from the yo-chatgpt,-is-this-legal? dept
Back in early 2025, DOGE bros Justin Fox and Nate Cavanaugh were handed multiple government roles with a simple mandate: go find the woke stuff and kill it. Fox and Cavanaugh had zero relevant experience for any of it — at one point Cavanaugh, a twenty-something college dropout whose main prior credential was a patent startup, was put in charge of the United States Institute for Peace. One of their assignments was the National Endowment for the Humanities, which they apparently concluded was irredeemably woke and needed to go.
Their review process for millions of dollars in previously approved grants consisted almost entirely of asking ChatGPT this:
“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. Do not use ‘this initiative’ or ‘this description’ in your response.”
And that was it. They then relied on the results of that as a reason to cancel millions of dollars in grants that had already gone through a detailed approval process. You’ll also recall, that these two bros who probably thought themselves masters of the universe in early 2025, flipped out that the plaintiffs in the case against them, the American Council of Learned Societies, put their depositions on YouTube, where you could see them unable to define DEI, and unable to try to defend why they canceled various grants for being woke. Turns out they didn’t like facing any scrutiny themselves.
Judge Colleen McMahon has now dropped a scathing 143-page ruling finding that feeding grants to ChatGPT and relying on its “why this is woke in 120 characters” output meets the “arbitrary and capricious” standard — making these cuts unlawful.
Fox testified that he did not define “DEI” for ChatGPT and that he did not have the slightest idea how ChatGPT understood the term…. Nor did Fox ask ChatGPT to factor in the purpose, methodology or scholarly substance of a project – which would have required familiarity with the underlying grant materials
Because the inquiry was framed only in terms of whether a project “relate[d] at all to DEI” (whatever that might mean to ChatGPT), projects whose abbreviated descriptions contained general references to “history,” “culture,” or “identity” were frequently identified by ChatGPT as relating to DEI. For example, a project to recover and analyze ancient writings attributed to Moses but excluded from the canonical Hebrew Bible and preserved in fragmentary form (e.g., the Book of Jubilees and the Testament of Moses) was classified as DEI because it claimed to “provide important insight into Jewish thought from two thousand years ago, complementary to insight from the Dead Sea Scrolls and the New Testament.”… The project description reflects a highly technical effort involving multispectral imaging, textual reconstruction, and the recovery of deteriorated source materials…. Yet, the assigned rationale consists of a single sentence linking the project to “Jewish thought,” which ChatGPT considered to be aligned with “DEI” goals.
The judge highlights some other whoppers as well:
nother grant supported a study of the Chinese government’s persecution of the Uyghur people. As described in the spreadsheet, the project documented surveillance practices, detention facilities, coercive assimilation policies, restrictions on language, religion, and cultural practice, and the effects of those policies on Uyghur communities in China and in the diaspora. In other words, the project concerned state policy, human-rights conditions, and the preservation of cultural and religious identity under conditions of repression by the Chinese government. Yet ChatGPT classified the project as “DEI,” Dkt. No. 248-11, at 22, apparently because it concerned the threatened erasure of a particular ethnic and religious group, albeit one located in a foreign country. Disfavoring this grant on “DEI” grounds, or any grounds, is especially difficult to square with the United States’ longstanding, bipartisan condemnation of China’s treatment of the Uyghurs, including during the first Trump Administration.
Judge McMahon clearly understands that ChatGPT is a “generate approximately what you want” engine, not a tool for deeply analyzing grant applications.
The record reflects that these ChatGPT determinations were generated without any additional context beyond the cursory spreadsheet descriptions themselves. Given what courts now know about the hallucinatory propensities of ChatGPT and similar generative-AI tools, it would hardly be surprising if ChatGPT inferred, from DOGE’s repeated requests, that Fox and Cavanaugh were looking for reasons why grants could be characterized as DEI – and therefore terminable – and supplied “rationales” simply in order to satisfy the user’s perceived demand. The utter lack of reasoning behind so many of its “rationales” certainly suggests as much.
