DOGE Used ChatGPT to Gut Humanities Grants. This Is Exactly What We Warned About.

DOGE Used ChatGPT to Gut Humanities Grants. This Is Exactly What We Warned About.

March 9, 2026 • 7 min read

Elon Musk’s short-lived Department of Government Efficiency (DOGE) just showed us exactly what happens when you let ChatGPT make policy decisions. Spoiler: it’s worse than you think.

The Conventional Wisdom

“AI will help governments make better, faster, more objective decisions.”

That’s the pitch. Strip away human bias. Use data-driven analysis. Let algorithms find waste and inefficiency that humans might miss. Pure rationality, zero politics.

Sounds great, right?

Why That’s Wrong

According to The New York Times, here’s what actually happened when DOGE rolled into the National Endowment for the Humanities (NEH) with orders to cancel grants deemed contrary to Trump’s anti-DEI agenda:

They didn’t read the grants.

They didn’t analyze the projects.

They didn’t consult experts.

Instead, they:

  1. Pulled short summaries off the internet
  2. Fed them into ChatGPT
  3. Asked: “Does the following relate at all to D.E.I.? Respond factually in less than 120 characters. Begin with ‘Yes’ or ‘No.’”
  4. Cancelled grants based on the output

The results were, according to the Times, “sweeping, and sometimes bizarre.”

What’s Actually Happening

This isn’t “AI-assisted decision making.” This is decision outsourcing—and it’s the laziest, most reckless version possible.

Let’s break down what went wrong:

1. Garbage In, Garbage Out

They didn’t even use the actual grant applications. They scraped “short summaries off the internet”—meaning random descriptions that could be incomplete, inaccurate, or written by someone who didn’t understand the project.

Imagine if your job performance review was based on your LinkedIn bio instead of your actual work. That’s the level of rigor we’re talking about.

2. The Prompt Was Terrible

“Does the following relate at all to D.E.I.?”

This is a yes/no question about an inherently complex, context-dependent concept. DEI can mean:

ChatGPT has no way to distinguish between these. It just pattern-matches on keywords.

3. ChatGPT Isn’t “Factual”

The prompt says “Respond factually”—as if that’s a magic word that turns ChatGPT into an objective truth machine.

It’s not.

ChatGPT is a language model. It predicts text. When you ask it to classify something, it’s making probabilistic guesses based on patterns in its training data—patterns that include all the biases, inconsistencies, and cultural assumptions of the internet.

Asking ChatGPT to be “factual” about whether a humanities grant “relates to DEI” is like asking a Magic 8-Ball to diagnose cancer. The format looks authoritative, but there’s no actual reasoning happening.

4. The 120-Character Limit

Why 120 characters? Because that’s a tweet. DOGE literally asked ChatGPT to make complex policy decisions in tweet-sized responses.

Nuance? Context? Proportionality? Gone. You get “Yes” or “No” and maybe a sentence fragment.

Why This Matters

This isn’t just about one badly-run government agency (though DOGE no longer exists, thankfully).

This is about what happens when decision-makers treat AI as a black box they can hide behind.

The Accountability Problem

When a human makes a bad decision, you can ask them why. You can challenge their reasoning. You can hold them responsible.

When an AI makes a bad decision, everyone shrugs and says “the algorithm decided.” No one’s accountable. No one has to defend their reasoning. The AI becomes a bureaucratic shield.

The Automation of Ideology

DOGE didn’t use ChatGPT to find waste or inefficiency. They used it to automate a predetermined ideological agenda.

The decision to target DEI-related grants was already made. ChatGPT was just the tool to scale it up without having to read, think, or justify individual cases.

This is AI as a force multiplier for lazy authoritarianism.

The Precedent

If this had worked—if DOGE had successfully used ChatGPT to gut the NEH without major backlash—it would’ve become the template.

Imagine:

All with the same level of rigor: scrape some text, ask a yes/no question, let the bot decide.

What You Should Do

If you work in government, tech, policy, or any field where AI tools are being proposed for decision-making:

Ask these questions:

  1. What data is the AI actually using?

    • Is it complete? Accurate? Representative?
    • Or is it “short summaries off the internet”?

  2. What’s the prompt?

    • Is it asking for nuanced analysis or binary answers?
    • Does it acknowledge uncertainty and context?
    • Or is it “respond factually in 120 characters”?

  3. Who’s accountable?

    • If the AI gets it wrong, who takes responsibility?
    • Can someone explain why a specific decision was made?
    • Or is it just “the algorithm said so”?

  4. What’s the human review process?

    • Are people actually checking the AI’s work?
    • Or are they rubber-stamping whatever it spits out?

  5. What’s the real goal here?

    • Is AI genuinely improving the process?
    • Or is it just automating an agenda while avoiding scrutiny?

The Counterargument

“But humans make biased decisions too!”

Absolutely. Human decision-making is flawed, inconsistent, and often biased.

But here’s the difference:

Humans can be interrogated. You can demand their reasoning. You can appeal their decisions. You can hold them accountable in court, in the media, or at the ballot box.

Humans can exercise judgment. They can weigh context, recognize edge cases, and say “this doesn’t fit the template.”

Humans have institutional knowledge. They understand the history, purpose, and nuances of the systems they’re working in.

ChatGPT has none of that. It’s a text prediction engine. It doesn’t understand what a humanities grant is, let alone whether canceling one is justified.

The problem isn’t that ChatGPT is worse than humans at everything—it’s that it’s being used for tasks that require human judgment and then being treated as if it provided that judgment.

Final Thoughts

The DOGE/ChatGPT debacle is a case study in what not to do with AI.

It’s not “efficiency.” It’s not “objectivity.” It’s outsourcing responsibility to a system that can’t be held accountable.

The good news? DOGE is gone. The grants are being reviewed by actual humans with actual expertise.

The bad news? This won’t be the last time someone tries to use ChatGPT (or Claude, or Gemini, or whatever) as a bureaucratic magic wand.

We need to stop pretending that asking an LLM a question is the same as doing analysis.

We need to stop treating AI-generated text as “objective” just because it came from a machine.

And we need to start holding people accountable for the decisions they’re delegating to AI—not letting them hide behind “the algorithm decided.”

Because the algorithm didn’t decide. Someone decided to use the algorithm. And they need to answer for what it does.

What Actually Works

If you do want to use AI responsibly in government or policy contexts:

✅ Use AI to assist experts, not replace them
Have ChatGPT summarize documents, flag potential issues, or generate draft analyses—then have humans review, contextualize, and decide.

✅ Use AI for triage, not final decisions
ChatGPT can help sort through thousands of applications to find ones that need closer human review. It should never be the final word.

✅ Document everything
If AI is part of your process, publish the prompts, the data sources, and the review procedures. Transparency is the only defense against algorithmic opacity.

✅ Build in appeals processes
Anyone affected by an AI-assisted decision should be able to request human review—and actually get it.

✅ Test for edge cases and bias
Before deploying AI at scale, test it on hundreds of real cases with known outcomes. If it’s making “bizarre” decisions in testing, don’t deploy it.

The DOGE approach violated every single one of these principles. Don’t be like DOGE.

Bottom line: AI is a tool. Like any tool, it can be used well or badly. DOGE used it badly—and the NEH’s grants paid the price. Let’s not let this become the new normal.