The Screenshot That Started a Firestorm
Last week, I was doom-scrolling through X (the platform formerly known as Twitter) when I saw a screenshot that made me do a double take. It showed what appeared to be an amendment summary for a major defense funding bill, complete with a telltale sign of AI involvement: a line that read "This amendment is drafted with the assistance of Claude, an AI assistant created by Anthropic." The post was from a user who had apparently been digging through the House's legislative repository. The account, which has since been inundated with replies, claimed that Rep. Anna Paulina Luna (R-FL) had let her staff use Anthropic's Claude to draft actual legislation.
Let me be clear: this is not a small thing. The National Defense Authorization Act (NDAA) is one of the most consequential bills Congress passes each year. It determines how the Pentagon spends hundreds of billions of dollars, funds weapons systems, sets troop levels, and often becomes a vehicle for all sorts of policy riders. If AI is drafting amendments to that bill, that raises some really uncomfortable questions about accountability, transparency, and who — or what — is actually writing the laws that govern our military.
According to www.theverge.com, Representative Luna responded quickly and forcefully. She said her staff had used AI only for "spellcheck" on the amendment summary, not for the actual bill text, and insisted that "NO Legislation is ever drafted with AI." Her office later clarified that the AI disclosure line was added by a staffer who was experimenting with the tool but that the substantive content of the amendment was written by humans.
Here's the thing: I believe her. Partially. The political cost of admitting you used AI to draft legislation is enormous right now, especially for a Republican who has positioned herself as a culture warrior against "woke" tech. But the denial reveals something deeper about how Washington is handling the AI revolution. Nobody wants to admit they're using the tools, but everyone is.
The Spellcheck Defense: Convenient or Cowardly?
Let's talk about that "spellcheck" defense for a minute. Because honestly, it's kind of wild when you think about it. If I told you I used Microsoft Word's spellcheck to write a 200-page legal document, you'd shrug. That's normal. But if I told you I used an AI assistant that generates entire paragraphs of text based on a prompt, you'd probably want to know where those paragraphs came from. The line between "assistance" and "authorship" is getting blurrier by the day, and Congress is not equipped to handle that nuance.
I've been covering tech policy for over a decade now, and I've watched this pattern repeat itself. A new technology emerges. Someone in government uses it sloppily. They get caught. They deny, deflect, or blame a junior staffer. Then, a few years later, everyone is using it openly, and we move on to the next scandal. Remember when politicians got caught using AOL email accounts? Or when they were found to be using private email servers? Each time, the initial reaction was outrage, followed by normalization.
The difference here is that AI is not just a communication tool. It's a creative tool. It can generate plausible-sounding text that is factually wrong, legally questionable, or politically biased. If an AI drafts an amendment that accidentally defunds a critical program or, worse, inserts language that benefits a specific contractor, who is responsible? The AI? The staffer who prompted it? The representative who submitted it? The company that made the AI? The answer right now is: nobody knows.
The Monetization Angle: Who Wins When AI Writes the Law?
Now, let's get to the part that interests me most: the money. Because this story is not just about a congresswoman's PR headache. It's about the massive, largely invisible ecosystem of companies that are positioning themselves to profit from AI in government.
Think about it. If AI can draft amendments, it can draft policy memos, regulatory impact analyses, and even entire bills. The market for legislative drafting software is already worth hundreds of millions of dollars. Companies like Bloomberg Government, FiscalNote, and Quorum already sell tools that help lobbyists and congressional staffers track legislation, analyze voting patterns, and draft documents. They are all racing to add generative AI features. Anthropic, the company behind Claude, has been aggressively courting government clients. According to www.theverge.com, Anthropic has a dedicated public sector team that has been meeting with congressional offices to pitch Claude as a tool for "research and drafting assistance."
Here's where it gets interesting. If a company like Anthropic can get its AI embedded in the legislative process, it gains something far more valuable than a software contract: it gains influence over the language of the law. Every time a staffer uses Claude to "improve" a paragraph, the AI is making choices about word selection, emphasis, and framing. Those choices reflect the biases baked into the training data. And those biases can have real-world consequences for defense contractors, tech companies, and every other industry that depends on federal funding.
I spoke with a former Senate staffer who asked to remain anonymous because he still works in policy. He told me: "The lobbyists are already using AI to draft proposed amendments and then submitting them to friendly offices. The staffers are overwhelmed. They just paste the text in, make a few tweaks, and submit it. Nobody has time to check every line. The AI companies know this. They're selling speed, but what they're really selling is the ability to shape the outcome."
That's the monetization angle that nobody is talking about. The real money in AI for government is not in selling chatbots to answer constituent emails. It's in selling the tools that write the rules. If you can control the drafting process, you can control the outcome. And the companies that get there first will have an enormous advantage.
