AI Tools Used by State Legislatures for Bill Drafting

State legislatures are past the debate stage on AI. They're using it right now to assist with drafting, research, and review of legislation moving through their chambers. The same technology getting written into law is the technology writing the law, and that loop is the actual story here, not some future risk to war-game.

Why the volume of legislation amplified the pressure to use AI for drafting it

Start with the numbers. In 2023, states introduced 191 AI-related bills. In 2024, that jumped past 600, with nearly 100 signed into law. By 2025, it hit 1,208, and for the first time, every single state had at least one AI bill on the table. As of March 2026, with sessions still running, the count already sits at 1,561 across 45 states, according to MultiState's tracking.

That's just the AI-specific bills. The broader legislative workload has been climbing in complexity for decades, long before generative AI showed up, so these drafting tools weren't solving a brand-new problem. They were answering pressure that had been building for years.

A Lawfare analysis by Sanders and Schneier makes a point worth sitting with: state legislatures probably need AI drafting help more urgently than Congress does, simply because state legislative staffs are so much smaller. Fewer people, more bills, less time. Add a Congress that's grown less productive, which pushes more actual policymaking down to the states, and drafting pressure concentrates exactly where staff resources are thinnest.

None of this is new in kind, only in degree. For years, under-resourced legislators leaned on model legislation and lobbyist-drafted text to fill the gap between what they needed to write and what their staff could produce. AI tools are stepping into that same role now, and that's the part worth being honest about: the risk was never that a machine writes badly. It's that whoever controls the prompt controls the interests baked into the language, the same way a lobbyist once did, just faster and with less fingerprint left behind.

One more thing lowered the barrier. The EU AI Act, in force since August 2024, gave American drafters a detailed, already-tested template to adapt instead of a blank page to fill. Nobody writes from scratch what someone else already built and stress-tested.

What legislative staff are actually doing with these tools day to day

Summarization is the on-ramp, and it's the honest, defensible use case. Most staff aren't asking AI to write a bill out of nothing. They're asking it to take a 300-page draft and turn it into three sentences a colleague can read before a vote.

Vermont legislator Priestley's documented workflow is a useful window into this. Priestley uses large language models to turn long, dense bills into plain problem-and-solution summaries for colleagues short on time. Draft language gets uploaded for review, with the model flagging specific provisions and suggesting revisions. While developing the State Information Practices Act, Priestley used LLM research support to study how other states structure protections around state-held data, the kind of cross-jurisdictional research that traditionally places heavy demands on limited staff time.

Other cases show how uneven this gets in practice. In Alaska, H.B. 86's sponsor tried using Microsoft's Bing AI to draft an amendment on money transmission, and the tool initially refused. The same sponsor, Sumner, later used Bing AI successfully to draft a separate bill on gambling aboard Alaska ferries. Same tool, same person, different outcome depending on subject matter. At the federal level, the sponsor of H.Res.66 reportedly prompted ChatGPT with something close to: "You are Congressman Ted Lieu. Write a comprehensive congressional resolution generally expressing support for Congress to focus on AI." Ohio stands out for longevity: the state has used an AI tool for wholesale revision of its administrative law since 2020, possibly the longest-running documented case of this at the state level. The U.S. House Office of the Clerk has also been reported to be exploring AI for internal legislative processing tasks.

Brainstorming and early drafting round out the common use cases that staff report experimenting with alongside the more established ones. Full bill text generated straight from a prompt still happens. It's the rare case, not the default, and that gap between what's technically possible and what staff actually trust the tool to do is worth noticing on its own.

The commercial platforms built specifically for legislative drafting

General-purpose chatbots are the entry point. A market of purpose-built tools is forming around this exact need, and the split between the two says something about where the technology is actually trusted.

FiscalNote Holdings (NYSE: NOTE) announced AI-powered legislative drafting inside its PolicyNote platform on July 10, 2025, aimed at government affairs teams, advocacy groups, and legislative staff. It can generate full bill text, model legislation, and amendments tailored to a specific sector or political context. Its Bill Comparison feature analyzes differences between legislative and regulatory texts in seconds, producing a redline of additions and deletions, useful for stacking bill versions against each other. FiscalNote says it counts clients across all three branches of the federal government.

Vulcan Technologies takes a sharper angle. Founded in 2025 and backed by Y Combinator, it calls its product a "regulatory operating system": an agentic platform pulling together laws, regulations, and court decisions across federal, state, and municipal jurisdictions, then analyzing statutory language and drafting proposed text with citations attached. Virginia Governor Glenn Youngkin mandated its use across all state agencies with a specific target: a substantial cut in existing regulation. That's not a pilot program. That's a governor telling every agency in the state to run its rulebook through an AI tool and come back with a number.

But for all the purpose-built platforms entering the space, general tools still dominate daily use. ChatGPT and Microsoft Copilot remain the most common, per NCSL's survey work, with Lexis+ AI showing up for legal research and Grammarly for basic proofing. Claude and AWS Bedrock get mentioned too. Westlaw Precision, Consensus, Workiva's Wdesk, and Gemini are reportedly under active consideration in offices that haven't adopted them yet. And this isn't confined to legislatures: the Trump administration has reportedly discussed using Google Gemini to help draft federal transportation regulations, with the Department of Education also experimenting with AI-assisted rulemaking.

