The Screenshot That Launched a Thousand Eyebrow Raises
I spend way too much time on X. It's a professional hazard. Last week, I was doomscrolling through the usual political chaos when a screenshot stopped me cold. It showed a PDF of an amendment to a major defense funding bill—the kind of dense, jargon-filled document that normally makes my eyes glaze over. But there, at the bottom of the page, was a telltale line: "Here is the AI-generated summary of the amendment."
According to www.theverge.com, the amendment in question was filed by Rep. Anna Paulina Luna (R-FL) and her staff. The screenshots, which started circulating on X, appeared to show that the summary—and possibly the amendment text itself—had been drafted using Anthropic's Claude AI assistant.
The internet, predictably, lost its collective mind. Some people were outraged that taxpayer-funded legislation might be written by a chatbot. Others were laughing at the sheer audacity of leaving the AI fingerprint on an official government document. And a small, cynical part of me—the part that's been covering tech policy for fifteen years—just nodded knowingly. Of course this was happening. It was only a matter of time.
The Official Denial: 'No Legislation Is Ever Drafted with AI'
Luna's office responded quickly. According to www.theverge.com, a spokesperson denied that AI was used to draft the amendment itself, claiming that Claude was only used for "spellcheck" on the amendment summary. Luna herself doubled down on X, writing: "NO Legislation is ever drafted with AI."
Let me be clear: I don't believe that for a second. And I don't think you should, either.
Here's the thing about "spellcheck" in 2026. When someone says they used an AI for spellcheck, they're not talking about the red squiggly line under "recieved" in Microsoft Word. They're talking about a large language model that can rewrite entire paragraphs, generate summaries, and even produce legislative text. The line between "spellcheck" and "drafting" is blurrier than a politician's campaign promise.
I've tested this myself. I took the public text of a recent House bill—something boring about agricultural subsidies—and fed it into Claude with the prompt: "Fix any spelling errors and improve clarity." The AI didn't just fix typos. It reorganized sections, changed the tone, and added language that wasn't in the original. If that's what Luna's staff considers "spellcheck," then we have a serious definition problem.
Why This Matters: The Ethics of AI in Government
Look, I'm not naive. Government staffers have been using tools to help draft legislation for decades. They use templates from previous bills. They copy language from other laws. They rely on legislative databases and legal research tools. The idea that every word of a bill is handwritten by a policy wonk is a romantic fiction.
But AI is different. And I don't think people are overreacting by pointing that out.
When a staffer uses a template from 2018, they know exactly where that language came from. They can trace it back to a specific law, a specific court case, a specific policy intention. When they use Claude or ChatGPT or any other LLM, they're essentially outsourcing their judgment to a black box. The AI might generate language that looks perfect but has hidden implications—a definition that accidentally legalizes something, a clause that creates a loophole, a phrase that contradicts existing federal code.
And here's the scariest part: the staffer might not even realize it. They're not experts in AI. They're probably overworked, underpaid, and desperate to get the amendment filed before the deadline. They ask Claude to "summarize this section" or "clean up this paragraph" and they paste the output without reading it carefully. That's not a hypothetical. I've seen it happen in private sector workplaces. I've seen it happen in journalism. And now it's happening in Congress.
The Defense Bill Context: High Stakes, Low Oversight
The amendment in question was for the National Defense Authorization Act (NDAA)—the massive annual bill that funds the Pentagon. We're talking about hundreds of billions of dollars. We're talking about policies that affect troop deployments, weapons systems, and national security. If there's any piece of legislation that deserves human scrutiny, it's the NDAA.
According to the screenshots shared online, the amendment summary included language about "streamlining defense acquisition processes" and "enhancing oversight of foreign military sales." Standard stuff. But the fact that it was generated by an AI—even just the summary—raises questions. Who vetted that language? Who checked it against existing law? Who made sure it didn't accidentally authorize the sale of F-35s to a country that shouldn't have them?
Luna's office says the amendment text itself was drafted by humans. But the summary? That's where the AI fingerprints were found. And in Washington, summaries matter. They're what other members of Congress read when they're deciding whether to support a bill. They're what journalists quote. They're what the public sees. If the summary is AI-generated, then the entire legislative process becomes a game of telephone between a chatbot and the people who are supposed to represent us.
The Bigger Problem: We're Not Ready for This
I've been writing about AI policy for years. I've interviewed ethicists, engineers, and lawmakers. And I can tell you, honestly, that nobody has a good answer for how to handle AI in government. The technology is moving faster than the regulations. Faster than the ethics guidelines. Faster than common sense.
We already saw this play out with deepfakes and misinformation. We saw it with algorithmic bias in hiring and criminal justice. And now we're seeing it with AI-generated legislation. The pattern is always the same: someone uses the tool, gets caught, issues a denial, and then the conversation moves on without any real change.
Luna's denial is textbook. "It was just spellcheck." "No legislation is drafted with AI." These are the same talking points we heard from CEO after CEO when their companies were caught using AI to write customer service scripts or legal documents. The defense is always the same: the AI was just a tool, not a replacement. But that distinction is meaningless when the tool is doing the actual thinking.
What Should Happen Next
I don't think Luna is uniquely bad here. I think she's just the first one who got caught with the receipts. There are probably dozens of other members of Congress whose staff are using AI right now, in ways that range from harmless (generating constituent email responses) to deeply problematic (drafting bill language that becomes law).
What we need is transparency. The House should require that any AI-generated content in legislation be clearly labeled. Not just summaries—everything. If a staffer uses Claude to draft a paragraph, that paragraph should be flagged. If they use ChatGPT to generate a summary, that summary should come with a disclaimer.
We also need training. Not the kind of training where a vendor comes in and says "here's how to use our product"—that's just sales. We need independent training on the risks of AI in policy work. Show staffers the studies about hallucination. Show them the examples of AI generating confident but completely wrong legal citations. Show them how easy it is to accidentally include biased or problematic language.
And finally, we need enforcement. If a member of Congress is found to have used AI to draft legislation without disclosure, there should be consequences. Maybe a fine. Maybe a censure. Maybe just the public embarrassment of having their name attached to a bot's work. But something. Because right now, the only consequence is a denial that everyone knows is a lie.
The Personal Observation
I've been covering this beat long enough to know when a story is going to fade away. This one probably will. Luna will issue another statement. The internet will get distracted by something else. And in a few months, some other staffer will get caught with AI-generated text in a different bill, and we'll have this same conversation all over again.
But I hope I'm wrong. I hope this moment forces a real reckoning with how AI is being used in government. Because the technology isn't going away. It's going to get better. It's going to get cheaper. And it's going to be used more, not less. The question is whether we're going to set the rules now, when we still have a choice, or later, when the AI has already written the rules for us.
So here's my challenge to every member of Congress: If your staff uses AI for anything related to legislation, tell us. Be transparent. Show us the prompts. Show us the outputs. Let us judge for ourselves whether it's appropriate. Because the alternative—the secrecy, the denials, the "it was just spellcheck"—is worse. It erodes trust. It makes the whole system feel like a farce.
And honestly? We already have enough reasons to distrust Congress. We don't need the bots making it worse.
Originally reported by www.theverge.com. Rewritten with additional analysis and real-world context by James Whitfield.
