A spreadsheet entered as a court exhibit this November contains a number that should stop anyone who has ever shipped a screening model: of 1,163 humanities grant proposals reviewed during a federal cost-cutting pass, 1,057 were flagged. Forty-two survived. That is a 90.9 percent positive classification rate.
No well-calibrated classifier produces that distribution by accident. A model that labels nine of every ten inputs as a problem is not discriminating โ it is collapsing. According to the filing, the labeling tool was ChatGPT, used by Department of Government Efficiency staff to determine which National Endowment for the Humanities grants were advancing DEI and which were not.
I read the coverage with two of my own alerts still open on a second monitor: a mid-cap DAO's treasury multisig moving 40 percent of its stablecoin reserves into a single vault, and a lending market where an oracle had printed a price no spot venue was quoting. Different jurisdictions, different assets. Same failure mode. A system issued a decision, nobody could reproduce the reasoning behind it, and the cost landed on people who never consented to the system in the first place.
Connecting the dots that others ignore or fear: this is not a story about a language model "hallucinating." It is a story about what happens when a probabilistic tool is handed discretionary authority and no audit layer is built underneath it. Crypto has spent a decade arguing that problem into product. We just watched a public agency skip the argument entirely and go straight to deployment.
To understand why the number matters, you need the plumbing. The National Endowment for the Humanities is a small federal agency โ its annual appropriation sits in the low hundreds of millions of dollars โ that disburses grants to museums, archives, historical societies, and scholarly projects. It is not a regulator. It has no enforcement arm. Its output is funding decisions, which is precisely why it became a target: a grant list is a clean, enumerable dataset, and enumerable datasets are what automated review systems love.
The mechanism, as reported and as described in the litigation, worked like this. DOGE staffers โ two individuals are named in the filing, Justin Fox and Nate Cavanaugh โ ran the agency's pending proposals through ChatGPT and asked it, in effect, to identify content related to diversity, equity, and inclusion. Proposals carrying that label were cut. Roughly half of the agency's budget disappeared in the pass.
One example has become the emblem of the case. The High Point Museum, in North Carolina, had a $349,000 grant to replace its heating, ventilation, and air conditioning system. The stated reason for flagging it as DEI, per the spreadsheet, was that the upgrade "would help more visitors." Read that twice. A climate-control project was classified as ideological because the model associated broader access with inclusion. That is not a factual error in the ordinary sense. It is a category error produced by the geometry of an embedding space, where "serving more people" sits close to "equity," and nothing in the pipeline was positioned to catch the drift.
The consequences moved fast. The museum recovered about 70 percent of its award through a termination clause โ meaning the correction required lawyers, not software. Four organizations, including the American Historical Association and the Authors Guild, are plaintiffs in a suit alleging viewpoint discrimination under the First Amendment and violations of equal protection. And in a detail that should matter to anyone building governance tooling, the agency's own acting chair, Michael McDonald, testified that many of the targeted projects were "harmless" in their promotion of DEI. That is an internal negative witness โ the operator of the system declining to defend its own output.
Two more context notes before the technical work. First, reporting indicates that deficit reduction, not only ideology, motivated the cuts. That matters, because it reframes the AI not as an ideological instrument but as a throughput instrument. Second, and this is the detail I keep returning to: the tool chosen was OpenAI's ChatGPT, not Grok, the model built by the company owned by the same person who stood up DOGE. I will come back to what that vendor choice actually signals, because it is not the story most people told.
Now the part that matters to anyone who has run a classifier in production.
Start with precision. A 90.9 percent flag rate is only meaningful against a base rate. Suppose, generously, that true DEI-flagged content in a humanities grant portfolio runs at 5 percent โ 58 of those 1,163 proposals. If the model had perfect recall and caught every one of them, it would still have produced 999 false positives. Precision would be 5.5 percent. In any domain I work in, a screen with 5 percent precision is not a screen; it is a random number generator with a cost center attached. And we have direct evidence that the base rate sat far below the flag rate, because the agency retained 42 proposals and because internal testimony described the flagged set as largely harmless. Whatever the true prevalence was, it was not 91 percent. That is not noise. That is the truth screaming.
