The window for shaping AI policy is open right now — and according to OpenAI’s chief policy officer Chris Lehane, that window won’t stay open forever. In a new essay published on OpenAI’s site, Lehane makes a blunt case: if governments, companies, and civil society don’t act while the political and technological moment allows it, the rules that govern one of the most consequential technologies in history will be written either too late or too badly. It’s a serious argument. And it deserves a serious read.
What Lehane Is Actually Arguing
Strip away the policy-speak and Lehane’s core point is surprisingly straightforward: capability and safety need to scale together. As AI systems get more powerful — and they are, fast — the evidence base for their safety needs to keep pace. Not lag behind. Not play catch-up. Keep pace.
This isn’t a new idea in AI safety circles, but hearing it framed explicitly by OpenAI’s top policy official in a public essay is notable. OpenAI has sometimes been criticized for pushing capability ahead of demonstrable safety guarantees, so Lehane putting this principle front and center feels like a deliberate signal — whether to regulators, to critics, or to OpenAI’s own internal culture.
He also argues for something that tends to get lost in the Washington-versus-Brussels debate: shared standards. Not one country’s rules imposed on everyone else, but genuinely interoperable frameworks that different governments and developers can work within. That’s harder than it sounds. The EU AI Act, the US executive orders, the UK’s more flexible approach — they’re already diverging in meaningful ways, and companies building global products are caught in the middle.
The Political Timing Argument
There’s a political science concept Lehane is implicitly drawing on here — the idea of a “policy window,” where a problem, a political will, and a viable solution all line up at the same time. These windows close. Sometimes because a crisis passes. Sometimes because political priorities shift. Sometimes because the technology moves so fast that the proposed rules are obsolete before they’re passed.
Lehane’s warning is that AI is at one of those inflection points right now. There’s enough public awareness, enough legislative interest, and enough industry willingness to engage that durable policy is actually achievable. Wait another two or three years — as models become more autonomous, as geopolitical tensions around AI intensify, as public trust becomes harder to rebuild — and the conditions for good policy may not exist anymore.
That’s not a neutral observation. It’s a call to action dressed up as analysis. And I think he’s largely right, even if you’re skeptical of OpenAI’s motives for making the argument.
What “Stronger Safety Evidence” Actually Means
One of the more substantive parts of Lehane’s essay is the demand for better safety evidence. This is where it gets technically interesting — and where OpenAI’s position carries some internal tension worth examining.
What does “stronger safety evidence” look like in practice? Currently, the field relies heavily on evaluations — standardized tests that probe model behavior across dangerous domains like bioweapons, cyberoffense, and manipulation. But evals have real limits. They’re often designed after the capability they’re testing for. They’re gameable. And they don’t necessarily predict real-world behavior when a model is deployed in complex, unpredictable environments.
Lehane doesn’t prescribe a specific technical solution, but the direction of travel is clear: the industry needs something more rigorous. Third-party audits. Mandatory incident reporting. Interpretability research that actually tells you why a model does what it does — not just that it passed a checklist.
This connects directly to why Paul Christiano’s recent addition to OpenAI’s board matters so much. Christiano is one of the most serious technical minds working on alignment and interpretability. His presence signals that OpenAI wants the safety infrastructure to be credible — not just the marketing around it.
The Shared Standards Problem
Getting shared standards right is genuinely hard. Here’s why: every major AI player has an incentive to shape standards in ways that favor their existing approach. OpenAI’s safety frameworks, Google DeepMind’s model cards, Anthropic’s Constitutional AI methodology — these aren’t neutral technical documents. They reflect architectural choices, research bets, and competitive positioning.
International coordination on AI standards is further complicated by the fact that China’s AI development is proceeding under entirely different assumptions about transparency, accountability, and state control. Any “global” standard that doesn’t account for that dynamic is incomplete by definition.
Lehane doesn’t fully resolve this tension — it’s fair to say no one has — but naming it clearly as a priority rather than handwaving toward “international cooperation” is at least an honest start.
