The Sentence Nobody Was Supposed To See
In June, a congressional aide added AI-drafted language to fix an amendment summary in the office of Representative Anna Paulina Luna, Republican of Florida. The published version still carried the words "Claude responded", the chatbot's own reply, left inside the public text.
Luna later acknowledged what happened in plain terms: her staff used AI to correct a draft and did not edit the result before it went out. The mistake became public because nobody read the final version closely enough to catch three words that did not belong in a piece of federal legislation.
Why Legislative Counsel Is Buried
The incident is a visible symptom of a much larger pattern inside the House Office of Legislative Counsel, the nonpartisan lawyers who turn policy ideas into bill text. According to interviews with eight current and former officials reported by Politico, congressional offices and outside advocacy groups increasingly go straight to Claude or ChatGPT to draft the legislative language itself, then hand it to the office for review.
AI can generate that language far faster than any team of lawyers can write it, which means the office's lawyers are now spending significantly more time rewriting flawed drafts than they ever spent writing from scratch, and the backlog compounds every time a new AI-assisted draft arrives with the same category of error already inside it.
What AI Cannot Judge
The reported failures are not typos, they are legal category errors that change how a law functions once enacted. Reviewers described AI drafts that could not distinguish whether a pot of money should legally be structured as a tax credit, a tax deduction, a tax exclusion or a grant, four instruments with different eligibility rules, different agencies enforcing them and different consequences for the people the law is meant to help.
Miscited statutes and incorrect legal definitions inside a bill do not stay contained to the page, they can trigger lawsuits, delay a bill's passage, or get discovered only after the law is already in force, at which point fixing it requires an entirely new piece of legislation.
The Decision Every Team Ignores
The lesson generalises well past Congress. Any organisation that has adopted generative AI for legal, contractual or compliance drafting is running the identical experiment: the tool makes producing a first draft nearly free, and that speed quietly pressures reviewers to spend less time checking it, exactly when checking matters most.
The Luna office incident is the cheap version of this failure, an embarrassing but survivable public correction. In a contract, a regulatory filing or a piece of enacted law, the same skipped review step produces a mistake nobody catches until it is expensive. The decision that matters is not whether to use AI to draft, it is whether the review step gets the same rigor it had before the draft became free.
Read next: Google Scrapped a Near-Finished AI Model | The Prompt Running Your Business Has No Owner