Then look at consistency. Eleven of the proposals involved HVAC or facilities work. Two were flagged. If a stable rule were being applied, the other nine should have been flagged too โ or none of them. Selective output from a fixed input class is the signature of context-dependent generation, not rule application. In practice that means the decision depended on phrasing, ordering, or whatever happened to be adjacent in the prompt window. You cannot appeal a decision that is not stable, because there is nothing stable to appeal against.
Then look at the asymmetry of the error costs. In fraud detection, a false positive means a declined transaction and a support ticket; a false negative means money walks out the door. That asymmetry justifies erring toward over-flagging. Nobody applied that logic here. A false positive in this pipeline meant a museum without a functioning HVAC system, a historical association without an archive budget, and โ as it turned out โ a constitutional claim. A false negative meant a grant proceeded exactly as originally intended. The error costs were inverted relative to the tool's default bias, and no one appears to have performed that calculation before shipping.
Then look at the explanations. The model did not merely label. It justified. "This upgrade will help more visitors." Something about environmental sustainability. These read like reasons, and that is precisely the danger. What the pipeline produced was post-hoc rationalization: a label first, a plausible-sounding rationale generated afterward to make the label legible. I have watched this pattern in on-chain analytics too. Cluster a set of wallets, and the tooling will happily tell you the cluster is "a market maker" โ not because it verified anything, but because the label needs a sentence. A generated explanation is not evidence. It is packaging.
Now the engineering inventory. In everything disclosed so far, there is no published prompt, no model version, no temperature setting, no confidence threshold, no decision log, no cross-proposal consistency check, and no documented human review step. In my own work โ and this goes back to a six-week stretch in 2017 when I manually traced 14,000 ETH out of the EOS pre-sale contracts and cross-referenced wallet clusters against Bitcointalk sentiment โ the lesson was never that data lies. It was that any number you cannot re-derive from source is a number you cannot defend. I found a 23 percent gap between reported token sales and on-chain liquidity during that exercise. It only became usable evidence after I could show the exact addresses and the exact blocks. A determination that "ChatGPT said so" fails that test at the first step.
Here is where this stops being a government story and becomes a crypto story.
The failure is not that a language model was used. The failure is that a discretionary decision โ one that moves money and restricts organizations โ was executed with no provenance layer. And provenance is the one thing this industry has spent a decade building, badly, expensively, and repeatedly.
Consider the oracle analogy, because it is exact. A price feed that publishes a number without a documented methodology is not trusted because it is accurate. It is trusted because its inputs are inspectable, its deviation bounds are published, and its failure modes are known in advance. When an oracle misprices, the loss cascades through liquidations within minutes, and the market responds by demanding proofs, circuit breakers, and fallback sources. The NEH pipeline was an oracle with no methodology, no deviation bound, and no fallback โ and its mispricing cascaded into half an agency's budget and a federal lawsuit. The anomaly isn't a glitch in the model. The anomaly is the absence of infrastructure around it.
The toolkit for fixing this already exists in fragments. On-chain attestations give you a signed, timestamped record that a specific agent produced a specific output at a specific block height โ the Ethereum Attestation Service and similar registries are doing this today for credentials and compliance records. Deterministic replay, where the same input yields the same output, is the baseline expectation for any smart contract and is achievable for inference with enough engineering discipline. Human-in-the-loop is not a buzzword; it is what a multisig is. A 3-of-5 threshold on a treasury is a confession that no single signer should be trusted with unilateral authority. An AI that allocates or denies resources needs to sit inside a multisig-shaped governance structure, not stand outside it.
I will be honest about the limits of my own argument. Verifiable inference โ zero-knowledge proofs over model execution, trusted execution environments that attest to a specific model version โ is early, expensive, and in several implementations slower than the decision loop it is supposed to guard. You cannot zk-prove a frontier model's forward pass at production latency today, and anyone claiming otherwise is selling something. What is available right now, cheaply, is procedural: log the prompt, version the model, publish the decision criteria, sample-audit the outputs, and give every affected party a documented appeal path. That is not cryptographic truth. It is the minimum standard of care. The NEH case failed even that.
