Product3 distinct publishers3 min readUpdated
Politico reports AI-drafted legislation is landing on an office of 78 attorneys who cannot decline it. The drafting cost did not disappear. It moved downstream.
The Product Desk · Product desk

Compiled by The Product DeskSomething wrong?How this is made
Politico reports AI-drafted legislation is landing on an office of 78 attorneys who cannot decline it. The drafting cost did not disappear. It moved downstream.
The lawyers who turn political intent into statutory text are now spending more time repairing AI-drafted bills than it would take them to write those bills from scratch, according to a person who has advised the House Office of Legislative Counsel on the technology and who spoke to Politico [2]. Owen Dahlkamp's report, published Monday and drawn from interviews with eight current and former officials connected to the office, is a clean instance of a pattern operators keep rediscovering: unreviewed model output does not remove work, it relocates it to whoever signs off [1]. The relocation lands somewhere with no slack. "When we talk about Congress's ability to legislate, we are talking about this office," Marci Harris of the POPVOX Foundation told Politico [7]. In the 118th Congress it prepared 30,494 bills, 188 percent more than the 10,564 that were actually introduced, per March testimony from legislative counsel Warren Burke [8]. So roughly 19,930 drafts, about 65 percent of the total, never reached the floor [1]. It also prepared 21,484 amendments, up 46 percent on the 114th Congress [9]. Doing that work were 78 attorneys and 23 support staff in the first session of this Congress [10]. Attorney headcount is up 59 percent since the 116th, but average experience has fallen 22 percent and 58 percent of attorneys now have under a decade of service [11]. Four congressional staffers involved in drafting told Politico that AI use is becoming routine [4]. Staffers and outside groups are using Claude and ChatGPT to write the legislative language itself [6]. "It's absolutely terrifying," one House staffer said, "and no one is thinking about it yet" [5]. The failure modes are the expensive kind: cheap to generate, slow to find. Wade Ballou, who led the office from 2016 to 2024, said AI struggles to determine whether a pot of money should be a tax credit, tax deduction, tax exclusion or a grant [12]. A person who works with the office described drafts defining a state as only the 50 states, which would exclude Washington DC and tribal nations from a federal programme [13]. 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" [14]. Daniel Schuman of the American Governance Institute put it plainly: AI "is not capable of drafting legislation that you would want to enact into law" [15]. The aggregate demand numbers do not yet show a flood, and that is worth stating. The 72 percent rise in requests cited from last year's testimony covers the first 60 days of this Congress, 5,623 against 3,271, a 71.9 percent increase over a narrow window [16][2]. Burke's most recent testimony covers 13 months: 23,239 requests against 21,979, a 5.7 percent rise, or about 1,260 extra requests [17][3]. Gizmodo rendered the same 72 percent as requests up in 2025 versus two years earlier [18]. The pressure described is per-document review cost, not raw intake. Burke wrote in March that the office is "bracing for AI to expand our workload" and that text produced by outside entities, human or computer, often takes significantly longer to review and revise; he declined to comment to Politico [19]. Meanwhile the tooling is inverted. The office has a working group evaluating Microsoft Copilot as authorised by the House [20], and already runs the Comparative Print Suite, which uses natural language processing to show how a bill changes the US Code and returns an error when it cannot place a change with confidence [21]. "Members and interest groups get to write things with AI, but then leg counsel themselves can't use [commercially available] AI," Charlotte Lee of Granicus, who helped design that tool, told Politico [22]. Watch whether Burke's next testimony separates AI-origin drafts from the rest, since intake counts alone will not surface this.
Ranked by verification strength, evidence, and original report placement.
Owen Dahlkamp reported the strain on the House Office of Legislative Counsel for Politico on Monday, drawing on interviews with eight current and former officials who work or have worked with the office.
Four congressional staffers involved in drafting told Politico that AI use is becoming routine in the House.
Congressional staffers and outside groups are using Claude and ChatGPT to draft legislation, including the legislative language itself.
One House staffer told Politico: "It's absolutely terrifying, and no one is thinking about it yet."
Marci Harris of the POPVOX Foundation told Politico: "When we talk about Congress's ability to legislate, we are talking about this office."
Wade Ballou, who led the office from 2016 to 2024, told Politico that AI struggles to tell whether a pot of money should be a "tax credit, tax deduction, tax exclusion or a grant", and that AI will miss those nuances.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Named testimony is solid; the headline comparison is one anonymous quote
The structural facts — 30,494 bills prepared versus 10,564 introduced, 21,484 amendments, 78 attorneys and 23 support staff, experience down 22 percent, requests up 5.7 percent over 13 months, Burke 'bracing for AI to expand our workload' — all trace to on-the-record congressional testimony and are consistent across the cluster. The load-bearing claim that fixing AI drafts costs more than drafting from scratch rests on a single anonymous adviser relayed through one Politico report, with no measured hours anywhere, and all three publishers derive from that one report rather than verifying independently. The 72 percent growth figure is accurate only for a 60-day window and is restated by Gizmodo as a two-year comparison.
Routine but undisclosed use, attested qualitatively
Adoption of AI drafting inside the House is attested by four staffers describing routine use, by the named tools (Claude, ChatGPT), and by one public episode involving a member's amendment summary. On the counter-side, no source quantifies what share of submitted bills is AI-drafted, staffers say use is not disclosed because it feels taboo, and Politico reported no bill voted on by either chamber has proved entirely AI-crafted. Institutional adoption of sanctioned AI is early: a Copilot evaluation group with no reported outcome, alongside one production tool, the Comparative Print Suite.
Language outruns the measured load
Headlines say 'swamped', 'riddled with errors' and 'absolutely terrifying', and the fix-costs-more-than-drafting line is presented as an established finding when it is one anonymous person's characterisation. The strongest quantitative signal in the cluster points the other way: requests rose 5.7 percent over 13 months, while the more dramatic 72 percent figure covers only 60 days and is restated by one publisher as a two-year rise. The direction of the story is credible — Burke himself is bracing for added workload and the failure modes are concrete and specifically described — so the gap is overstatement of magnitude and certainty, not fabrication.
Anonymous insiders, advocacy groups and a tool designer shape the account
The account is built mainly from people with stakes in how the office is resourced and tooled: anonymous officials who cannot speak on the record, advocacy organisations arguing for congressional capacity (POPVOX Foundation, American Governance Institute), and a Granicus designer of the office's own tool who uses the story to argue that legislative counsel should get commercial AI access. The office's leader declined to comment while his testimony — itself a budget-facing document requesting resources — supplies the load-bearing numbers. No source discloses vendor spend, and none of the three publishers disclose commercial relationships with the AI vendors named.
Direction well supported, magnitude weakly held
Confidence is moderate: the institutional facts are on the record and internally consistent, the failure modes are described concretely by a former head of the office and by lawyers who work with it, and two publishers independently relay the same error taxonomy. It is capped by single-origin reporting for the whole cluster, anonymity on the decisive quantitative comparison, an unresolved conflict between the two growth figures, and the absence of any measurement of review effort, AI-drafted bill share, or the Copilot evaluation's outcome.
Follow any of these and your For You feed starts watching them — no settings page required.
product
Incogni ranks 13 AI assistants by privacy risk: bigger is worse, except ChatGPT1 distinct publisher
science
Text watermarks land on 2 December. The detection they imply does not.1 distinct publisher
science
Claude's watermark is a compliance artefact, not a cheating detector1 distinct publisher
security
The nationalization argument is really a vendor-continuity memo1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 18, 2026
1 article · August 18, 2026
1 article · August 18, 2026