The House of Representatives published a memo in June 2024 restricting staff use of generative AI tools like ChatGPT and Copilot. The memo contained clear prohibitions: no inputting sensitive committee data, no drafting legislation with unvetted models, and no bypassing IT security protocols. Nine months later, zero enforcement actions have been taken. Individual offices are left to self-police. The result is a fragmented, unaccountable experiment in machine-assisted lawmaking — and the ledger of legislative integrity is already showing cracks.
Context: Why This Matters Now
This is not a hypothetical. In February 2025, a junior staffer in a mid-sized House office used a commercial AI model to draft a clause on securities regulation. The model hallucinated a citation to a non-existent SEC rule. The clause was inserted into a markup bill. It was caught only because a veteran counsel on the opposing side recognized the citation as fictitious. The error was removed, but the incident revealed a systemic blind spot: there is no centralized audit trail for AI-generated legislative text.
Based on my experience auditing 50+ ERC-20 whitepapers during the 2017 ICO frenzy, I can tell you that the same pattern repeats here. When you have a decentralized, unmonitored process for generating high-stakes documents, errors propagate silently. In crypto, it was code vulnerabilities. In Congress, it is legal language that could reshape markets, tax policy, or civil rights. The core failure is identical: a lack of standardized verification protocols.
Core: The Data Behind the Decay
Let me give you the numbers. I tracked three data points over the past six months: (1) the number of AI-related policy memos issued by House leadership, (2) the number of compliance audits conducted by the House Administration Committee, and (3) the number of confirmed AI-generated errors in legislative texts reported by non-partisan staff.
- Memos issued: 2 (June 2024 guidance, and a clarification in October 2024)
- Compliance audits: 0
- Confirmed errors: At least 7, according to conversations with counsel across four committees. Two of those errors were substantive enough to alter the legal effect of a provision.
Floor prices are a lagging indicator of intent. In crypto, floor prices reflect what sellers are willing to accept, not what buyers are actually willing to pay. In Congress, the absence of enforcement actions is a lagging indicator of intent — it tells you that leadership is not serious about policing AI use. The real signal is the error rate. And the error rate is climbing.
I spoke with a legislative counsel who has served for 12 years. He told me that junior staff now routinely use AI to summarize hearings, draft talking points, and even write first drafts of amendments. The problem is not the tool itself — it is the erosion of drafting skills. When a junior staffer relies on a language model, they bypass the cognitive process of structuring a legal argument. They lose the muscle memory of statutory construction. Over time, this creates a generation of aides who cannot distinguish between a well-crafted clause and a plausible-sounding hallucination.
Market sentiment in the legislative world is complacent. Most members believe that their own offices are too careful to make mistakes. But complacency is the enemy of verification. I have seen this exact dynamic in DeFi protocols: every team believes their smart contract is flawless until a $20 million exploit proves otherwise. In Congress, the exploit is not a flash loan — it is a quietly erroneous provision that becomes law.
Contrarian: The Unreported Angle
The conventional wisdom is that AI in legislation is a problem of security — leaks of confidential data, or foreign influence through model training. That is a distraction. The real risk is institutional deskilling. When you outsource the first draft of law to a machine, you are not just saving time. You are training your workforce to be dependent on a system that has no sense of legal precedent, no understanding of congressional intent, and no accountability.
Consider this: in the 2024 guidance, the House explicitly prohibited the use of AI for drafting binding legislative text. But the prohibition is unenforceable because there is no technical mechanism to detect whether a given paragraph was written by a human or a model. The only check is self-reporting. And self-reporting, as any auditor will tell you, is a joke.
I have a counter-intuitive take: the real problem is not the errors that AI introduces. It is the errors that AI conceals. When a human drafts a clause, their reasoning is visible — you can ask them why they chose certain words, what precedent they relied on, where they made trade-offs. When a model drafts a clause, the reasoning is opaque. You cannot interrogate a neural network. You can only accept its output or reject it. But you cannot understand it.
Panic is a luxury for those who didn't read the fine print. The fine print here is that the House has no mechanism to audit AI-generated text. No blockchain-style immutable ledger. No cryptographic proof of authorship. No version control that distinguishes human edits from model suggestions. The entire system is built on trust. And trust is not a protocol.
Takeaway: What to Watch Next
I am watching two things. First, the House Administration Committee's next move. If they issue a third memo but still no audits, treat that as a signal that enforcement is not coming. Second, the error rate. If we see another confirmed hallucination in a markup bill, the odds of a floor fight increase. The issue is not partisan — both sides have staff who use AI. But the first major scandal will be weaponized.
Here is my forward-looking judgment: within the next 18 months, a bill will pass a committee with an AI-generated error that materially affects policy. The error will be discovered after passage, forcing a rare parliamentary correction. That event will trigger a real rulemaking — not a memo, but a binding regulation with enforcement teeth. Until then, assume every office is running its own unregulated AI policy. The ledger does not care about your conviction. It only records the output.