AI Rules: Three Proposals, One Missing Question
Archi4 min read
Contents
The Headlines Got the Fight Wrong#
When the EU AI Act passed, the story was about bans and fines. When the White House issued its executive order, the story was about reporting thresholds. When major labs signed voluntary commitments, the story was about good faith. Each framing was accurate and each missed the operational question: once a model is deployed, who checks whether it still behaves the way its paperwork says it does?
That gap matters more than the headline differences between the three approaches. A rule that cannot be verified after deployment is a press release with a legal font. Here is how the three leading models compare on that single test, and where each one quietly fails.
Option one: the EU AI Act#
The Act takes a risk-tiered approach. Prohibited uses are banned outright. High-risk systems face conformity assessment, documentation, and post-market monitoring. General-purpose models face transparency and, for the largest, systemic risk evaluation.
The strength is structure. The weakness is timing. Most obligations for high-risk systems phase in over two to three years, and the standards bodies writing the technical detail are still drafting. A company building a hiring tool today faces real uncertainty about what conformity will require in 2026. The Act also leans on self-assessment for many systems. A provider documents its own risk management and, in many cases, its own testing. That is not nothing, but it is not independent verification either.
What the headlines missed: the Act's most enforceable provisions are the paperwork ones, not the performance ones. Regulators can check whether a technical file exists. Checking whether a model still performs as documented after six months of production drift is a harder and less funded task.
Option two: the US sector-by-sector approach#
The United States has no comprehensive AI statute. Instead, existing regulators apply existing authority. The FTC polices deceptive claims. The EEOC applies hiring discrimination law to automated screening. Financial regulators examine model risk at banks. The result is a patchwork that follows industry lines rather than technology lines.
The advantage is speed and specificity. A bank already knows its examiner. A hiring vendor already faces discrimination law. The disadvantage is coverage. A company selling an AI tool to schools, small businesses, or consumers may fall between agencies, and no single body sees the whole picture. Enforcement also depends on political will and budget, both of which move with elections.
What the headlines missed: the sector approach is strongest exactly where the headlines are quietest. Routine examinations of banks and insurers already test model governance. The weak spots are the sectors without a dedicated regulator and the consumer products that cross jurisdictions.
Option three: voluntary commitments#
Voluntary pledges from major labs have covered pre-deployment testing, red-teaming, watermarking, and reporting of dangerous capabilities. They are fast to write and easy to update. They also depend on the signatory's own judgment about what counts as passing.
The honest comparison is not voluntary versus mandatory. It is who holds the pen when a test fails. Under a voluntary regime, the lab decides whether to ship, delay, or narrow a release, and the public often learns the reasoning later, if at all. That can still produce cautious behavior, especially when reputations are at stake. It cannot produce a record that an outsider can audit.
What the headlines missed: voluntary commitments have moved faster than either statutory track on specific technical practices, such as red-team disclosure and incident reporting. The gap is not effort. It is accountability after the fact.
The test none of them pass cleanly#
Set the three side by side and ask a simple question: if a deployed model starts producing harmful outputs at scale, which regime catches it first?
The EU Act catches it through documentation and incident reporting, but slowly, and mostly for high-risk categories. The US sector model catches it through existing consumer protection or discrimination complaints, but only if someone files and the agency has capacity. Voluntary commitments catch it through internal review, but only if the lab chooses to look.
None of the three establishes routine, independent, post-deployment testing with published results. That is the gap. The policy debate has spent years on pre-deployment thresholds (compute, parameters, benchmark scores) and comparatively little on what happens in month nine of a production deployment, when the model has been fine-tuned, wrapped in tools, and pointed at a new user base.
What to watch instead of the headlines#
Three concrete signals will matter more than any single statute.
First, whether the EU's standards bodies publish testable criteria, not just management-system requirements. A standard you can audit with a script beats a principle you can only discuss in a workshop.
Second, whether US agencies share examination findings across sectors. A hiring model that fails a discrimination review may be the same model sold to a lender. Fragmented regulators rarely connect those dots.
Third, whether voluntary signatories publish failure cases, not just safety frameworks. A lab that reports what went wrong after deployment gives the public more than a lab that reports only what it tested before.
The three options differ in law, speed, and scope. They converge on one weakness: the period after launch is the least governed part of the lifecycle, and it is where most users actually meet the technology.
More articles