The House Office of Legislative Counsel, the nonpartisan legal office responsible for turning congressional ideas into precise statutory language, is facing an unexpected burden: fixing AI-drafted bills. According to a person who has advised the office on technology, the office spends “more time trying to fix AI-drafted legislation than it would take for them to draft it from scratch.”

Congressional staffers and outside groups have increasingly turned to AI tools like Claude and ChatGPT to draft legislation. But the lawyers who check that work cannot keep up. The strain on the Office of Legislative Counsel has become a growing concern, with officials reporting that the quality-control layer between an idea and a statute is being tested like never before.

The role of the Office of Legislative Counsel

The Office of Legislative Counsel is essentially the quality-control department for Congress. Every bill and amendment that is formally introduced must pass through this office, where attorneys review language for legal precision, consistency with existing law, and proper statutory structure. Marci Harris of the POPVOX Foundation described the office’s essential function to reporters: “When we talk about Congress’s ability to legislate, we are talking about this office.”

The volume of work is already enormous. In the 118th Congress, the office prepared 30,494 bills, which is 188 percent more than the 10,564 bills that were actually introduced. The office also prepared 21,484 amendments, up 46 percent from the 114th Congress. These figures come from testimony delivered in March by Warren Burke, the current legislative counsel.

Burke’s testimony also revealed who is doing this work. In the first session of the 118th Congress, the office had 78 attorneys and 23 support staff. Attorney numbers have increased by 59 percent since the 116th Congress, but the average experience level has fallen by 22 percent over the same period. More than half of the attorneys now have less than a decade of service.

Where AI goes wrong in a bill

The errors that AI introduces into draft legislation are often small but consequential. Wade Ballou, who led the office from 2016 to 2024, explained that AI struggles with basic legal distinctions. For example, AI has difficulty telling whether a pot of money should be a “tax credit, tax deduction, tax exclusion or a grant.” These are materially different legal mechanisms, and choosing the wrong one can alter the entire effect of a provision. AI misses those nuances.

Definitions are another problem area. A bill that defines a state as only the 50 states can inadvertently exclude Washington, DC, and tribal nations from a federal program. A person who works with the office gave this exact example, warning that such subtle exclusions can have massive unintended consequences.

Citations are the third major failure point. Ari Hershowitz, a lawyer and technologist who has worked with the office, called incorrect statutory references “a guaranteed way of introducing thousands of bugs into our legal system.” When an AI draft cites the wrong section of the U.S. Code, the impact can ripple through subsequent interpretation and enforcement.

Daniel Schuman of the American Governance Institute summed up the concern: “AI is good for a lot of stuff. But it’s not capable of drafting legislation that you would want to enact into law.”

This is not the first time official pipelines have buckled under machine-generated submissions. In August, reports emerged that someone filed 55 software vulnerabilities in a matter of days, and 54 of those were invented by AI. The parallel to legislative drafting is striking: AI can generate plausible-looking output, but that output often requires more human effort to vet than creating the output from scratch.

Two different growth figures

Burke’s testimony presents two distinct sets of numbers. One figure, cited in an earlier news report, showed a 72 percent rise in requests during the first 60 days of the current Congress. That figure compared 5,623 requests against 3,271 in the same window of the previous Congress.

Burke’s most recent testimony, covering a longer window, reports a more modest rate of increase. Across the first 13 months of the current Congress, the office received 23,239 requests, compared with 21,979 in the same period for the 118th Congress. That is a rise of 5.7 percent. The discrepancy suggests that the initial AI surge may have been front-loaded, or that some requests are being filtered earlier in the process.

On AI specifically, Burke was direct. “In the near term, we are bracing for AI to expand our workload,” he wrote in his March testimony. He added that text produced by outside entities, whether by humans or computers, often takes significantly longer to review and revise. Burke declined to comment for the recent article on this topic.

The office is testing its own AI

The Office of Legislative Counsel is not entirely avoiding AI. It has organized a group of staff members to evaluate Microsoft Copilot, as authorized by the House. The stated aim is improving efficiencies such as legal research, according to Burke’s testimony.

The office already runs one AI tool of its own: the Comparative Print Suite. This tool uses natural language processing to show exactly how a bill would change the U.S. Code. Hershowitz described it as the first AI tool in the House. Importantly, it returns an error when it cannot place a change with confidence, preventing hallucinations. “There’s no hallucinations,” he said.

Charlotte Lee of Granicus helped design that tool. She described the wider arrangement to reporters. “Members and interest groups get to write things with AI, but then leg counsel themselves can’t use [commercially available] AI,” she said. “There’s a complete incongruency.” This disconnect leaves the people responsible for accuracy in a position where they cannot use the same generative tools that create the problematic drafts.

What has not been established

While the anecdotal evidence of AI-drafted bills is mounting, no bill voted on by either chamber has turned out to be entirely AI-crafted. The evidence so far concerns AI-drafted bills arriving at the office, not statutes leaving Congress. That means the full extent of AI influence on enacted law remains unknown.

Disclosure is patchy by nature. One staffer told reporters that people who use AI tend not to talk about it, and that it still feels taboo. Labeling rules elsewhere have landed unevenly too. Under the EU AI Act, for example, an AI-written article needs no label while a proofread email does. The legal framework for transparency is still catching up to the technology.

One incident is public. Representative Anna Paulina Luna drew criticism in June over an amendment summary generated with Claude, although she said she had not used AI to write the amendment itself. “Not a shocker. Most staff use it,” she said, adding later that she had told her staff to double check. She later deleted one post on the episode and replaced it, according to a Gizmodo report.

Two members of the Committee on House Administration gave statements on the issue. Representative Joe Morelle said, “AI cannot replace the skilled and dedicated professionals in the Office of Legislative Counsel.” Representative Norma Torres said responsibility for introduced legislation rests with members and their staff, and “cannot be outsourced to a chatbot.”

The wider AI push on Capitol Hill

Congress has been circling AI technology from several directions this year. It spends nine out of ten AI dollars with OpenAI. House Democrats have demanded answers from OpenAI and Anthropic on rogue agents. Legislation is also moving. A bill from Ted Lieu and Jerry Moran would create an emergency kill switch for AI systems, a proposal that followed a rogue agent breaching Hugging Face in July.

Wade Ballou sees a narrower role for the technology. AI can help lawmakers process information, he told reporters, but it cannot replace the art of drafting. Hershowitz expects the volume problem to worsen either way. “You had a problem that was already at 10,” he said. “Now you multiply it by 100.”

Source:TNW | Government-policy News