The House Office of Legislative Counsel is spending more time fixing AI-written bills than it would take to write them from scratch. Drafts produced by chatbots arrive with wrong statutory citations, flawed definitions and legal structures that would not survive contact with existing law, according to reporting by Government Executive.
The office received 5,623 legislative requests in the first 60 days of the current Congress, up 72% from the same period two years earlier. Generative tools did not create that workload, but they multiply it: producing text is now nearly free, while checking it is not.
What goes wrong in AI-drafted legislation
The problems are not stylistic. AI tools miss the fine distinctions that determine how a law gets interpreted and applied. Wade Ballou, who ran the office for almost a decade until 2024, gave a concrete example: AI cannot work out whether a pot of money should be structured as a tax credit, a tax deduction, a tax exclusion or a grant. Those are four different legal instruments with four different consequences.
Definitions cause similar damage. A draft that defines "state" as the 50 states quietly writes Washington DC and tribal nations out of a federal programme. A lawyer who worked with the office described AI systems citing previous statutes incorrectly in drafts, an error that can tie up passage and invite litigation.
More representatives' offices and outside groups are using publicly available tools such as ChatGPT and Claude to draft legislative proposals. The office employs 61 attorneys and 19 support staff, and the cleanup work now absorbs a larger share of their time than original drafting does.
A second, quieter problem
The drafting failures are visible. The knowledge failure is not. Congressional staffers reportedly no longer know their own bills as well as they used to. Writing legislation forced people to learn an issue in depth. Prompting for legislation does not, and the understanding that used to come with the work has thinned out.
Disclosure makes the scale hard to measure. One staffer described AI use as something people tend not to discuss, still carrying a taboo. No bill voted on by either chamber has turned out to be entirely AI-crafted, so the evidence concerns what arrives at the office rather than what leaves Congress.
Lawmakers have not agreed on whether any of this should be happening. One representative argued that relying on the technology could introduce errors, create unintended legal consequences, or produce language that does not match a member's actual policy intent. A commentator quoted in coverage of the story put the governance gap more bluntly, describing a world where AI writes laws without anyone having agreed that it should.
The office fights AI with AI
The response to a workload created by automation is, awkwardly, more automation. A working group inside the office developed the Comparative Print Suite, which uses natural language processing to show exactly how a proposal would change the US Code. Its design differs from a general chatbot in one respect that matters: when the tool cannot determine with confidence where a change belongs, it returns an error rather than producing something plausible. That keeps invented content out of draft bills.
An adviser who helped design the tool pointed at the contradiction running through the whole situation. Members and interest groups get to write with AI, while legislative counsel themselves cannot use commercially available AI. The arrangement, she said, "is completely incongruent."
The pattern extends beyond Congress. Courts have run into the same problem with AI-generated filings. The difference in legislation is that an uncaught error becomes law rather than a sanctioned attorney. The same drafting assistants now reaching Capitol Hill are being sold into offices everywhere, with vendors racing to ship agents that produce finished documents and presentations.
For legal professionals, the lesson is in the contrast between the two kinds of tool. A system that flags what it cannot resolve gives reviewers something to inspect. A system that produces polished language without exposing uncertainty creates a false sense of completion. The Comparative Print Suite works because it admits ignorance - a design principle worth applying to any AI tool that touches legal work. For lawyers weighing similar tools in their own practice, the distinction between generation and verification is the one that determines whether automation helps or creates liability. Resources like AI for Legal and AI for Government can help professionals assess these tools before they become part of a workflow.
Why this matters for legal professionals
If congressional drafters - among the most careful legal writers in the country - cannot trust AI output without extensive review, the same caution applies to contract drafting, brief writing and regulatory analysis. The burden of verification falls on the lawyer, not the tool. Any AI system that produces confident, polished text without surfacing its uncertainty should be treated as a draft from a junior associate who cannot be questioned. The tools arrived before the rules did, which is the normal order of events and rarely the useful one.
Your membership also unlocks:
