The 3-part plan to regulate AI before it regulates us

A heavyweight coalition of researchers from OpenAI, Google DeepMind, Microsoft, and a half-dozen top universities has dropped a detailed blueprint for regulating what they call “frontier AI” — the most advanced foundation models that could, without warning, develop capabilities dangerous enough to threaten public safety. The paper, authored by more than 30 people including OpenAI’s Jade Leung and Miles Brundage alongside Google DeepMind’s Joslyn Barnhart, doesn’t mince words: these models pose a unique regulatory headache because their dangerous abilities can emerge unexpectedly, misuse is tough to prevent once a model is deployed, and once those capabilities are out in the wild, there’s no putting the genie back in the bottle.

The authors lay out three essential building blocks for any serious regulatory regime. First, standard-setting processes that define what’s actually required of developers building at the frontier. Second, registration and reporting rules that give regulators genuine visibility into what’s happening inside AI labs — not just what those labs choose to share. And third, real enforcement mechanisms to ensure companies actually comply with safety standards for both development and deployment. The paper is careful to nod toward industry self-regulation as a useful starting point, but it’s blunt about the limits: voluntary efforts won’t be enough when the stakes are public safety.

What makes this paper land differently than the usual think-tank fare is the specificity around enforcement. The authors float granting supervisory authorities real teeth and even explore licensure regimes for frontier models — essentially, you’d need a permit to build or deploy the most capable systems. That’s a far cry from the current landscape, where labs largely decide for themselves what’s safe enough to release. The paper also proposes an initial set of safety standards that reads like a minimum viable regulatory checklist: pre-deployment risk assessments, external scrutiny of model behavior, deployment decisions that actually use those risk assessments, and ongoing post-deployment monitoring for new capabilities or misuse patterns.

The tension here is obvious and the authors don’t pretend otherwise. These are researchers who span industry, academia, and policy — and they’re trying to thread a needle between enabling innovation and preventing catastrophe. The paper explicitly frames itself as a contribution to a broader conversation about balance. But reading between the lines, the message to policymakers is clear: the window for getting ahead of frontier AI risks is narrow, and the current patchwork of voluntary commitments isn’t going to cut it. Whether governments will act with the urgency this paper implicitly demands remains the trillion-dollar question.