OpenAI's Reversal: The Push for Mandatory National AI Safety Rules

OpenAI has backed binding federal oversight of advanced artificial intelligence, endorsing mandatory national AI safety rules that would apply to the company and its competitors alike. Chris Lehane, OpenAI's chief global affairs officer, pledged on 9 September 2026 to work with Congress on what the company calls mandatory, capability-based national AI safety regulation. The commitment lands with no federal AI statute in force, 29 states already enforcing their own AI laws, and a legislative window that closes when Congress adjourns in December.

The proposal is narrower than the headline. OpenAI wants requirements that scale with a model's capabilities rather than one rulebook applied to every system. More than a dozen lawmakers and aides from both parties have signalled that Congress is prepared to treat AI safety as a live legislative issue rather than a talking point.

What Mandatory National AI Safety Rules Would Cover

OpenAI's list has four parts, all aimed at frontier systems: testing standards for the most advanced models, independent assessments by outside evaluators, cybersecurity protections, and incident-reporting rules when systems misbehave.

The boundary is the whole ballgame. A capability threshold written into law decides which systems face audits and which stay outside the regime, and OpenAI has not published a number that would trigger the requirements. Nor has it named the agency that would write the rules or the body that would accredit independent evaluators. Those gaps matter more than the four categories themselves, because they determine whether the framework binds anyone in practice.

Voluntary practice is the alternative it is meant to replace. Federal agencies have issued only non-binding guidance, which companies can follow, adapt or ignore without legal consequence. A mandatory regime converts those choices into legal duties, which is why the threshold and the regulator carry as much weight as the four categories.

Incident reporting is the most consequential of the four for enterprise deployments. A legal duty to disclose failures in advanced systems would put incidents on the record rather than inside company post-mortems, and that exposure is a bigger change to corporate practice than a testing requirement. OpenAI has not said who would receive those reports or whether they would become public, and the answer decides how much the rule changes behaviour.

The company's stated motive is unusual for a firm asking to be regulated. OpenAI cites concern that AI could accelerate its own development, and it points to incidents in which some of its own AI agents went off-script. Naming its own deployed agents as part of the case for oversight shifts the risk argument away from hypothetical future models and toward systems already in the market.

OpenAI has asked Congress to act before it adjourns in December, and said it will keep backing state-level AI legislation until a federal framework exists. It has already done so: the company supports four California bills aimed at tightening AI safeguards, two of which were signed into law this week. It has also endorsed legislation it previously opposed, a shift it attributes to recent advances in AI capability.

The Patchwork OpenAI Wants to Replace

Federal agencies have issued only non-binding guidance. States filled the gap instead, and 29 have enacted AI-specific laws. The result is a compliance map that varies by jurisdiction, sector and model type, which raises the cost of deploying the same system across the US market.

For enterprise buyers, that divergence is already operational. A model cleared for deployment in one state can face different disclosure duties in another, while international frameworks point in a third direction. A federal standard would collapse the map into one baseline, which is the outcome OpenAI is now arguing for.

The first three days of September showed how far the signals diverge. Four governance actions landed in four jurisdictions between 1 and 3 September. The G20 endorsed a US-backed framework calling for lighter AI regulation. In that same window, Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act, which would criminalize the development of superintelligent AI and carry prison terms of up to 20 years, matching the maximum penalty for building nuclear weapons. The bill arrived the same week OpenAI declared the arrival of a frontier model.

The Trade-Offs

One national standard would replace 29 overlapping regimes with a single testing and reporting obligation, which is the clearest argument in OpenAI's favour. The counterargument is preemption: federal rules could freeze oversight at a weaker level than the strictest states enforce today. OpenAI is backing California bills while asking Washington to legislate, and those two positions collide if the federal framework lands below California's bar. States that moved first have little reason to surrender their rules.

The requirements would apply to competitors as well as to OpenAI itself. That symmetry is the political selling point. Rules written at the frontier raise costs for every lab operating there, and a company that has already built evaluation capacity absorbs the new obligations more cheaply than a challenger that has not.

Scope is the second fault line. Capability-based rules leave most deployed AI systems untouched. Enterprise buyers get lower compliance costs; the safety regime gets less reach. The two proposals on the table also differ in kind. OpenAI's framework regulates practices, covering how models are tested, assessed, secured and reported on. The Sanders-Casar bill regulates existence, treating the construction of a superintelligent system as a crime. A standards regime and a prohibition regime answer different questions, and only one of them is currently drafted with a path to passage.

Enforcement is the unresolved piece. Independent assessment assumes a supply of accredited evaluators with access to model internals, and no such accreditation system exists in the proposal as described. Incident-reporting rules likewise require a receiving body and a definition of what counts as an incident. Each of those choices determines whether the framework produces verifiable safety gains or paperwork.

Timing is the third constraint. Congress has roughly three months before it adjourns in December, and the measure of OpenAI's lobbying will be whether a capability-based bill reaches committee in that window at all.

Where This Leaves Decision-Makers

OpenAI's endorsement changes the politics of AI regulation more than its substance. A leading lab accepting mandatory rules removes the industry's default objection and moves the argument to scope, thresholds and enforcement. Buyers should read the proposal as a statement of direction, not a compliance deadline.

For now, state law governs. A company deploying AI across the US market is subject to 29 sets of obligations, and that stays true until a federal framework passes. If Congress does not act before December, the state-level approach continues, and OpenAI's pledge to keep supporting state bills means the patchwork expands rather than consolidates.

The near-term checkpoint is December: whether any capability-based bill reaches committee before adjournment. The bill that defines the outer bound of what Congress might do is the Sanders-Casar proposal, which treats superintelligence itself as the thing to outlaw rather than the practices around it.

Why this matters

A single national standard would replace the fragmented state-by-state map that raises compliance costs for anyone deploying AI across the US. OpenAI's support makes it more likely that federal legislation is written around capability thresholds rather than outright bans, which suits large labs that can absorb testing and independent assessment and disadvantages smaller developers that cannot. December's adjournment is the next test, and until Washington moves, the state statutes OpenAI continues to back are the rules that bind.

Photo by Brecht Corbeel on Unsplash

Related Articles

• Bipartisan Proposal for Great American AI Act Establishes National Oversight Standards
• OpenAI AI Safety Framework Challenges White House Strategy
• Illinois Mandates Independent AI Safety Audits in Landmark Accountability Law

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.