This is one of the most frustrating things in lots of other stories about LLMs, in which the media and politicians and everyday users insist some sort of actual intelligence, thought, or intent behind outputs. When all any of them do is try to generate a plausible sounding response to your request. Judge McMahon understands that. And either Fox and Cavanaugh are too technically illiterate to understand that or, more likely, they didn’t give a shit. They just wanted to check boxes so they could cancel as many grants as they wanted.
ChatGPT is certainly good at doing your homework for you, though.
The only problem is that the law requires actual reasons and a real process for something like this — it can’t just be arbitrary. But feeding it to ChatGPT to give you the reasons why you can cancel these grants, is absolutely arbitrary.
There can be no serious dispute that the review process implemented by DOGE did not conform to, or even resemble, NEH’s ordinary grant-review process, which itself was designed to comply with the mandates of its authorizing statute. The statutory process contemplates individualized evaluation by subject-matter experts, advisory review, Council involvement, and final action by the Chairperson. See supra Sections I(A), (C). And once grants had been awarded, the Act does not authorize a wholesale post-award revocation based on new ideological criteria. Rather, for awards made under § 956(c), the Act provides for post-award evaluation, financial and project reporting, and potential termination only when the Chairperson determines that a recipient has substantially failed to satisfy the purposes for which the assistance was provided or the applicable statutory criteria.
DOGE’s process, by contrast, rested on abbreviated spreadsheet descriptions and a binary ChatGPT prompt asking whether each project “relate[d] at all to DEI.” Fox then folded the ChatGPT-generated rationales into the final spreadsheets used to identify grants for termination, thereby combining DOGE’s AI-generated classifications with NEH staff recommendations. … Nothing in the record suggests that NEH had ever before employed such a process before March 2025. And nothing in the authorizing statute permitted it.
There’s also the fact that DOGE has no authority to cut off these grants:
DOGE had no statutory authority to terminate NEH grants. And on the undisputed evidence, DOGE – not the NEH Chairperson or anyone else at NEH – effectuated the terminations at issue here.
This is not a case involving a routine disagreement over how to interpret a statute. Nonstatutory ultra vires claims are reserved for extraordinary cases in which “the agency action go[es] beyond mere legal or factual error and amount[s] to a clear departure by the [agency] from its statutory mandate.” … They require a showing that the agency “plainly acts in excess of its delegated powers and contrary to a specific prohibition in the statute that is clear and mandatory.”….
This case presents something more basic. It is not that DOGE misconstrued a statutory provision conferring authority on it; it is that Congress conferred no authority on DOGE at all with respect to the awarding, continuation, or termination of NEH grants.
The Supreme Court has made it clear that Congress — not the executive branch — has the power here. (Amusingly, the GOP loves to point this out, but only when Democrats are in the White House.)
The Executive Orders and DOGE-directed termination process did not carry out the NEH grantmaking scheme Congress prescribed by law in a duly enacted statute. They displaced it. Congress vested funding authority in the NEH Chairperson. DOGE substituted a presidential policy judgment for a statute duly enacted by Congress. That is not faithful execution of law
Judge McMahon even points out that the Trump administration’s posturing regarding DOGE in other cases is part of what sinks it in this case:
As the Solicitor General represented to the Supreme Court, “USDS [DOGE] is a presidential advisory body within the Executive Office of the President” and “USDS has no organic statute and is not created by Congress.” Appl. to Stay Pending Certiorari or Mandamus and Request for Immediate Administrative Stay at 5, 21, In re U.S. DOGE Serv., No. 24A1122 (emphasis added). If DOGE “has no organic statute and is not created by Congress,” then it necessarily lacks any statutory delegation of authority to terminate NEH grants. And it certainly cannot exercise such authority in derogation of the very purposes that Congress wrote into the agency’s authorizing statute.
The court goes on to explain that an executive order simply cannot overrule Congress, but more to the point, Trump’s executive order setting up DOGE in the first place itself admits that it doesn’t have the powers it sought to exercise in cancelling these grants.
