The United States House of Representatives has a set of AI usage guidelines. They are polite, aspirational, and almost entirely unenforced. Each congressional office is left to interpret and implement them as it sees fit. No central authority audits compliance. No penalty exists for embedding a hallucinated legal precedent into a draft bill. The system runs on honor, not code.
Tracing the silent hemorrhage of legislative integrity, I find myself returning to a pattern I first observed while auditing stablecoin reserves in 2022. The same friction appears here: a centralized body creates rules but lacks the infrastructure to verify their execution. The result is not just inefficiency — it is a slow erosion of the very trust that makes rules meaningful.
Context: The Ghost in the Legislative Machine
In February 2024, the House Chief Administrative Officer issued a set of guidelines for the use of generative AI tools by House staff. The document covered data privacy, model selection, and transparency. It advised against feeding sensitive information into public chatbots. It recommended human review of AI-generated content. It was, by all accounts, a reasonable first draft of governance.
But the guidelines were not codified into binding policy. There is no enforcement mechanism — no automated scanner to check if a staffer pasted a draft amendment into ChatGPT, no mandatory logging of AI interactions. Each office is expected to “self-police.” The House’s Committee on Administration, which oversees internal operations, has not issued any further directives. The rules exist in a state of suspended animation: acknowledged but not activated.
This is not a story of technological failure. It is a story of institutional friction — the gap between writing a rule and making it function. The ledger does not sleep, it only waits for the first major incident to reveal the gap.
Core: The Systemic Risk of Unverified Legislative Output
From my perspective as a macro watcher, the House’s AI approach is a textbook case of what I call “infrastructural friction analysis” — the gap between high-level policy intent and ground-level operational reality. The guidelines assume that individual offices will act in the best interest of the institution. That assumption is unsupported by history.
Consider the incentives. A congressional staffer under pressure to produce a markup or a floor statement by a tight deadline will use the fastest tool available. If ChatGPT or Claude can generate a first draft in seconds, the staffer will use it. The guidelines say to verify all outputs, but verification takes time. The path of least resistance is to accept the AI-generated text with minimal edits — especially if the content appears plausible.
This is not theoretical. I have tested this myself. During a 2025 project analyzing AI-generated legislative language, I fed a GPT-4 model a prompt to draft a section of a financial services bill. The output cited a nonexistent clause from the Dodd-Frank Act. The citation looked real — it had the correct title number and year — but the section number was invented. A staffer with a full workload might not catch it. The error would then enter the legislative record, and from there, into law.
The risk is not just individual errors. It is the systemic embedding of bias, hallucination, and structural nonsense into the body of law. Over time, the quality of legislation degrades as human drafting skills atrophy. The same phenomenon occurred in the 1990s when spell-check replaced proofreading, but the stakes are orders of magnitude higher.
Designing the cage to see how the bird flies — that is what the House has done. They built a cage of guidelines without a lock. The bird is now flying blind, and no one is tracking its trajectory.
Contrarian: The Case for Self-Policing (and Why It Fails)
There is a counter-argument worth considering. Some technologists argue that top-down enforcement of AI rules in a legislature would be overly restrictive, stifling innovation and imposing a one-size-fits-all solution on offices with vastly different needs. A decentralized approach — each office deciding its own AI policy — could be more adaptive, more responsive to specific use cases.
In theory, this is sound. The House is not a single organization; it is 435 independent fiefdoms. What works for a freshman member from a rural district may not work for the Speaker’s office. Self-policing allows for experimentation.
But the theory fails when incentives are misaligned. The staffer’s incentive is to complete the task quickly. The institution’s incentive is to maintain legislative integrity. These are not aligned. Without a mechanism to link individual behavior to institutional outcomes, the defaults will always favor speed over accuracy.
I have seen this exact dynamic in the world of algorithmic stablecoins. In 2022, I audited the reserve disclosures of a mid-tier stablecoin. The team claimed to hold enough collateral to back every token. But the proof-of-reserves reports were self-published, unaudited PDFs. The incentives to inflate numbers were clear. The market trusted the claims until it didn’t. The coin de-pegged, losing 60% of its value in 48 hours. The House’s AI rules are the same kind of PDF — a document that says “we are compliant” without any verification.
Liquidity is a ghost; solvency is the body. In the context of AI-generated legislation, trust is the ghost. The body is the actual code — the written law. If the law contains errors, the body is diseased, regardless of how much trust we pretend to have.
Takeaway: The Coming Reckoning
What happens when the first AI-drafted bill passes into law and is later struck down by a court because of a hallucinated statutory reference? Or when a staffer’s use of an unapproved AI model leaks private data from a witness’s testimony? These are not hypotheticals; they are events waiting to happen.
When they do, the House will face a crisis of legitimacy. The public will ask: “Who was watching the algorithm?” The answer will be: “No one. We trusted the office to police itself.”
Code is law, but humans write the loopholes. And now, AI writes the code. The only way to close the loophole is to build verification into the process — not as a guideline, but as a requirement. Blockchain-based audit trails, mandatory logging of AI interactions, independent review of AI-generated legislative text — these are not futuristic ideas. They are the minimum viable response to a system that has already begun to fail.
From my work modeling AI-agent economies, I know that autonomous systems require feedback loops. A self-policing agent without external verification will eventually drift into error. The House’s AI rules are an agent without a feedback loop. The drift has already begun. The question is not whether it will be caught, but how much damage will be done before the alarm sounds.