The AI Excuse Just Died in Court

Introduction

A DOGE staffer fed $100 million in humanities grants into ChatGPT one at a time, using a prompt that capped ChatGPT's answer at 120 characters and never defined "DEI." The chatbot flagged a Holocaust anthology, a $349,000 HVAC replacement, and a study of the plastics industry (rationale: "#DEI") as DEI. The government's defense was that the AI did it. Judge Colleen McMahon's ruling says otherwise: "ChatGPT was the Government's chosen instrument. DOGE selected the AI tool, formulated the prompt, and defined the operative viewpoint-based criterion."

This piece is analysis based on court records, depositions, and published reporting. DOGE, Fox, Cavanaugh, the White House, and NEH did not respond to requests for comment.

How Two Staffers Killed the NEH in 22 Days

The National Endowment for the Humanities has awarded more than 70,000 grants since 1965. Between April 1 and April 3, 2025, 1,477 of them (roughly 97% of the active portfolio, per Inside Higher Ed) were placed on a termination list representing over $100 million in congressionally appropriated funds, per the ruling. NEH also fired 65% of its own staff in the same window, per ACLS plaintiff materials. The McMahon ruling calls it "the largest mass termination of previously awarded grants in the history of the NEH since its establishment."

The decision-makers were two GSA employees seconded to DOGE: Justin Fox and his superior Nate Cavanaugh. Per NPR, Fox was in his early 20s, with a CV running from investment banking at Nexus Capital Management to founding a self-described "cutting-edge blockchain consulting firm." No background in grant administration or the humanities. By March 12, 2025, the same day DOGE first met with NEH leadership, Cavanaugh had already sent acting chair Michael McDonald a spreadsheet labeled "Biden Grants" with 1,295 awards totaling $322 million flagged as "remaining (can presumably be clawed back)."

The Prompt That Never Defined DEI

Here's the prompt Fox used, quoted verbatim from page 21 of the ruling: "Does the following relate at all to DEI? Respond factually in less than 120 characters. Begin with 'Yes.' or 'No.' followed by a brief explanation."

That's the whole instruction, with no definition of DEI and no reference to the executive orders that define it. Fox testified under oath that he "had not the slightest idea how ChatGPT understood the term," a concession the government formally admitted in Dkt. No. 277. He fed 1,162 cursory grant descriptions into the chatbot one by one. No government witness could identify a single instance where Fox read ChatGPT's rationale, disagreed, and chose not to terminate. The ruling puts it plainly on page 94: DOGE "simply adopted ChatGPT's rationales wholesale as their own."

The rationales themselves, reported in depth by Inside Higher Ed, read like a parody of algorithmic governance. The Holocaust anthology was flagged because it "explores Jewish writers' engagement with the Holocaust in the USSR." The HVAC system at the High Point Museum was flagged because better climate control "aligns with the goal of providing greater access to diverse audiences." A study of ancient Moses manuscripts using multispectral imaging was flagged for the words "Jewish thought," and a study of the Chinese government's persecution of Uyghurs was flagged for DEI.

McMahon, in a footnote, noted that ChatGPT was probably doing exactly what consumer LLMs do when you keep prompting them in the same direction: hallucinating rationales to satisfy the perceived demand. Fox kept asking which grants were DEI, and the chatbot kept finding reasons they were.

The "Detection List" the Ruling Spent 12 Pages On

The ChatGPT process was only half of it. Fox also built what he called a "Detection List," a keyword filter he ran against every grant description independently. The Fifth Amendment section of the ruling itemizes the keywords: "gay," "BIPOC," "indigenous," "tribal," "melting pot," "equality," "LGBTQ," "homosexual," "minorities," "native," "immigrant." He did not search for "white" or "heterosexual." McMahon's verdict on that asymmetry: it "could not be more obvious" evidence of discrimination based on protected characteristics. This part had no AI in it at all — just a 20-something with a keyword list and no subject-matter expertise in any of the fields whose grants he was flagging.

The Pressure from the Top Was Made Up

Neither Fox nor Cavanaugh had statutory authority to terminate a federal grant. The NEH authorizing statute, 20 U.S.C. §§ 956(c) and 957(f), vests grant authority solely in the NEH Chairperson with advice from the National Council on the Humanities. DOGE isn't in that statute at all. So how did two unqualified GSA employees end up making the calls?

On March 31, 2025, the day before the mass terminations, Fox emailed McDonald: "We're getting pressure from the top on this and we'd prefer that you remain on our side but let us know if you're no longer interested." Cavanaugh later admitted in his January 23, 2026 deposition there was no actual White House directive: he and Fox invented the framing to coerce McDonald into compliance. McDonald yielded anyway, writing back: "as you've made clear, it's your decision on whether to discontinue funding any of the projects on this list."

The termination notices then went out from Fox's personal Microsoft email account, with letters drafted by DOGE that cited a non-existent executive order mandating NEH to "eliminate all non-statutorily required activities and functions." Grant recipients learned their funding was gone from an email sent by someone they had never heard of, citing a law that didn't exist.