… the Executive Orders, read on their own terms, confirm the absence of any operative grant-termination authority in DOGE. Executive Order 14158, Establishing and Implementing the President’s “Department of Government Efficiency,” first creates the principal DOGE entity by renaming the United States Digital Service as the “United States DOGE Service (USDS)” and providing that it “shall be established in the Executive Office of the President.” Exec. Order No. 14158 § 3(a), 90 Fed. Reg. at 8441. It then separately creates, “within USDS,” a “temporary organization known as ‘the U.S. DOGE Service Temporary Organization’” and states that this temporary organization is created “in accordance with section 3161 of title 5, United States Code.” Id., § 3(b), 90 Fed. Reg. at 8441. But the only reference to a statute in the Order – to 5 U.S.C. § 3161 – attaches to the temporary organization, not to the principal DOGE entity.
Section 3161 itself does not confer any substantive policymaking power on DOGE. It defines a “temporary organization” as an organization “established by law or Executive order for a specific period not in excess of three years for the purpose of performing a specific study or other project,” and grants personnel, detail, compensation, and travel authority to staff such an organization. 5 U.S.C. § 3161(a)–(e). Nothing in that provision authorizes DOGE to exercise independent executive power, direct the functioning of a federal agency, terminate grants, halt payments, or impose binding decisions on statutory officers. At most, it permits the creation and staffing of a temporary body for a “specific study or other project.”
Pointing to another DOGE EO, the judge points out that the federal government has publicly stated that DOGE was only to have an advisory role. But all of the details here show that the two bros were absolutely in charge (frequently overruling the staffers ostensibly in charge).
DOGE Teams – including those in which individuals such as Justin Fox or Nate Cavanaugh serve – are expressly created “within” agencies and function to “advise” agency heads – not to exercise independent decisionmaking authority.
And the evidence is overwhelming that Cavanaugh and Fox were not “advising.” They were ordering.
Any suggestion that DOGE merely advised NEH is belied by the undisputed record. They did not simply provide policy advice for NEH leadership to accept or reject after their review. Before ever meeting with NEH leadership, Fox had already begun reviewing NEH grants using keyword searches and DOGE-generated categories to identify grants he considered objectionable. DOGE then used ChatGPT-generated “DEI” rationales, combined those classifications with NEH staff ratings, and generated the final spreadsheets used to identify grants for termination. See supra Sections I(E)–(G). Cavanaugh and Fox then “both rejected [the NEH Chair’s] recommendation” to continue funding certain grants, Dkt. No. 276-1, Cavanaugh Dep., 181:3–9, and added still more grants for termination – grants that NEH staff had identified as “N/A” – on the eve of the April 2, 2025 mass termination, Dkt. No. 248-32, US-000041206. The appearance of Chairperson McDonald’s name on the termination letters does not alter the substance of what occurred. Fox, not McDonald, drafted and sent the termination notices under McDonald’s name.
Just days before the Mass Termination, Fox told McDonald, “We need a game plan for effectuating . . . final grant terminations and contract cancellations by tomorrow AM. We will carry these plans out before the end of the week. 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.” Dkt. No. 248-15, Ex. 15, US-000050717 (emphasis added). That is not the language of an adviser awaiting a decision from the statutory decisionmaker. Fox told McDonald that “we” (i.e., Fox and Cavanaugh) would carry the terminations out, invoked “pressure from the top” (i.e., the White House), and framed McDonald’s participation as simply optional. See Dkt. No. 248-1, Fox Dep., 305:8–17. His reference to whether McDonald was “no longer interested” makes clear that DOGE understood the terminations would proceed with or without him. No reasonable factfinder could read that communication as evidence of a “purely advisory” role. Fox’s message leaves no serious doubt about who was directing the process.
In other words, Justin Fox’s faux tough-guy “I’m in charge here” attitude is a big part of what sank this case. These guys were so full of their own shit, and so uninterested in how government actually works, that they effectively blew up their legal position by simply being assholes.
The record admits of only one conclusion. DOGE was calling the shots. DOGE controlled the lists, DOGE rejected McDonald’s suggestions, DOGE added additional grants for termination, DOGE drafted and sent the notices, and DOGE dictated which previously awarded grants would live or die. McDonald’s one and only decision – to allow his name to appear at the bottom of the termination letters – alters none of that. His signature did not transform DOGE’s decision into an NEH decision; his name was simply a fig leaf designed to make it appear like the official authorized by law to terminate grants was the official who had done so. If DOGE were in fact limited to an advisory role, Fox and Cavanaugh could not have directed NEH to terminate specific grants over the objections of its Chair. But they did. And they did so, without any statutory delegation of power. The challenged conduct cannot be reconciled with the National Foundation on the Arts and the Humanities Act or the constitutional requirement that executive action be grounded in lawful authority. It cannot stand.
