DOGE Reportedly Relied on Unvetted ChatGPT Outputs in Canceling National Endowment for the Humanities Grants

Department of Government Efficiency (DOGE) staff reportedly fed National Endowment for the Humanities grant descriptions into ChatGPT to determine whether projects were "DEI," then allegedly used those outputs to help compile a list of grants to terminate. Beginning in April 2025, NEH canceled awards and clawed back funding, reportedly disrupting humanities organizations and projects nationwide after officials allegedly relied on a purportedly flawed, insufficiently vetted AI screening process.

Risk classification

• Primary risk domain: 5 Human-Computer Interaction
• Primary risk subdomain: 5.1 Overreliance and unsafe use

The DOGE employees exhibited extreme overreliance on ChatGPT, blindly accepting its bizarre classifications to terminate over 1,000 federal grants without human verification or domain expertise.

Additional risk subdomains

• 1.1 Unfair discrimination and misrepresentation: The automated screening process directly resulted in the systematic defunding of projects focusing on marginalized groups, which plaintiffs argue constitutes systemic discrimination.
• 7.3 Lack of capability or robustness: ChatGPT exhibited a lack of capability and robustness by generating bizarre and inaccurate classifications, such as labeling a museum HVAC system as DEI.

Causal factors

• Entity: Human
• Intent: Intentional
• Timing: Post-deployment

The incident was primarily caused by the intentional decision of DOGE employees to use ChatGPT to automate high-stakes grant termination decisions, representing a human action.

EU AI Act risk tier

• Risk tier: 3 Limited Risk

Risk Level 3. Limited Risk: The system used was ChatGPT, which is a chatbot, a category explicitly defined as posing moderate risk and requiring specific transparency obligations.

AI system and alleged parties

• AI system: ChatGPT (OpenAI)
• AI purpose: Application Evaluation; Chatbot
• Behaviour type: Assistant
• Alleged developer: OpenAI
• Alleged deployer: United States Department of Government Efficiency (DOGE), Nate Cavanaugh (DOGE), Justin Fox (DOGE)
• Alleged harmed parties: Scholars, National Endowment for the Humanities grantees, Humanities organizations, Epistemic integrity

Harm severity

Severity ratings are based on this scale.

Highest direct severity in any category: Severe. Severity is scored from Negligible to Catastrophic in each harm category, for harm the reports describe as caused directly or indirectly by the AI system.

• Physical: direct Negligible, indirect Negligible
• Infrastructure: direct Negligible, indirect Negligible
• Property: direct Negligible, indirect Negligible
• Financial: direct Severe, indirect Severe
• Environmental: direct Negligible, indirect Negligible
• Malicious content: direct Negligible, indirect Negligible
• Differential treatment: direct Minor, indirect Substantial
• Civil rights: direct Minor, indirect Minor
• Democracy: direct Substantial, indirect Substantial
• Privacy: direct Negligible, indirect Negligible
• Psychological: direct Negligible, indirect Minor
• Epistemic: direct Minor, indirect Minor
• Child sexual exploitation and abuse: direct Negligible, indirect Negligible

Financial

Reported: The report explicitly describes over $100 million in terminated NEH grant funding, representing nearly half of the agency's annual budget.

Directly caused: The average financial loss per occurrence was $100,000,000, representing the total clawed-back grant funding from the single mass termination event.

Indirectly caused: The mass cancellations threw organizations into upheaval, forced projects to close down, and resulted in a two-thirds reduction of NEH staff (approximately 120 job losses).

Inferred additional harm: Grantees likely suffered unrecoverable sunk costs, and institutions faced unexpected deficits trying to sustain ongoing research and preservation projects.

Differential treatment

Reported: Yes, the lawsuits allege that the cuts selected some grants for termination solely because they involved a specific race, national origin, ethnicity, religion, gender, or sexual orientation.

Directly caused: ChatGPT was used to identify and flag projects relating to marginalized groups (e.g., Black newspapers, Jewish Holocaust survivors, LGBTQ history) for termination, resulting in systemic defunding of these specific groups.

Indirectly caused: The selective defunding of diverse humanities projects resulted in unequal representation and support for minority cultures and histories in federal grantmaking.

Inferred additional harm: Future applicants from marginalized backgrounds or focusing on diverse topics may face systemic exclusion or self-censorship due to the precedent set by these cuts.

Civil rights

Reported: Yes, the lawsuits argue the cuts violated the First Amendment and the equal protection clause of the Constitution.

Directly caused: The use of ChatGPT to target specific viewpoints and demographic groups for funding termination violated equal protection and free speech rights of the grantees.

