Legislating With Autocomplete

New reporting indicates congressional offices are leaning on AI chatbots to draft and analyze legislation more heavily than at any previous point. The commentary has divided immediately and predictably into two camps that are both, I think, asking the wrong question.

Camp one is horrified. Camp two points out that legislative text has been assembled from templates, precedent and lobbyist-supplied language for a century, and that a chatbot is a faster version of a process nobody was romantic about. Camp two is more right than camp one and is still not asking the question.

The question is: which model, running where, retaining what?

A staffer pasting a draft bill into a consumer chat interface has transmitted pre-publication legislative text to a third party, on a consumer plan, under terms most people have not read, with retention behavior that varies by product and tier and changes without notice. The policy problem is not that a machine helped write a bill. It is that nobody can currently answer, for any given bill, where the text went.

The second question is provenance, and this one has teeth. These systems produce fluent, confident, plausible text — including citations to statutes that do not exist, in exactly the register of statutes that do. There are documented instances of fabricated case law reaching actual court filings, from lawyers, who are trained to check. A legislative aide at eleven at night, three days before markup, is under materially worse conditions than those lawyers were.

The failure mode is not a bad law. It is a correct-sounding cross-reference to a subsection that was repealed in 2011, surviving into enacted text because it reads exactly like the eleven real cross-references around it.

None of this argues for prohibition, which will not happen and should not. It argues for three unglamorous rules: an approved tool with a proper data agreement, a disclosure that a draft was machine-assisted, and mandatory citation verification before markup. Every one of those is boring. Every one is achievable this year. None of them is in any of the coverage, which is still arguing about whether robots should write laws.

The United States Code is machine-readable and public. Verifying that every citation in a draft resolves to a real, in-force provision is a lookup, not a research project — an afternoon's work for one competent person, and it would catch the entire class of failure that everyone is worried about.

It does not exist as standard tooling in any office I can find reference to. Cross-reference checking remains a human task performed under deadline by the most junior person in the room.

This is the pattern across every field that has adopted these tools quickly: the risk is real, the mitigation is dull and mechanical, and the discussion stays at the level of philosophy because philosophy is more interesting than a lookup table.

The question is not embarrassing and there is no wrong answer to it, which is what makes the uniform non-response interesting.

A norm forms when the first few people answer and nothing bad happens to them. Nobody wants to be the first few.