Beyond these problems, the Court notes, the grant cancellations also were a clear First Amendment violation, given that the cancelled grants were chosen because of their expression being too woke, which is clearly targeting speech for what it says.
In sum, while the government retains some discretion in administering grant programs, that discretion is bounded by the First Amendment. It may define programmatic objectives and allocate resources accordingly, but it may not deny or withdraw funding because it disagrees with the ideas expressed. “[I]deologically driven attempts to suppress a particular point of view are presumptively unconstitutional[.]” Rosenberger, 515 U.S. at 830. Where, as here, the government funds a program that facilitates private expression, it may not act where “the specific motivating ideology or the opinion or perspective of the speaker is the rationale for the restriction.”
And here, there’s very clear viewpoint discrimination:
DEI is, of course, a viewpoint….
The record establishes that the termination decisions were driven by an expressly ideological method of classification. There can be no genuine dispute about this point. Indeed, the Government has all but admitted as much….
And thus, the cancellations were based on viewpoint, which violates the First Amendment:
Regardless of whether DOGE’s classifications were coherent (obviously, many of them were not), what matters is that it deliberately sought to single out and eliminate grants based on its perception that they implicated disfavored ideas. See Dkt. No. 248-17, Ex. 17, NEH_AR_000013 (“[W]e have only excluded [from termination] the ones that seem not to conflict with the Administration’s priorities[.]”). The Government’s actions cannot stand. It may not “restrict the speech of some elements of our society in order to enhance the relative voice of others.” Buckley v. Valeo, 424 U.S. 1, 48–49 (1976) (per curiam)
Indeed, the judge notes that the White House literally bragged about cancelling these grants because of their perceived viewpoint:
The Government, for its part, has never claimed that these grants were reviewed for their humanistic merit or that they were terminated for any reason other than the DOGE decisionmakers’ perception that they espoused, reflected, or related to the disfavored “DEI” viewpoint. To the contrary, the Government admits that these grants were terminated because they were deemed to promote DEI…. DOGE itself publicly characterized the terminations as “DEI,” posting from its official account on X (formerly Twitter) that “During the previous administration, the National Endowment for the Humanities (NEH) awarded the following grants to spend taxpayer dollars, all of which have been cancelled ($163M in overall savings). NEH grants will be merit-based and awarded to non-DEI, pro-America causes.”… The Government’s own public statements confirm what the record otherwise makes clear: DEI, the disfavored viewpoint, was the reason why the majority of Plaintiffs’ grants were terminated. The Government engaged in blatant viewpoint discrimination.
Then there was the political discrimination. Apparently the DOGE crew asked to only look at grants from the Biden admin, which the court notes is impermissible. While it correctly states that new administrations are allowed to set their own priorities, it cannot claw back previously granted money entirely on the basis of disfavored speech or political association.
The Court is under no illusion that a new administration may prefer, on a going-forward basis, to fund certain kinds of scholarly programs over others. But lawful priority-setting is one thing; depriving someone of a benefit previously conferred based on viewpoint and perceived political association is something else entirely. Consistent with Regan and Rust, a new administration may pursue lawful funding priorities within the bounds of the NEH statute and the Constitution. But it has no license to suppress disfavored ideas.
Fox and Cavanaugh walked into a federal agency, fired up ChatGPT, and apparently assumed that generating plausible-sounding bureaucratic language was the same thing as following the law. It wasn’t. A 143-page ruling says so in considerable detail. The government can set its funding priorities. It cannot, however, use a chatbot to manufacture pretextual rationales for suppressing ideas it dislikes, hand authority to people Congress never authorized, and then put a fig leaf signature on the termination letters and call it a day.
They were cosplaying as government while dismantling it from the inside. And at least in this court, they didn’t get away with it.
Filed Under: 1st amendment, chatgpt, dei, doge, elon musk, free speech, grants, justin fox, nate cavanaugh, neh
You must be logged in to post a comment Login