Chris Lehane just published OpenAI's most aggressive policy push yet, arguing that the "AI policy window is open" and demanding immediate action on mandatory safety requirements. The framing is urgent, the rhetoric is sweeping, and the subtext is fascinating: OpenAI wants regulation now, but only regulation that applies to a "handful of well-resourced laboratories" and explicitly avoids becoming "open-weights policy by another name."
This is capability-based regulatory capture dressed up as democratic governance, and we should call it what it is.
The Framing: Defenders Window Meets Policy Urgency
Lehane's post anchors itself to two technical developments: the capabilities of Astra (OpenAI's latest model) and Jakub Pachocki's recent essay warning about recursive self-improvement. The argument is straightforward: AI is advancing quickly, "AI-accelerated AI development" is coming, and we need mandatory federal regulation before capabilities outpace institutions.
The "defenders window" framing—borrowed from Greg Brockman—is clever. It positions frontier AI as a temporary defensive advantage before "powerful offensive capabilities become widespread." This creates urgency without acknowledging the obvious: OpenAI is the entity developing those offensive capabilities, and they're asking for rules that cement their position before anyone else catches up.
The post explicitly states that "AI is already accelerating parts of the research used to develop and align the next generation of models," with research showing AI agents can perform tasks "that would take skilled researchers several days." But the logical conclusion—that OpenAI should slow down their own development—is never seriously entertained. Instead, the answer is policy that "can evolve as the technology does."
The Ask: Targeted Regulation for "Frontier" Labs Only
Here's where it gets interesting. Lehane lays out four policy priorities:
- Mandatory national AI safety requirements
- State-level momentum (supporting four California bills)
- Industry-led standards
- Global standards on when development should "slow or stop"
But buried in the details is the actual regulatory model OpenAI wants: "Frontier safety requirements should apply to the handful of well-resourced laboratories developing the most capable systems—not to startups, small developers, or researchers operating nowhere near the frontier."
This sounds reasonable until you realize it's a moat-building exercise. OpenAI gets to define what "frontier" means, advocate for "capability-based" thresholds that conveniently exclude competitors, and wrap it all in democratic governance language.
The post explicitly rejects regulation that might constrain open-weights models: "Nor should frontier safety policy become open-weights policy by another name." OpenAI even cites their support for the "Open Weights and American AI Leadership letter" as evidence of their balanced approach. Translation: regulate us (but only a little), don't regulate open models (because that would strengthen Meta and Mistral), and definitely don't create rules that would slow our commercial deployment.
The California Bills: Strategic State Support
OpenAI announces support for four California bills: SB 813 (independent safety assessments), AB 1405 (AI-auditor standards), SB 1119 (youth safety protections for companion chatbots), and AB 1864 (biological threat safeguards).
These are genuinely useful bills, and OpenAI's support matters. But notice what's not here: binding compute thresholds, mandatory pre-deployment government review, or any mechanism that would give regulators the power to actually stop a dangerous model from shipping.
SB 813's "independent assessment" infrastructure is explicitly framed as second-best to federal standards, and OpenAI makes clear they prefer assessments that maintain their "access to highly sensitive information"—which in practice means assessors who sign NDAs and can't publicly report findings.
The "reverse federalism" framing—where state laws create a de facto national baseline Congress can codify—is smart politics. But it's also a way to ensure the eventual federal framework looks like what OpenAI already supports, rather than more aggressive alternatives.
The Omissions: What They're Not Saying
Lehane's post claims OpenAI will "slow or stop the development or deployment of systems we cannot sufficiently safeguard," citing their Preparedness Framework as evidence. But the framework itself is self-governing—OpenAI gets to decide when risks are "unacceptable," and there's no external enforcement mechanism beyond bad PR.
The post mentions "universal monitoring of full trajectories, including chains of thought" for Astra, but doesn't explain whether those chains of thought are auditable by external researchers, what the decision criteria are for escalation, or who ultimately decides when to ship.
There's also zero discussion of compute governance, export controls, or data-center-level monitoring—the kinds of hard interventions that would actually constrain frontier development. Instead, we get "industry-led standards" and voluntary commitments that "will continue pursuing technical solutions to alignment."
Recursive Self-Improvement: The Real Question
To Lehane's credit, the post acknowledges that "fully autonomous recursive self-improvement" isn't happening yet, and that "we should not pursue it unless and until it can be done safely." This is the right stance.
But the next paragraph immediately walks it back: "AI is already accelerating parts of the research," and OpenAI's aim is to "safely build automated AI researchers" that make "the next one safer, more aligned, and easier to control, not simply more capable."
This is definitional jujitsu. If you define recursive self-improvement narrowly enough ("fully autonomous"), you can claim it's not happening while simultaneously pursuing AI-accelerated AI research. The distinction between "AI agents that automate research tasks" and "recursive self-improvement" is real, but it's also a convenient rhetorical buffer.
The honest question is: at what capability level does AI-accelerated research become recursive self-improvement? And who decides when that threshold is crossed? Lehane proposes "common ways to measure this progress" and "shared safety bars," but offers no details on enforcement or independent verification.
What Good Regulation Would Look Like
If OpenAI were serious about democratic governance of frontier AI, they'd support:
- Binding compute thresholds with pre-deployment government review above certain FLOP counts
- Public incident databases with mandatory reporting and external investigation
- Auditable alignment eval results that independent researchers can verify
- Compute governance that tracks training runs and gives regulators visibility into frontier development
- Liability frameworks that create real financial consequences for deployed harms
Instead, we get capability-based self-regulation, voluntary industry standards, and state bills that focus on auditor credentials and youth safety—important, but nowhere near the hard governance questions.
The Real Policy Window
Lehane is right about one thing: the policy window is open, and it's closing fast. But the danger isn't that we'll regulate too slowly—it's that we'll lock in a framework designed by frontier labs to protect their competitive position while offloading accountability to "independent assessors" and "industry-led standards."
OpenAI's policy push is sophisticated, well-timed, and strategically coherent. It's also exactly the kind of regulatory framework you'd design if your goal was to slow down competitors, avoid binding constraints on your own development, and maintain the appearance of responsible governance.
The tell is in what they're willing to give up: auditor standards, yes. Youth safety, sure. Biological threat safeguards, absolutely. But binding compute limits, mandatory pre-deployment review, or any enforcement mechanism with real teeth? Not in this proposal.
If we want regulation that actually governs frontier AI development rather than rubber-stamping it, we need to demand more than voluntary commitments and industry-led standards. The policy window is open. Let's not waste it on regulatory capture with good branding.