Congress Is Drafting Bills With AI, and AI Is Bad at the Part That Matters
Lawmakers and lobbyists have started using generative tools to draft legislation. Reporting on the practice found that the tools are currently poor at the technical mechanics of federal law and regulation, which is an awkward finding, because the technical mechanics are most of what a bill is.
A statute is not an argument. It is a set of instructions for amending other statutes. The operative text of a typical bill consists of directions to strike a phrase in a specific subsection of an existing code title and insert different words in its place, plus definitions, effective dates, severability, and appropriations language. Getting it right requires knowing what the current code says, what pending bills would change about it, how courts have construed the terms of art, and which committee has jurisdiction over the resulting text.
Generative models are good at producing text that resembles legislation and comparatively weak at all of the above, because those things depend on precise reference to a body of law that changes constantly and is documented in ways that reward exact lookup rather than fluent recall. The failure mode is a bill that reads perfectly and cites a subsection that does not exist, or amends language that was already repealed.
The interesting question is who catches it. Legislative counsel offices exist for exactly this and they are small, oversubscribed, and not growing. Volume produced with a tool that generates in seconds meets a review capacity that is measured in staff hours. That is the pressure point, and it is the same pressure point appearing everywhere generative tools land on top of a human verification step: the generation scales, the checking does not.
The lobbying half is the part with sharper implications. Draft text supplied by outside interests has always been part of how legislation gets written, and the constraint on it was cost. Producing polished statutory language required lawyers billing by the hour, which limited the volume and concentrated the practice among well-funded interests. Cheap generation removes the cost constraint without removing the accuracy problem, so the plausible near-term result is more outside drafts of lower average quality arriving at offices with no additional capacity to evaluate them.
Nobody involved is doing anything improper. That is what makes it hard to address. The tooling changed the economics of one input into a process that was never designed with a check for volume.