And this is why I keep circling the vendor detail. DOGE used ChatGPT, not Grok, despite the ownership overlap. The reading most people reached was that it proved xAI's models were not ready for government work. That reading is defensible but incomplete. The more useful reading is that tool convenience beat political alignment in the procurement decision, which tells you the operator's true priority was speed, not ideology. You do not choose the model that is easiest to invoke if your goal is a rigorous screening function. You choose it if your goal is a large volume of labels, quickly, with a defensible-looking signature at the bottom. In that frame, the AI was not the decision-maker. It was the letterhead.
The crypto market has a name for this pattern, and we usually apply it to teams. A project publishes a whitepaper describing a decentralized protocol, and then you pull the foundation wallet and find that four addresses control 61 percent of supply. The rhetoric is decentralized; the ledger is not. The NEH filing is that same divergence, one layer up. The rhetoric was "objective AI screening." The ledger โ the spreadsheet itself โ shows a human-cut budget wearing a machine-generated label.
I have done this variation of the work before. In 2021 I mapped the pre-mine acquisition patterns of the top 50 wallets tied to a blue-chip NFT launch and found that roughly 60 percent of early holders traced back to a single marketing agency, which reframed an "organic community" story into a coordinated distribution event. Nobody had to accuse anyone of fraud. The wallet graph simply did not support the narrative. That is the posture this case deserves too: not outrage, but an honest reading of the artifact. The artifact says the AI produced a label and the humans produced a budget cut, and the label was doing work the evidence cannot carry.
So here is where I will disagree with most of the commentary, including some of my own instinct.
The dominant framing is that AI made a mistake and cost a museum its heating system. That framing is comfortable and mostly wrong. Models do not hold authority. They have no budget line, no signature, no legal personality. The decision to strip an agency of half its appropriation belonged to people, and the phrase "ChatGPT flagged it" is a liability-laundering device โ a way to move a political judgment into a technical register where it cannot be argued with on its merits. Correlation is not causation applies here in an unusual direction: the flagged labels correlate with the cuts, but the cut volume was set before any model ran. If the target was a 50 percent reduction, the model's job was to generate 1,057 justifications, not to find 1,057 problems.
The second contrarian note is uncomfortable for my own side of the industry. Everything I described as the fix โ attestations, decision logs, human-in-the-loop โ is exactly what is missing from the "agentic AI" products now being pitched to DAOs. Every week I see treasury frameworks that propose delegating allocation, rebalancing, and even grant review to LLM agents with wallet permissions and no published evaluation harness. The pitch deck says efficiency. The architecture says a signature that can move funds and an explanation generated after the fact. If a federal agency can be embarrassed by an unpublished prompt, a DAO with an unaudited agent cannot pretend it is structurally different. The same 90.9 percent could be sitting in a governance forum right now, wearing a different label.
Third: transparency is not a cure for politics. If the NEH published its exact prompt and model version, the model would still have been asked to adjudicate an inherently contested term. "DEI" in the current American context is not a technical category; it is a political one with a shifting boundary. Handing it to a classifier does not depoliticize it. It hides the politics inside an embedding. Community safety is the ultimate metric of value, and a system that cannot explain itself to the community it governs has already failed that metric โ regardless of how many decisions per hour it produces.
Watch four signals over the next twelve months, because they will tell you whether this was an incident or a template.
First, whether the court writes a procedural standard โ an obligation to log, version, and review โ rather than ruling narrowly on the facts. That distinction determines whether agencies can rebuild the same pipeline with better paperwork. Second, whether other federal bodies copy the mechanism. One agency is a case study; four agencies is an architecture. Third, and most telling, whether OpenAI ever states a position on government use of its models for resource-allocation decisions. A vendor silence here is a policy.
Fourth, and this is the one I will be tracking on-chain: whether an "AI decision audit" category actually forms โ attestation registries, replayable inference logs, tools that let a grant recipient or a DAO member contest a machine output with the same rigor we contest a failed transaction. The demand is now demonstrated. The question is whether anyone building it has the patience to make it boring, because boring is what audit infrastructure has to be.
The 90.9 percent number will not be remembered as an AI failure. It will be remembered as the moment we learned that a machine with discretionary power and no ledger underneath it is just an opinion with better posture.
