When AI Writes the Law: How AI Slop and Legal Errors Threaten U.S. and EU Lawmaking

That gap is becoming the more important story.

The bottleneck is no longer writing

The House has long had more proposed legislation than it could ultimately process. During the previous Congress, the Office of Legislative Counsel prepared more than 30,000 drafts, almost three times the number that were actually introduced.

AI adds a new dimension to that imbalance. Congressional offices, advocacy organisations and outside interest groups can now produce legislative language at a speed and volume that would once have required substantial staff time.

The result is not necessarily better legislation. It is more legislation that looks finished.

Commercial systems such as OpenAI's ChatGPT and Anthropic's Claude can generate convincing statutory language while making mistakes that would be unacceptable in a legal document. They can invent or misidentify sections of the US Code. They can confuse tax credits with deductions, exclusions or grants. They can define "state" in a way that inadvertently leaves out Washington, D.C., or tribal nations.

These are not cosmetic defects.

A statute can be grammatically perfect and legally wrong.

That distinction is central to the problem confronting congressional lawyers. Correcting an AI-generated bill can take longer than drafting one from scratch because the lawyer must first identify which parts of the text can be trusted. The apparent efficiency of automated drafting therefore shifts work rather than eliminating it, placing the burden on the small group of professionals responsible for validating the result.

Former House Legislative Counsel Wade Ballou, who served from 2016 to 2024, described the danger of applying error-prone generative AI to foundational statutory language as "the highest risk at the apex of failure."

The phrase captures the institutional problem: the closer AI gets to the final legal text, the less room there is for plausible mistakes.

The hidden cost is expertise

Another consequence cannot be measured simply by counting defective drafts.

Traditional legislative drafting forces congressional staff to understand the mechanics of the policy they are trying to enact. Working through definitions, cross-references and interactions with existing law is part of understanding the legislation itself.

Delegating that work to an AI system can break the connection between policy knowledge and statutory language.

A staffer may know what a proposal is intended to achieve without fully understanding how the wording needed to achieve it interacts with existing statutes. If the machine produces the language, the human role can gradually become one of selecting, editing, and approving text rather than mastering its legal architecture.

That creates a particularly awkward dependency. The Office of Legislative Counsel is then expected to act as both a drafting service and a final human firewall against technology used elsewhere in the legislative process.

Rep. Norma Torres has warned that excessive reliance on AI-generated bill language could introduce errors, create unintended legal consequences or produce wording that does not accurately reflect a member's policy intent. Rep. Joe Morelle has similarly argued that AI cannot replace the professionals in the legislative counsel's office and has backed providing the institution with the resources it needs.

The issue is therefore not whether congressional lawyers will continue to be necessary. It is whether they can absorb the consequences of a technology that dramatically increases the amount of text they are expected to police.

A different kind of legal technology

The House response points towards an important distinction between using AI to generate language and using technology to verify it.

The Comparative Print Suite, developed for the House by technologist Ari Hershowitz and strategic designer Charlotte Lee of Granicus, is designed to visualise changes to the US Code using natural language processing. Its crucial feature is not that it produces more text. It is that it can signal uncertainty rather than invent an answer.

That design philosophy matters.

Commercial large language models are built to generate plausible continuations. Legislative systems require something closer to disciplined refusal: when a citation cannot be established or a legal relationship is uncertain, the system must say so.

The difference is fundamental. In ordinary writing, a confident approximation may be harmless. In statutory drafting, an invented section number can alter the meaning of an entire provision.

The emerging lesson for government is therefore less about whether public institutions should use AI than about which tasks are suitable for probabilistic systems. The most valuable technology may not be the system capable of producing a bill in seconds, but the one capable of telling a lawyer precisely where the machine - or the human - may have gone wrong.

Europe is confronting the same problem differently

The United States is not alone in discovering that AI can increase the supply of legislative text faster than institutions can verify it.

Officials in the European Parliament have also raised concerns about "AI slop", hallucinated legal references and growing use of AI-assisted amendments. But the institutional response is developing along a different path.

The EU's AI Act has been supplemented by the July 2026 Digital Omnibus on AI, while Article 50 transparency requirements took effect on August 2. In the European Parliament, however, disclosure of AI use by members remains voluntary. The institution is also developing its EPGenAI Hub to provide secure access to approved models.

The contrast is revealing. Europe is building a framework around controlled use and transparency while the US federal legislative system remains without a comprehensive framework governing AI deployment in lawmaking. Proposed federal legislation, including the Great American AI Act, remains stalled amid disputes over federal preemption.

Neither approach resolves the underlying capacity problem.

Rules can require disclosure. Secure platforms can control access. Neither automatically supplies the lawyers needed to verify millions of lines of machine-generated legal language.

Who gains from the new abundance?

The shift also changes the balance between lawmakers, civil servants and the organisations trying to influence legislation.

Advocacy groups and special interests can now generate highly customised statutory language rapidly, reducing one of the practical barriers between lobbying and legislative drafting. The ability to arrive with ready-made language has always mattered in Washington; reducing the cost of producing that language makes the advantage easier to scale.

The corresponding pressure falls on nonpartisan congressional staff.

If the volume of proposed language rises while the capacity for legal scrutiny remains largely fixed, the institution's scarce resource becomes verification. The people who possess that expertise become gatekeepers of an increasingly automated legislative pipeline.

That creates a peculiar inversion of AI's promise. The technology is supposed to remove bottlenecks. In this case, it can make the bottleneck more consequential.

The danger is not that Congress will suddenly allow a chatbot to write the law without human involvement. The more plausible institutional failure is subtler: humans remain formally responsible while the quantity of machine-produced material becomes too large for them to examine with the same depth.

If defective provisions survive that process and become law, the consequences move downstream. Federal agencies may have to implement ambiguous provisions; courts may be asked to resolve contradictory language; businesses and citizens may face prolonged uncertainty over what Congress actually intended. The cost of a cheap first draft can then reappear as litigation, administrative delay, and regulatory confusion.

Congress has spent decades building institutions to make statutory language precise. AI has made producing that language cheap.