The Grant That Wasn't Killed

While ChatGPT was tagging Jewish-themed humanities grants as DEI and DOGE was canceling them, NEH simultaneously awarded the Tikvah Fund — a conservative Jewish think tank — the largest single grant in the agency's 60-year history: $10.4 million for a "Jewish Civilization Project to combat antisemitism," announced September 2025. Per ACLS discovery materials and Washington Post reporting, McDonald's office reached out to invite the application; he then overruled his own advisory council to approve it. Tikvah's CEO says the proposal went through "multiple rounds of rigorous review." What's documented in the ruling is the council vote against it and McDonald's override. The contrast isn't that the Tikvah grant was wrong — it's that the same process that killed grants algorithmically approved one by executive fiat.

Why "ChatGPT Did It" Lost

The legal innovation in McMahon's opinion sits on pages 92–94. The government argued the viewpoint discrimination belonged to the AI rather than to the agency, with DOGE just processing the rationales ChatGPT generated. McMahon's response reads like genuine impatience: "There is no distinction to be drawn here between the Government and ChatGPT. ChatGPT was the Government's chosen instrument for purposes of this project. DOGE selected the AI tool, formulated the prompt, and defined the operative viewpoint-based criterion." She compared the defense to Flip Wilson's Geraldine Jones claiming "the devil made me do it," and said it worked no better for the federal government than for Geraldine.

That comparison is the precedent now: the first federal opinion to hold that using an AI tool to launder a politically motivated decision into something that looks algorithmic does not insulate the government from constitutional liability. Every agency is integrating generative AI into decision-making right now, and every contractor is selling some version of "let the model decide so nobody has to take the blame." McMahon's answer, in language that will end up in every brief filed against this practice: the model is an instrument and the government owns what it does with one.

Who Benefits

The mechanics here are about cover. DOGE's incentive wasn't to find DEI; it was to cancel $100 million in grants quickly. Two staffers with no subject-matter expertise can't manually review 1,162 scholarly applications in days, but they can paste 1,162 descriptions into a chatbot in days. The resulting spreadsheet looks systematic, like an algorithm produced the kill list rather than two people with a political agenda. The AI's role was to provide a paper trail that obscured the human judgment underneath, which is the laundering mechanism McMahon's ruling calls out by name.

The downstream beneficiary worth noting is OpenAI — not because the record shows any awareness of Fox's usage, but because of what the timeline reveals about sequencing. The government was using ChatGPT for consequential federal decisions in March 2025. No FedRAMP-authorized ChatGPT product existed until April 2026, over a year later. In between, the GSA-OpenAI OneGov partnership opened ChatGPT Enterprise to every federal agency at $1 per agency — brokered by the same GSA Fox and Cavanaugh were detailed from. The connection is institutional, not conspiratorial: the government normalized using commercial AI tools in consequential decisions before the authorization framework existed to govern them, and OpenAI benefited from that normalization.

And Michael McDonald, who yielded statutory authority he was supposed to exercise, got nominated by Trump in February 2026 for the permanent NEH chairperson role he was filling on an acting basis (NEH.gov). He was awaiting Senate confirmation when the ruling dropped.

What We Still Don't Know

The ruling settles the constitutional question and leaves the data question wide open. Nothing in the 143 pages, the publicly released depositions, or ACLS's discovery materials specifies which ChatGPT product Fox used. That matters because in March 2025, no FedRAMP-authorized ChatGPT product existed yet. Consumer ChatGPT, the most accessible option, trains on submitted conversations by default unless the user actively opts out. ChatGPT Enterprise wouldn't receive FedRAMP Moderate authorization until April 27, 2026, over a year after Fox processed the grants.

The version question is the one I keep coming back to — while being clear about what's known versus what isn't. If Fox used consumer ChatGPT, the default data policy in March 2025 would have used submitted conversations to improve the model unless he actively opted out. Whether that happened with his 1,162 grant submissions is genuinely unknowable from the public record: OpenAI doesn't disclose which conversations are used for training, and the ruling doesn't address it. This is speculation with a factual basis, not an established finding. It's the difference between "DOGE made a politically motivated decision using a chatbot" and a harder question nobody has formally asked yet. No inspector general at GSA, OMB, or NEH has opened a FISMA compliance investigation, which means the data question may stay unanswered.

What the Injunction Does and Doesn't Do

McMahon's order permanently enjoins the government from enforcing the terminations or re-terminating the same grants on the same impermissible basis. It does not order immediate payment, which requires a separate suit in the Court of Federal Claims, and the administration contested rather than complied with the related preliminary injunction McMahon issued July 25, 2025 (ACLS press release; Authors Guild). The practical fight over whether the money actually flows back is still on.

The administration's position — stated publicly through White House spokespeople and echoed in DOGE statements — is that the grants were wasteful DEI spending that shouldn't have existed regardless of how they were identified. That's a substantive policy argument the ruling explicitly doesn't resolve: McMahon's opinion is about how the decision was made, not whether the grants were worth funding. If Congress wanted to cut NEH's budget, it could. The question the ruling answers is narrower: DOGE used a viewpoint-based filter without statutory authority, and that's unconstitutional regardless of whether you think the grants were good ones.

The harder question is the one the ruling forces every agency to answer next time. Federal agencies are going to keep using AI tools for consequential decisions, because the procurement contracts are being signed right now. When one of those decisions violates the Constitution, the government does not get to point at the model. Whether the administration appeals, and whether a higher court narrows McMahon's "chosen instrument" precedent, is the thing still in motion.