- Capability-safety parity: Stronger models require proportionally stronger evidence that they’re safe to deploy
- Shared technical standards: Interoperable frameworks across jurisdictions, not fragmented national rulebooks
- Third-party verification: Safety claims need external validation, not just self-certification
- Durable legislation: Policy built to last through multiple political cycles and model generations
- Incident transparency: Mandatory reporting when things go wrong, so the field can learn from failures
Who This Essay Is Really For
Reading between the lines, this piece is doing several things at once. On the surface, it’s a policy argument. But it’s also OpenAI positioning itself as a responsible actor at a moment when that positioning is commercially valuable. Enterprises buying into OpenAI’s growing suite of products want to know the company isn’t going to blow up their compliance programs. Governments considering OpenAI partnerships — and there are many — want the same assurance.
Lehane, who came from a political background before joining OpenAI, knows how to write for multiple audiences simultaneously. This essay works as a genuine policy brief and as a reputation management document at the same time. That’s not a criticism — it’s a skill. But readers should hold both interpretations in mind.
What the Competition Is Doing
OpenAI isn’t alone in making policy noise right now. Anthropic has been aggressive in Washington, pushing its own safety frameworks and participating in the AI Safety Institute’s model evaluations. Google DeepMind has similarly engaged with the EU AI Act process. Even the UK’s AI Safety Summit framework drew participation from across the competitive landscape.
The race to write the rules is real. Companies understand that whoever helps shape the standards will have a structural advantage when those standards become mandatory. If your existing safety program already satisfies the regulatory threshold, you’re ahead. If your competitor has to rebuild their entire evaluation pipeline to comply, you’ve just made their life harder.
This isn’t cynical — it’s just how regulatory capture works in industries where technical expertise is concentrated. The question is whether the resulting standards are genuinely good for the public, or just good for the incumbents who helped write them.
Why Timing Actually Matters Here
Lehane’s urgency is worth taking seriously independent of OpenAI’s motives. We’re at a genuinely unusual moment. AI capabilities are advancing faster than most regulatory timelines can track. The systems being deployed today — autonomous agents, multi-modal models running real-world tasks, AI integrated into critical infrastructure — are qualitatively different from what existed when most current policy proposals were drafted.
I’ve been covering this space long enough to remember when “AI regulation” meant rules about algorithmic hiring decisions. The policy conversation has had to sprint just to keep up with where the technology actually is. Lehane’s point that this window won’t stay open is, in my view, the strongest part of his argument — and the part that deserves the most attention from people outside OpenAI’s immediate orbit.
The decisions made in the next 18 to 24 months about how AI is governed will be harder to undo than most people realize. Institutions calcify. Jurisdictional precedents get locked in. And if the frameworks that emerge prioritize speed to market over accountability — or alternatively, prioritize risk-aversion over beneficial deployment — the costs will be measured in years of foregone progress or years of avoidable harm.
Given how much is already riding on getting this right — from AI agents accelerating research itself to models reshaping entire professional categories — the stakes of getting the policy layer wrong are genuinely high. Lehane’s essay won’t resolve the hard questions, but it names them clearly enough that it deserves to be part of the conversation. The harder work — building the coalitions, agreeing on technical standards, passing legislation that survives contact with reality — is still ahead.
Frequently Asked Questions
What is Chris Lehane’s role at OpenAI?
Chris Lehane is OpenAI’s Chief Global Affairs Officer, responsible for the company’s policy, communications, and government engagement. He joined OpenAI in 2024 after a career in Democratic politics and crisis communications, and has become one of the company’s most visible public voices on regulatory issues.
What does “AI policy window” mean?
It refers to the political science concept of a brief period when the right conditions align for meaningful policy action — public awareness, legislative interest, and workable solutions exist simultaneously. Lehane argues AI is in that window now, but that it will close as the technology outpaces political capacity to respond and as geopolitical dynamics harden.
How does this relate to the EU AI Act and US regulation?
The EU AI Act is already in force, with phased implementation through 2026 and beyond. The US has taken a more fragmented approach through executive orders and agency guidance rather than comprehensive legislation. Lehane’s call for shared standards is partly a response to this divergence, which creates compliance complexity for global AI companies and risks producing incompatible regulatory regimes.
Is OpenAI’s safety record strong enough to make this argument credibly?
That’s genuinely contested. OpenAI has faced criticism — including from former employees — that commercial pressure has at times outpaced safety review. But the company has also invested significantly in safety research and recently added credible external voices like Paul Christiano to its governance structure. Lehane’s essay is best read as a statement of intent as much as a description of current practice.