The Defense Industry's Quiet AI Gold Rush
Let's zoom out from the Luna kerfuffle and look at the bigger picture. The defense industry is already in the middle of a massive AI gold rush. The Pentagon's budget for AI-related projects has grown from essentially zero a decade ago to over $1.8 billion in the 2025 fiscal year, according to a report from the Government Accountability Office. Companies like Palantir, Anduril, and Shield AI are building everything from autonomous drones to AI-powered targeting systems. But the less glamorous side of this boom is the software that helps manage the bureaucracy.
Every year, the Department of Defense produces thousands of pages of regulations, acquisition guidance, and policy documents. The people who write those documents are overworked, underpaid, and increasingly turning to AI tools to help them keep up. If a mid-level Pentagon analyst uses Claude to draft a memo about procurement rules, and that memo eventually becomes the basis for a formal policy, we have just outsourced a piece of defense policy to a private company's algorithm. And that algorithm is not transparent. It is not accountable. And it is certainly not subject to the Freedom of Information Act.
Representative Luna's amendment was about defense funding. I don't know the specifics of what it proposed, but the fact that it was part of the NDAA means it had the potential to shift millions of dollars. If AI helped write it, even just for spellcheck, the process by which that text was generated matters. The disclosure line, if it was real, suggests that someone on her staff was aware of the ethical gray area and tried to flag it. But the denial suggests that the official position is to pretend the tool doesn't exist.
The Accountability Problem No One Wants to Solve
Here is the core tension: Congress is supposed to be the branch of government that writes the laws. That is its primary constitutional function. If members of Congress are delegating that responsibility to AI systems, even partially, they are undermining the very idea of representative democracy. We elect people, not algorithms. We hold people accountable at the ballot box. You cannot hold an algorithm accountable.
But the truth is more complicated. Congressional staffers are stretched thinner than ever. The average House member has 18 staffers to cover everything from constituent services to policy research to press relations. The idea that every amendment is handcrafted by a policy expert is a fantasy. Most amendments are recycled from previous years, copied from lobbyist proposals, or drafted by committee staff who are already overwhelmed. AI is not replacing human judgment in most cases. It is filling a gap that already existed.
The question is whether we are okay with that gap being filled by for-profit companies whose incentives do not always align with the public interest. Anthropic, for all its talk of "constitutional AI" and safety, is a venture-backed startup that needs to grow revenue. Its investors include Google, Salesforce, and a host of Silicon Valley venture firms. They are not in the business of improving democracy. They are in the business of selling software.
What This Means for the Rest of Us
I realize this story might feel inside-baseball. A congresswoman's staffer uses AI for spellcheck. So what? But I think it's a canary in the coal mine. The Luna incident is going to be the first of many. As AI tools become more integrated into the legislative process, we are going to see more scandals, more denials, and more revelations that the laws we live under were written with significant machine assistance.
The companies that make these tools are going to make a lot of money. Anthropic, OpenAI, and Google are all positioning themselves as essential partners for government. They are hiring former congressional staffers, building compliance teams, and crafting messages about how AI can "help democracy work better." But the real question is: who audits the output? Who checks for bias? Who ensures that the AI is not subtly favoring one industry over another, one political party over another, one set of values over another?
We don't have answers to those questions yet. And the people who should be asking them — members of Congress — are the same people who are using the tools. It's a classic regulatory capture problem, but it's happening before the regulation even exists.
A Personal Observation
I'll end with a confession. I use AI tools every day. I use them to brainstorm, to edit, to rephrase awkward sentences. I am literally writing this article with the help of an AI assistant that suggests better ways to structure arguments. I am not a hypocrite. But I also know that I am ultimately responsible for every word that appears under my byline. If I get a fact wrong, I cannot blame the AI. I cannot say, "My AI assistant made me do it."
Congress needs to apply the same standard. If a representative submits an amendment, they are responsible for its content. They cannot hide behind a staffer's experiment with Claude. They cannot claim it was just spellcheck when the output looks like it was generated from a prompt. The public deserves to know when AI is involved in the legislative process. And the companies that sell these tools need to be held to a higher standard of transparency.
Because here is the uncomfortable truth: the next time a defense funding amendment quietly shifts billions of dollars toward a specific contractor, it might not be because a lobbyist wrote it. It might be because an AI, trained on data that favors that contractor, suggested the language. And nobody will know until it's too late.
Sarah Chen-Morrison is a veteran tech journalist who has covered the intersection of technology and policy for Wired, The Verge, and Ars Technica. She has been tracking the AI industry's influence on government since before it was cool.
Originally reported by www.theverge.com. Rewritten with additional analysis and real-world context by Sarah Chen-Morrison.
