AI is drafting laws badly. As of Aug. 18, 2026, the technology is creating extra work inside the House Office of Legislative Counsel, where staff must review and rewrite bills filled with wrong terms, incorrect citations, and other mistakes.
The problem is not limited to awkward wording. AI tools miss small details and legal nuances that can change how a law is interpreted and enacted, leaving congressional staffers less familiar with a bill’s content and objectives after using AI to draft it.
Wade Ballou, who led the office until 2024 after almost 10 years in the role, described a basic failure that sounds simple until it reaches statutory text: “AI can’t determine whether a pot of money should be a ‘tax credit, tax deduction, tax exclusion or a grant,’ for instance.” Those choices are not interchangeable, and pretending they are is how legislation acquires expensive footnotes.
AI can also misclassify who qualifies for a federal program. In some drafts, its classification of “state” includes only the 50 states, leaving out DC and tribal nations even when a federal program should include them. The tools may also cite previous statutes incorrectly, forcing legal staff to trace the mistake before a proposal moves forward.
The office wants useful automation, not blind trust
The House Office of Legislative Counsel is exploring AI to improve efficiency, but its approach puts verification at the center. The office is developing a tool called “Comparative Print Suite,” which helps staff visualize how a proposal would change current laws and returns an error when it cannot determine where those changes should be made.
That distinction matters. A system that shows the changes it understands—and flags the ones it cannot map—gives staff a way to inspect its work. A system that produces polished language without exposing uncertainty gives everyone a false sense of progress, which is a familiar feature of automated systems dressed up as assistance.
The same problem is appearing in court filings, where AI-generated text can move from error into manipulation. A court litigant injected prompts into filings in an attempt to influence case outcomes after suspecting the court was using AI, while some filings have included hidden AI prompts designed to deceive third parties that monitor court documents and process them with AI.
Hidden prompts turn bad drafting into misconduct
Formatted court documents can hide information from ordinary readers while leaving it available to software that processes the file. Some court systems accept those formatted submissions, though plain-text filings could block this particular method of concealment. Courts have also faced a suggestion that filings should include AI notification prompts to flag when AI played a role.
The legal risk does not stop at hidden text. Pro se litigants can build arguments backward, asking chatbots to support only their position instead of testing it against opposing arguments. Litigants and lawyers can then become trapped inside falsehoods or confabulated information because the system keeps producing arguments that reinforce the original claim.
One case involved an unnamed litigant who continued adding hidden text to filings despite warnings and faced modest sanctions. Courts are generally gentle with pro se litigants who breach procedure, but deliberate attempts to game the system are treated as serious offenses. Injecting hidden prompts into a filing is not a formatting quirk; it is an attempt to manipulate the court’s process.
The dividing line is clear. AI can help compare legal text, expose changes, and identify an error when it cannot locate the right place to work. It cannot replace judgment about what a law means, who it covers, or whether an argument survives contact with the other side.
That leaves public institutions with an unglamorous requirement: every AI-assisted document needs human review by someone who understands its purpose. The Capitol’s bill-writing problems and the courts’ prompt-injection problems point to the same lesson—polished output is not proof of legal accuracy, and hidden instructions are not cleverness. They are evidence that the system needs stronger controls.
Based on
