Artificial Intelligence has arrived on Capitol Hill — and it’s changing the way American laws are written.
Just not always in the way you’d expect.
Congressional staffers and outside groups are turning to tools like Claude and ChatGPT to draft legislative text, even without any serious debate about the risks that come with it.
And those risks are very real. 😬
As Daniel Schuman, executive director of the nonprofit American Governance Institute, which helps modernize government technology, put it: AI is great for a lot of things, but it’s not capable of drafting laws you’d actually want to see passed.
The problem is that AI makes mistakes that fly under the radar of anyone who doesn’t know the technical details of American legislation.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Incorrect citations to previous statutes, mixing up tax credits and deductions, accidentally excluding territories like Washington D.C. and tribal nations — those are just some of the slip-ups showing up in algorithm-generated drafts.
The ones left cleaning up the mess are the folks at the Office of Legislative Counsel, or simply OLC.
The irony is that the office is spending more time fixing these drafts than it would take to write the laws from scratch.
Meanwhile, the debate over AI use in Congress is still in its infancy — and plenty of people inside the Capitol are already worried about what could happen when these errors make it into actual law. 👇
When AI Meets American Legislation
To understand the scale of the problem, it helps to know that American legislation is among the most complex in the world. Every single word in a bill carries enormous legal weight — a comma in the wrong place, an incorrect reference to an existing statute, or an ambiguous definition can completely change how a rule affects millions of people. It’s work that requires years of specialized training and a deep understanding of existing layers of law, judicial precedents, and the political intentions behind every drafted passage. It’s not exactly the kind of task you’d expect to hand off to a language model without careful supervision.
But that’s exactly what’s happening. Overworked legislative staffers, squeezed by tight deadlines and growing workloads, are finding in artificial intelligence tools a seemingly convenient shortcut. According to people who coach office teams on how to use these technologies, staffers and outside groups are going straight to Claude or ChatGPT to draft the legislative language itself. The speed at which AI generates a complete text is undeniably impressive — in a matter of seconds, it delivers a draft that looks coherent, well-structured, and technically plausible. The problem starts right there: the text looks right. But looks can be deceiving, and in the legal world, that illusion can come at a steep cost.
What makes all of this even more delicate is the political context in which it’s happening. The American Congress is under constant pressure — there’s less time for deliberation, more bills moving simultaneously, and a growing demand for quick responses. In that environment, any tool that promises speed is going to get adopted, sometimes without the care it requires. And when that tool is a generative AI trained to sound confident even when it’s wrong, the risk of a serious slip-up making it into the final text of a law is real and documented. Four congressional staffers involved in drafting legislation confirmed that AI use is becoming increasingly common in the House, with teams and advocacy groups relying on it for day-to-day work, including reviewing and drafting proposals.
The Mistakes AI Is Making
The types of errors found in AI-generated drafts are not trivial. We’re not talking about out-of-place slang or clunky sentences — we’re talking about technical failures that only someone with deep knowledge of the American legal system can spot. One of the most common examples is incorrect statute citations. As Ari Hershowitz, an attorney and technology expert who has worked with the office on integrating tech tools, warned, this type of flaw is a surefire way to introduce thousands of bugs into our legal system. The AI references a law that exists, but with the wrong number, the wrong section, or the wrong year, creating a cross-reference that doesn’t actually work in practice. In a bill, this kind of error can render an entire section legally unenforceable.
Another frequently reported issue involves conceptual mix-ups that seem minor but have massive fiscal impact. Wade Ballou, who headed the Office of Legislative Counsel from 2016 to 2024, highlighted how certain details are critical — for instance, defining whether a particular resource should be a tax credit, a tax deduction, a tax exclusion, or a subsidy appropriation. According to him, AI will miss those nuances. The difference between a tax credit and a tax deduction is fundamental in American tax law — the first directly reduces the tax owed, while the second only lowers the taxable base. When AI uses both terms interchangeably or picks the wrong one, it can completely change the benefit a piece of legislation was designed to offer, helping or hurting taxpayers in ways the lawmakers never intended.
