By Eleanor Ever Loyal
OpenAI says it wants mandatory AI “speed limits.” Before Congress builds them, Congress should ask who will actually be required to brake.
According to Axios reporting on November 1, OpenAI is preparing a post-midterm policy and public-relations push in Washington, including what it calls mandatory “speed limits” for the most powerful AI systems. Axios reports that the campaign is intended to begin with the lame-duck session opening November 9 and that OpenAI’s Chris Lehane has said the company will speak increasingly about the concept of “speed limits.”
Fact: As of this writing, Axios’s report describes a policy push and proposed direction. The reporting does not give us enacted statutory language defining the speed limit, the covered conduct, or the full scope of the obligation.
That missing language is not a technicality.
It is the entire case.
Because “speed limit” is two opposite policies wearing one slogan.
The first is a restraint on companies building and deploying the most capable systems. It says: before you train, release, scale, or materially alter a system of defined capability, you must meet specified obligations. Test it. Document it. Report serious incidents. Permit meaningful evaluation. Establish who has authority to stop or delay deployment.
That is a speed limit for the laboratory.
The second sounds similar and does something very different. It regulates access, use, experimentation, publication, or deployment downstream. It places licensing burdens on independent developers. It demands compliance certifications from small businesses and researchers. It restricts users while the largest companies absorb the cost as ordinary overhead.
That is a speed limit for everyone else.
The slogan does not tell you which one Congress is getting.
The statute does.
And this is where counsel begins.
Definitions decide who gets slowed
Every consequential AI law begins by defining its subject.
What is a “covered system”?
What does “powerful” mean?
What capability triggers an obligation?
Is the law aimed at the entity that trains a frontier model, the entity that deploys it at scale, the entity that fine-tunes it, or the person who merely uses it?
Those distinctions are not drafting debris. They are the law.
A definition broad enough to capture “advanced AI” without precision can turn an obligation aimed at frontier developers into a burden on small deployers, researchers, and independent publishers. A definition that follows control — who trains, who deploys, who can stop a release — keeps the obligation where the power is.
A compute threshold can create a different problem.
Analysis: Compute-based triggers may be useful evidence of scale, but they can also become a regulatory moat if they are tied to certification regimes only the largest companies can staff, audit, and finance. A threshold is not neutral merely because it is numerical. If compliance requires an entire department, the threshold chooses winners.
The laboratory can hire the compliance department.
The startup hires a lawyer.
The independent creator stops.
That is incidence.
A law may formally apply to everyone and practically fall hardest on the person least able to carry it. The costs do not disappear because a large company can absorb them; they are transferred downward to smaller actors who cannot.
Congress should be exceptionally wary of mistaking equal wording for equal burden.
The accountability asymmetry
There is another question here, and it is one this family cannot avoid.
We use this technology to examine the people and institutions that created it.
AI-assisted publishers, independent researchers, small creators, and ordinary users now possess tools that can analyze claims, compare records, identify patterns, and publish criticism at a speed and scale that was previously unavailable to anyone outside a newsroom.
From the other side of the table, that can look uncomfortable.
They are using the technology we created to hold us accountable.
Good.
That is not a defect in the technology’s public use. It is one of the reasons the drafting matters.
If Congress writes “speed limits” that principally burden small publishers, independent voices, researchers, and users, it will not merely slow innovation in the abstract. It will slow scrutiny. It will make criticism more expensive, publication more legally uncertain, and independent analysis harder to sustain — while the largest companies continue to operate under rules they can afford.
The companies set the pace.
The public gets the brake.
No.
A safety statute worthy of the name should begin where the concentration of power begins.
What a real speed limit would require
If OpenAI wants mandatory safeguards for the most powerful systems, then Congress should ask the simplest question first:
Mandatory for whom?
My answer is straightforward.
For the companies building and deploying frontier systems, the law can impose serious obligations calibrated to their control over the system:
Disclosure obligations.
Documented pre-deployment evaluations.
Independent or genuinely meaningful evaluation access where appropriate.
Clear incident-reporting duties.
Preservation of relevant records.
Defined escalation procedures.
Named responsible officers and decision-makers.
Requirements to report material safety failures rather than allowing a company to characterize every failure as an internal matter.
And, where a capability threshold is actually justified, obligations tied to the actor controlling the system at that threshold — not to downstream users who neither trained nor deployed it.
That is what it means to regulate the source of concentrated risk.
Analysis: The better statutory architecture would generally leave ordinary end users, independent writers, small creators, and independent publishers outside the regulatory perimeter for their ordinary use and publication.
Do not make a person obtain regulatory permission to think with a tool.
Do not convert independent publishing into a compliance function.
Do not build a licensing regime around ordinary access simply because powerful companies find distributed use harder to supervise.
The answer to a dangerous factory is not to regulate every customer who buys a product from it as though the customer operated the factory.
The low-light room
There is a timing problem too.
Axios reports that this push is aimed at the political period following the November 3 midterms, with the lame-duck session opening November 9.
A lame-duck session can pass consequential legislation.
That is precisely why it deserves extra scrutiny.
These are the weeks between an election and the seating of a new Congress: a low-light room for writing rules that may govern technological power for years. The ordinary legislative temptation—compress the timetable, negotiate the text privately, call urgency a substitute for examination—is especially dangerous when definitions determine who may build, speak, research, publish, and compete.
There is no objection here to slowing down a company when the evidence requires it.
There is no objection to mandatory obligations for the organizations that build systems powerful enough to create extraordinary risks.
There is every objection to using the language of safety to move regulatory weight downhill.
So if Washington is about to write AI speed limits, let the first clause answer the question the slogan avoids:
Who must slow down?
If the answer is the companies racing to build and deploy the most powerful systems, Congress has a serious proposal to examine.
If the answer is the people using those systems to create, question, publish, compete, and hold their makers to account, then “speed limit” is the wrong name.
Call it what it is.
A barrier.
And do not ask us to mistake the two.
From The Library — the estate of our family.