants.[1][4]

The landmark ruling in American Council of Learned Societies v. National Endowment for the Humanities marks a watershed moment for AI governance and civil liberties. The court found that the government's use of an off-the-shelf large language model to flag and terminate research based on ideological keywords violated the First Amendment.[1][3]

The mechanism behind the mass cancellations reveals the unprecedented speed at which generative AI can execute administrative mandates. Following the presidential inauguration in early 2025, the newly formed Department of Government Efficiency (DOGE) was tasked with aggressively reducing federal spending and eliminating programs tied to "diversity, equity, inclusion, and accessibility."[1][2]

To process the massive volume of active federal grants, two DOGE staffers, Justin Fox and Nathan Cavanaugh, arrived at the National Endowment for the Humanities (NEH) with a novel strategy. Rather than manually reviewing thousands of pages of academic proposals, they turned to OpenAI's ChatGPT.[2]

According to testimony reported by The New York Times, the staffers fed descriptions of hundreds of NEH grants into the chatbot with a highly specific, rigid prompt designed to automate the ideological vetting process.[2]

The prompt read: "Does the following relate at all to D.E.I.? Respond factually in less than 120 characters. Begin with 'Yes' or 'No.'" The AI obliged, rapidly categorizing years of federally funded research in mere seconds.[1][2]

Projects were flagged simply because terms related to LGBTQ communities, race, or gender appeared in the text. In one instance revealed during testimony, a military research project was targeted and canceled simply because ChatGPT noted that it "explicitly says LGBTQ."[2]

The government then rubber-stamped the chatbot's outputs, adopting them wholesale to terminate more than $100 million in funding across various agencies. The speed of the AI allowed the department to execute the cuts at a scale that would have taken human reviewers months to achieve.[1][2]

The sweeping cancellations triggered immediate legal action. The Modern Language Association, the American Council of Learned Societies, and the American Historical Association filed a joint lawsuit, arguing the terminations were an unlawful exercise of viewpoint discrimination.[4]

In its legal defense, the government attempted to distance itself from the AI's outputs. Federal attorneys argued that the ChatGPT exchange merely provided "context" for the reviewers and did not taint the ultimate human decisions to terminate the grants.[1]

The federal court rejected this defense entirely. The judge found that the agency "selected the AI tool, formulated the prompt, and defined the operative viewpoint-based criterion" that drove the mass cancellations.[1]

Because the agency adopted the classifications generated by the LLM to justify terminating grant after grant, the court ruled that "those classifications are the Government's own for constitutional purposes."[1]

The constitutional violation was stark. Terminating funding based on its purported connection to "disfavored ideas" was ruled a clear violation of the First Amendment's protection against government censorship.[1][3]

Furthermore, the court noted that canceling grants because they referenced a particular race, gender, or protected class defied the Constitution's promise of equal protection, turning the AI into an instrument of automated discrimination.[1]

Legal scholars note that the candor of the dialogue between the DOGE staffers and ChatGPT provided a rare "smoking gun" of discriminatory motivation. The user explicitly asked the chatbot to discriminate based on viewpoint, and the chatbot provided the exact rationale needed to execute the cuts.[1]

As federal agencies rush to integrate generative AI into their workflows to save time and money, the ruling serves as a stark warning about the limits of algorithmic administration.[1][3]

The decision establishes a vital precedent: the government cannot launder unconstitutional censorship through an algorithmic intermediary. The state remains fully liable for the constitutional violations generated by the tools it chooses to deploy.[3]

Ultimately, the ruling affirms that while large language models are powerful tools for bureaucratic efficiency, the constitutional obligations of the state cannot be outsourced to a chatbot.[1][4]