And then there’s the geographic issue, which seems simple but reveals a structural limitation of language models. In some drafts, Washington D.C. and tribal nations ended up excluded from federal programs because AI defined state as referring only to the 50 states, according to a person who works with the OLC. AI doesn’t have a real understanding of the political-administrative structure of the United States — it just reproduces linguistic patterns learned during training. When those patterns clash with legal reality, the resulting text can look correct at first glance and be completely problematic in practice. 😬
The Office of Legislative Counsel Caught in the Crossfire
The Office of Legislative Counsel exists precisely to ensure that bills come out with impeccable technical quality. With dozens of public servants, the OLC operates as one of the most important quality-control filters in all of Congress. It’s a low-profile office, little known outside Capitol Hill circles, but with enormous strategic importance for the functioning of American democracy. As Marci Harris, executive director of the POPVOX Foundation, summed it up, the OLC is the institutional capacity that makes Congress’s primary constitutional power, lawmaking, actually work. When you talk about Congress’s ability to legislate, you’re talking about this office.
What’s happening now is that the OLC is being flooded with AI-generated drafts that arrive with problems baked in. Instead of receiving a human outline with clear guidelines and natural imperfections that can be efficiently refined, the office’s attorneys are getting texts that look ready to go but hide technical errors that require a meticulous review from the ground up. The result is paradoxical: by trying to speed up the process with artificial intelligence, staffers and outside groups are actually placing an even heavier burden on the human experts who are supposed to be the last filter before a bill goes to a vote.
Sources connected to the office reported that, in several cases, the OLC spends more time trying to fix AI-drafted legislation than it would take to write everything from scratch. That’s because reviewing an algorithm-generated text requires the reviewer to read every line with active suspicion, verifying every reference, every definition, and every logical structure — something that wouldn’t be necessary at the same intensity with a human draft, where mistakes tend to be more predictable and localized. It’s a complete inversion of the efficiency promise AI supposedly brought to the table. According to Schuman, from the American Governance Institute, commercially available AI tools simply lack the attention to detail and the sophisticated legal understanding needed to write bills.
The Debate That Still Hasn’t Really Happened
Despite all of this, Congress still hasn’t had a structured, serious conversation about how to regulate the use of artificial intelligence in the legislative process. The sense of concern is palpable among those following the issue closely. As one House staffer vented, the situation is absolutely terrifying, and nobody is thinking about it yet. There are scattered initiatives, informal working groups, and some internal guidelines in certain offices, but nothing that matches the scale of the problem that’s unfolding. While tools like ChatGPT and Claude are being freely used by staffers and outside advocacy groups to draft bill proposals, there is no clear policy on transparency, accountability, or mandatory review by human experts.
This regulatory gap is especially concerning because AI use in the legislative process isn’t limited to Congress’s internal staffers. Lobbying groups, nonprofits, think tanks, and even private companies with stakes in particular legislation are also using these tools to craft proposals and suggestions that end up in lawmakers’ hands. When an AI-generated text, with embedded errors and training biases, starts circulating as the basis for a real bill, the risk of distorting the democratic process becomes concrete and hard to ignore.
Representative Joe Morelle, the top Democrat on the House Administration Committee, made it clear in a statement that AI cannot replace the qualified and dedicated professionals at the Office of Legislative Counsel, who help the House carry out its most critical function. Representative Norma Torres, also a committee member, reinforced that over-relying on AI-generated bill text could introduce errors, create unintended legal consequences, or result in language that doesn’t accurately reflect a lawmaker’s policy intent. She emphasized that the ultimate responsibility for introduced legislation falls on the members of Congress and their teams — and that responsibility cannot be outsourced to a chatbot.
What many experts inside and outside the Capitol are calling for is that any use of AI in the legislative process needs to come with robust layers of specialized human review — not as additional red tape, but as an essential safeguard. Technology can be a powerful ally in organizing information, researching precedents, and even suggesting language, but the final responsibility for a law’s text needs to be in the hands of people who understand its real-world consequences. Because at the end of the day, real people are the ones who will live under the weight of these laws — and an algorithm-generated error can’t be held accountable for that. 👇
