DOGE wields ChatGPT against humanities grants

This is a textbook failure mode for institutional AI: vague prompts, ideological keyword filters, and human operators treating probabilistic output like administrative judgment. For developers, it is a sharp reminder that the real risk surface around LLMs is often process design, not just model quality.

• –Discovery materials described by the American Historical Association say DOGE used ChatGPT responses and a spreadsheet workflow to identify grants for cuts
• –Grants reportedly flagged included projects tied to Holocaust history, Native languages, LGBTQ history, and other culturally or historically specific subjects
• –The core technical lesson is brutally familiar: if you ask an LLM a politically loaded classification question with no rigorous criteria, you get confident garbage at scale
• –This kind of misuse strengthens the case for audit logs, human review, domain experts, and strict limits on using general-purpose chatbots in high-impact decisions