AI-Generated Bills Flood House Legislative Counsel Office, Creating More Work Than They Save
The US House Office of Legislative Counsel is being inundated with AI-generated legislative drafts riddled with incorrect terminology, fabricated citations, and structural errors, according to a report published by Politico on August 17, 2026. Lawyers at the OLC now spend more time reviewing and rewriting these drafts than they would have spent writing the bills from scratch, sources told reporter Owen Dahlkamp. (Politico)
More congressional offices are turning to publicly available general-purpose AI tools, specifically ChatGPT and Claude, to produce legislative proposals, Politico reported. The result is a volume of defective draft legislation arriving at the OLC, the nonpartisan office responsible for translating policy ideas into legally sound statutory language. (Engadget)
The errors are not cosmetic. Wade Ballou, who headed the OLC for nearly a decade until 2024, told Politico that AI cannot determine whether a sum of money should be structured as a tax credit, a tax deduction, a tax exclusion, or a grant. These are not interchangeable labels; each carries distinct budgetary and legal consequences under federal law. A model that picks the wrong mechanism has not drafted a bill with a minor flaw. It has drafted a different bill.
One recurring error reported by Politico is that AI drafting tools sometimes classify "state" to mean only the 50 US states, which would exclude the District of Columbia and tribal nations from federal programs. A lawyer who worked with the OLC told Politico that AI also incorrectly cites previous statutes in draft bills, inserting references to laws that do not exist or do not say what the draft claims.
Beyond the technical defects, Politico reported a subtler concern: congressional staffers who use AI to draft their bills are not as deeply familiar with what their bills are about and hope to accomplish. The OLC's work has always depended on a collaborative process in which staffers articulate policy goals and legislative counsel translates those goals into precise statutory text. When the staffer's starting point is a machine-generated draft, the comprehension gap between the proposal and its author widens, and the OLC's review work shifts from collaboration to forensic correction.
The OLC is not rejecting AI outright. A working group within the office developed an internal tool called the Comparative Print Suite, which helps staffers visualize how a proposed change would alter current law. The tool is designed with a guardrail that reflects hard-won caution: when it cannot determine where changes in a proposal should be made, it returns an error rather than guessing. That design choice prevents hallucinated text from entering draft bills, a failure mode the office has apparently seen enough of from consumer-grade tools.
The contrast between the two approaches is stark. General-purpose LLMs, trained on broad internet corpora, are optimized to produce fluent text regardless of whether that text is legally accurate. The Comparative Print Suite, built within the legislative context it serves, constrains its output to what it can verify and refuses to generate when verification fails. One approach prioritizes completion; the other prioritizes correctness.
This is a pattern technology professionals will recognize from other domains. The same dynamics that make LLMs useful for drafting marketing copy or summarizing documents, their fluency and speed, become liabilities when the output must conform to a precise, high-stakes specification with no tolerance for plausible-sounding error. Legislative drafting is closer to code generation than to copywriting: a confidently wrong citation is not a stylistic problem, it is a functional defect. The difference is that a faulty function call fails fast. A faulty statutory citation can propagate through the legislative process until someone catches it, or does not.
What the OLC's experience surfaces is not that AI is unsuited to legislative work. The Comparative Print Suite demonstrates that an appropriately scoped tool, embedded in its domain's constraints, can add value. The problem is the gap between what general-purpose chatbots can produce and what statutory drafting requires, and the assumption by some congressional offices that the former is a substitute for the latter. The OLC's lawyers are now absorbing the cost of that gap, one hallucinated citation at a time.
