Politico reports that generative AI has created a new problem for the House Office of Legislative Counsel: lawmakers are inundating these nonpartisan expert drafters with low-quality bill text, requiring the experts to “spend significantly more time reviewing and rewriting the proposals”—perhaps more time than it would take to draft them “from scratch.”
But the current wave of AI slop isn’t a freestanding data point. The broader context is a decades-long erosion of Congress’s ability to channel policy through institutional expertise before it becomes statutory text. In technical policy areas such as taxation, the costs are concentrated and tangible.
Are today’s AI-produced “mountains of text” the end of history, or merely a midpoint in a broader transformation? And can we solve the problem simply by starting all tax-drafting chatbot prompts with, “You are Ward Hussey.”? Links, literature, and more, below the fold.
In Owen Dahlkamp’s coverage at Politico, tax law provides a motivating example:
People familiar with AI tools and the [Office of Legislative Counsel]’s work said the task of writing legislation is filled with crucial details that can trip up artificial intelligent agents—for instance, whether a particular pot of money should be a “tax credit, tax deduction, tax exclusion or a grant,” said Wade Ballou, who headed the legislative counsel’s office from 2016 to 2024. “AI is going to miss those nuances,” he said.
Those nuances, of course, tripped up congressional staffs long before the advent of generative AI. Indeed, one might date the decline in legislative drafting to the weakening of committee chairs in the 1970s, the ideological polarization of the 1990s, or the gridlock-oriented legislative strategies of the 2010s. Whatever the origins, the arc is well-supported: compared to prior decades, today’s congressional drafting process draws less heavily on concerntrated, longitudinal expertise in existing and prospective legislative text.
Again, tax law provides a revealing case study. In 2019’s Constituencies and Control in Statutory Drafting, Shu-Yi Oei and Leigh Osofsky interviewed government tax counsels, documenting an older model in which members of Congress and their staffs made high-level policy choices and a small group of specialists—especially House and Senate Legislative Counsel—controlled the translation of those choices into statutory language. Oei and Osofsky’s interviewees describe the transition from that expertise-driven model towards more diffuse control over legislative text.
The older model, of course, is associated with longtime House Legislative Counsel Ward Hussey, the principal tax drafter during much of the postwar period. The process might start with a “conceptual markup” with members of Congress and Joint Committee on Taxation staff. Next, Hussey would grill the staff on policy choices, ramifications, and alternative implementations. Only then would Hussey and John Buckley translate the policy substance into statutory language. Essentially, policy development preceded text—with issues like those raised by Ballou addressed in advance of drafting.
The purely conceptual markup ended at Ways and Means under Chair Bill Archer in the mid-1990s, when the committee began requiring actual legislative language before markup. Over time, the drafting process grew more amorphous, with earlier involvement of Legislative Counsel, more proposals from outside groups, and the dispersion of drafting authority previously concentrated in one or two dominant drafters. The backdrop to these changes was rising partisanship, and the process reflected that shift. But text could precede policy development, leaving the two to evolve in tandem.
Indeed, this type of co-evolution characterized legislation in a less overtly partisan era. Before Hussey were Middleton Beaman and Stanley Surrey, who played a role in a World War II-era shift from narrative tax legislation to something more “surgical.” As George Yin writes in 2023’s Textualism, the Authoritativeness of Congressional Committee Reports, and Stanley Surrey, a bill could comprise “just a series of snippets,” targeted amendments that only an acolyte of tax law could read in context. But the expert drafters, including staff at the Joint Committee on Taxation, authored committee reports alongside the more-surgical legislative text. Members of Congress could—and did—use these reports to make decisions about statutory text. The point is that, in a functional legislative process, there’s a relationship between statutory text and underlying policy, mediated by expertise.
And it’s the expertise element that Yin highlights in identifying the causes of a multi-decade decline in “structural legislation”—”good government” bills that address fundamental plumbing rather than interventionist policy. In 2018’s Crafting Structural Tax Legislation in a Highly Polarized Congress, Yin finds that, after a boom in committee staff in the 1970s linked to congressional reforms, the number of committee staff declined after the Republican Revolution in 1994—and this decline was particularly sharp for Ways and Means. Moreover, Yin traces a growing proportion of partisan staff participants in the tax legislative process after 1990, as well as the relegation of JCT staff to a role focused more on revenue estimation than “helping directly in the formation of legislation.” Over time, expertise has taken a backseat to partisan interests in the tax legislative process, with implications for the technical sophistication of legislative work product.
From these perspectives, generative AI may represent the endpoint of a long process of deskilling in the legislative process. When a congressional office prompts a large language model to generate statutory text from a policy idea, the output largely skips the expert-mediated dialogue between text and policy that characterized the Hussey and Beaman eras. It’s beep-boop, here’s a bill. If expert staffers spend their time checking AI-produced text for hallucinations and errors, then there’s less capacity for meaningful contributions to advance policy. Generative AI yields text at alarming volume and velocity, while the capacity of human experts remains relatively fixed.
But there’s also a more optimistic possibility. Right now, generative AI represents a workflow problem—that’s effectively the point Politico makes without explicitly saying so. When generative AI occupies the sole step between a policy idea and presumptive statutory text, the result almost certainly leans towards an ocean of AI slop. But if generative AI were just another tool, used iteratively in a dialogic process, then I suspect there’s a chance to recapture some of the expert legislative capacity that Congress has lost. It’s mapping this layered workflow—and convincing lawmakers to use it—that presents the core challenge as generative AI enters public policy. Lawmakers don’t have to turn Claude into Ward Hussey. But they may be able to replicate some functions of Hussey’s drafting room using these new technological tools.
Related TaxProf Blog coverage:
- Artificial Intelligence and the Future of Tax Law (May 10, 2026)
- The New Places of Legal Expertise in an AI World (Feb. 17, 2026)
- Weekly SSRN Tax Article Review & Roundup: Speck Reviews Yin’s Textualism, Tax Legislative History, and Stanley Surrey (Dec. 16, 2022)
- Oei & Osofsky: Constituencies and Control in Statutory Drafting—Interviews with Government Tax Counsels (May 26, 2019)
- Remembering Ward Hussey (Dec. 4, 2009)
- Death of Ward M. Hussey (Nov. 23, 2009)
