Claude & ChatGPT Draft Laws on Capitol Hill: 8 Officials Cite Error-Filled Bills

For AI practitioners, the Capitol Hill episode is a textbook case of the fluency-precision gap: models that can produce pages of plausible prose cannot reliably cross-reference the U.S. Code or define terms consistently. As congressional offices feed Claude and ChatGPT output directly into the legislative drafting pipeline, the House Office of Legislative Counsel is absorbing the hallucination cost in the form of extended review and rewrite cycles. The lesson for model builders is that domain-specific verification — not raw generation — is the bottleneck for regulated, high-stakes text.

On August 17, 2026, Politico reported — and outlets including Townhall and WCBM quickly amplified — that the House Office of Legislative Counsel (OLC) is straining under a flood of artificial-intelligence-drafted legislation. The reporting, drawn from interviews with eight current and former officials who work or have worked with the office, describes a quiet transformation of the American lawmaking process: congressional offices and outside advocacy groups are no longer merely using AI to research or summarize policy, but are feeding prompts into Anthropic's Claude and OpenAI's ChatGPT to produce the legislative language itself. The output arrives in mountains of text that the OLC's lawyers must then review and rewrite, spending significantly more time than before on cleanup rather than on careful drafting.

“

As congressional offices feed Claude and ChatGPT output directly into the legislative drafting pipeline, the House Office of Legislative Counsel is absorbing the hallucination cost in the form of extended review and rewrite cycles.

That inversion of workflow is the story's most important detail. The House Office of Legislative Counsel is the nonpartisan shop that turns policy intentions into precise statutory language — the prose that courts will later parse for meaning. It is a choke point where accuracy is the entire product. A statute must cross-reference the United States Code exactly, define terms consistently across hundreds of pages, and withstand decades of adversarial judicial interpretation. General-purpose language models, by contrast, are optimized for fluency and plausibility, not for the verifiable, structured precision that legislative drafting demands. As a result, the productivity dividend that generative AI promises elsewhere — faster output, fewer human hours — is inverted on Capitol Hill: the front end generates text almost instantly, but the back end must absorb the cost of correcting it.

The consequences, according to the officials cited by Politico, extend well beyond internal inefficiency. The AI-generated drafts are plagued by mistakes, including erroneously cited statutes and incorrect legal definitions. Those are not cosmetic errors. A miscited provision can render a section ambiguous or unenforceable; an incorrect definition can cascade through an entire regulatory regime, altering who is covered, what conduct is prohibited, and how agencies may act. The sources warned that the defects risk tying up the passage of legislation and inviting lawsuits over statutory errors that survive into enacted law. For an institution whose work product governs millions of Americans, the margin for error is effectively zero, yet the volume of machine-generated text is expanding the attack surface for exactly the kind of defects the OLC was created to prevent.

The governance problem is also a problem of misaligned incentives. The benefit of AI drafting — speed and volume — accrues to the individual congressional office or outside group that uses it. The remediation cost falls on the shared, under-resourced OLC. Daniel Schuman, executive director of the American Governance Institute, which works to modernize government technology, put the point bluntly: "AI is good for a lot of stuff. But it's not capable of drafting legislation that you would want to enact into law." Staffers and outside groups are, in his account, "going to Claude or ChatGPT to draft the legislative language itself." Schuman's framing matters because it comes from a technologist otherwise sympathetic to government modernization; the objection is not to AI in principle but to its unverified application to the highest-stakes text in the country.