Court Rebukes DOGE's ChatGPT DEI Vetting

This is a clean example of why high-stakes AI use collapses when the model is treated like a reason generator instead of a noisy classifier. Courts do not care that the output sounds plausible if the process behind it is undefined, undocumented, and unauthorized.

• –DOGE reportedly asked ChatGPT whether grant summaries related to DEI, but gave it vague prompts and no meaningful definitions
• –The court found the resulting rationales were too thin to support a lawful termination decision
• –The ruling also says DOGE lacked statutory authority to terminate NEH grants in the first place
• –For builders, the lesson is simple: model outputs need constrained inputs, explicit criteria, and auditable human decision ownership
• –This is a policy warning shot for any government or enterprise team trying to launder discretionary decisions through LLM prose