What AI does well in drafting and where it reliably falls short

Give credit where it's due, because the strength here is real and specific. Lawfare's analysis points to something worth taking seriously: these models have what amounts to superhuman attention spans for enforcing syntactic and grammatical rules. A human drafter gets tired proofreading cross-references on page 40 of a bill. A model doesn't get tired at all. That consistency, paired with the ability to research how other states handled a similar provision far more quickly than a staffer working manually could, is a genuine and measurable strength.

Summarization and research are the reliable use cases, then. Generating bill text that's both legally sound and contextually right from a cold start is a much harder problem, and it's the one where trust should be lowest, not highest, because that's exactly where the failures are hardest to catch.

The concern, raised in that same Lawfare piece and echoed elsewhere, is that legislatures may be deploying this technology without fully understanding its limits. Training data carries bias. Models hallucinate citations that look completely legitimate until someone checks them by hand. Uploading sensitive draft legislation into a commercial model raises security questions nobody's fully answered.

Governance hasn't kept pace with adoption, and the gap is the real problem, not the tool itself. NCSL's Will Clark pointed out that the number of legislative offices with a formal AI use policy has barely moved: 16 respondents reported having one in the 2024 survey, and 16 again in 2025, even as the share of staff actually using generative AI nearly doubled, from 20% to 44%. Clark also flagged, at the August 2025 national legislative conference in Boston, that real usage is almost certainly higher than either number suggests. Survey respondents are just the ones willing to admit it.

No rule currently requires a human to draft legislative text. No regulation limits how much of a bill an AI tool can produce. The lobbying parallel resurfaces here: model legislation written by industry groups has long been a way to embed particular interests into statutory language before real democratic debate even starts. AI drafting tools risk doing the same thing, quietly, at a scale and speed no lobbyist could ever match on their own.

What it means that laws are now being written by the same kind of tool they are meant to regulate

State the loop plainly: legislatures are using AI to write AI regulation faster than anyone can fully understand either the tools doing the writing or the laws coming out the other end.

The volume makes this concrete rather than abstract. 1,208 AI bills introduced in 2025, 145 of them enacted. That's an output rate that strains human drafting capacity on its own, which is precisely the condition that makes AI assistance feel not just convenient but necessary. The pressure and the tool feed each other, and neither one slows the other down.

What's missing is any disclosure requirement, and this is the part that should actually worry people. When a bill reaches a floor vote, there's no reliable way to know what share of its text came from a prompt, what got revised by a human afterward, and what got accepted wholesale. That's not a hypothetical gap. It's the current state of practice, right now, in bills already signed into law.

A further question worth sitting with is whether AI-assisted drafting could shift the balance of power between the legislative and executive branches, if it enables more granular statutory language that limits agency discretion at a scale and speed human drafters rarely achieve. Human drafters, working under deadline pressure, rarely hit that level of granular specificity at speed. AI might make it routine, which would quietly hand legislatures more control over agency behavior than they've historically been able to exercise.

Colorado is worth watching as a live case of the loop closing on itself. The state enacted one of the first comprehensive AI risk laws in the country, and as of August 2026, its Attorney General is releasing proposed implementing regulations for the revised law (SB 189) and the Chatbot Safety Act (HB 1263), covering transparency requirements for automated decision-making in employment, housing, and healthcare, effective January 1, 2027. A state actively using AI tools to help draft the very accountability framework that AI systems will have to meet: that's just where the road leads when the tool and the subject of regulation are the same thing.

The absence of a disclosure rule isn't neutral. It's a policy choice, made by inaction, and legislatures are making it right now, bill by bill, without ever having to vote on it directly.

How citizen-facing civic technology fits into the same shift AI is driving inside legislatures

So who else gets to use this capability? Once you see how deeply AI has embedded itself on the drafting side, that question stops being rhetorical.

Civic technology is broadly understood as digital tools that reinforce democracy: helping the public stay informed, participate in policymaking, and hold government accountable. Look at what that framework describes, and the overlap with what's happening inside legislatures is hard to miss. The same capability that lets a legislator compress a dense bill into a plain-language summary, or lets an agency work through regulatory revision, can help an ordinary resident turn a plain-language concern into a properly structured legislative proposal. Platforms like pollsee, a civic tech startup that uses AI to draft actual bills from citizen ideas, are built on exactly that premise. The technical barrier is identical on both sides of that transaction, which makes the current gap a choice, not a limitation.

Access right now is lopsided, and that's the part worth naming directly instead of softening. Legislatures have drafting tools, purpose-built platforms, and increasingly, in-house AI policy. Citizens navigating that same legislative process generally have none of it. Closing that gap is exactly the opportunity the OECD framework points to: civic technology that hands ordinary people the same drafting leverage that state agencies and lobbying groups already have.

What makes a citizen-drafted proposal credible once it lands on an official's desk isn't the AI polish on the language. It's verification: real people, attached to real names, co-signing a specific proposal. That structure is what separates something an office has to take seriously from a petition that gets filed and forgotten.

A caution worth repeating rather than glossing over: AI inside civic systems can just as easily increase opacity as reduce it, and complicate accountability rather than strengthen it. Any tool built for citizens needs to stay transparent about what the AI is actually doing, and keep human judgment, and human signatures, at the center of the process instead of automating them away.

Which leaves a fairly direct question. If lobbying groups and well-funded agencies can use AI to shape legislation before the public even sees it, shouldn't the public get the same kind of leverage, pointed back the other way?