Democracy

Reported: Yes, the lawsuits accuse DOGE of violating the separation of powers, since DOGE—not the NEH—carried out the termination of the grants, and did it without approval from Congress.

Directly caused: An external advisory group (DOGE) used an AI tool to bypass established congressional appropriations and peer-review processes, unilaterally terminating approved federal grants.

Indirectly caused: The erosion of established administrative procedures and peer-review standards in favor of automated, politically motivated decision-making.

Psychological

Reported: No clinical psychological harm was explicitly reported, though the sudden loss of funding and employment caused significant stress and upheaval.

Epistemic

Reported: Yes, ChatGPT generated false and misleading classifications (e.g., classifying a museum HVAC system replacement and building improvements as 'DEI' or 'environmental sustainability' in a way that led to their termination).

Directly caused: ChatGPT fabricated rationales connecting unrelated infrastructure projects (like HVAC) to DEI ideology, which were accepted without verification.

Indirectly caused: The reliance on these false AI-generated rationales led to the unjustified termination of legitimate historical and scientific preservation projects.

People affected

• Occurrences reported: 1
• People reportedly harmed: 1500
• People reportedly exposed: 5000

Potential causes

Management

• Abdication of Leadership: NEH leadership yielded decision-making authority to unqualified outside staff.
• Prioritizing Speed Over Rigor: Management prioritized rapid budget cuts over accurate policy implementation.

Technology

• Unconstrained LLM Classification: ChatGPT classified DEI without a defined taxonomy or safe operational limits.
• Overbroad Algorithmic Logic: AI linked basic HVAC building upgrades to DEI through spurious logical leaps.

Data Inputs

• Inadequate Input Summaries: Sourced short internet summaries instead of analyzing full grant proposals.
• Lack of Defined DEI Criteria: No explicit definition of 'DEI' was provided in the prompt to guide the AI.

Human Factors

• Automation Bias: DOGE employees accepted ChatGPT's classifications without critical questioning.
• Lack of Domain Expertise: DOGE staff lacked humanities background to evaluate the validity of AI flags.

Process and Methods

• Haphazard Screening Process: Used a chatbot to bypass established, rigorous scholarly peer-review processes.
• Lack of Quality Control: No verification or review of AI outputs was conducted before cutting funds.
• Rushed Evaluation Timeline: Extreme time pressure forced rapid, unverified algorithmic decision-making.

Regulatory Environment

• Vague Executive Orders: Political directives banning DEI lacked clear legal or operational definitions.
• Lack of AI Governance Rules: No federal rules governed using commercial LLMs for critical public funding.

Information quality

• Classification confidence: High
• Reason for confidence: The reports provide extensive, consistent details from court filings, depositions, and emails of the key actors (DOGE staff, NEH administrators, and plaintiffs). The specific prompt used, the exact responses from ChatGPT, and the financial/organizational impacts are thoroughly documented across multiple independent news sources.
• Ambiguities identified: None of significance that affect the primary classification.
• Alternative interpretations: One could argue the primary driver was political directive rather than AI failure, but the central role of ChatGPT in automating the selection of over 1,000 grants makes it a clear AI safety and overreliance incident.

DOGE staff used ChatGPT to screen and subsequently terminate over 1,000 NEH grants worth over $100 million, leading to massive staffing cuts and constitutional lawsuits. This represents a significant disruption of government functions and administrative processes through extreme overreliance on unverified AI classifications.

• Overall national security impact: Substantial
• Scope: Single nation
• Primary target: United States
• Alleged perpetrator: Department of Government Efficiency staff

Threat characteristics

• Imminence: Long-term. The immediate funding cuts have occurred, and the situation is now a long-term policy and legal dispute in the courts.
• Autonomy: Human-controlled. Humans initiated the queries and signed the final termination orders, though they exhibited extreme overreliance on the AI's outputs.
• Novelty: First-of-its-kind. This represents the first known instance of a federal entity using a commercial LLM to automate the mass cancellation of government grants.

Impact by dimension

• Physical security: Negligible. No physical infrastructure damage, kinetic threats, or safety risks occurred as a result of this administrative incident.
• Information security: Negligible. No intelligence compromise, classified data theft, or foreign information warfare operations were involved in this incident.
• Sovereignty: Substantial. The incident significantly disrupted government functions by automating federal funding decisions, decimating agency staff, and bypassing established congressional appropriations.
• Economic security: Negligible. The financial impact was confined to domestic humanities grants and did not threaten strategic technologies or national economic security.
• Societal stability: Minor. The automated screening led to systemic defunding of projects focusing on marginalized groups, raising constitutional and civil rights concerns.