Chris Hughes proposes ‘Office of AI Supervision’ to regulate AI labs

Chris Hughes, one of Facebook’s co-founders, is calling for the creation of a new government body called the “Office of AI Supervision” that would place regulatory teams directly inside the world’s most advanced AI laboratories. The proposal, laid out in a Financial Times opinion piece, envisions a model borrowed from the financial sector, where embedded examiners have been a fixture of bank oversight for decades.

The core idea is straightforward: put government staff inside companies like OpenAI, Anthropic, and Google DeepMind for fixed three-year rotations. After those three years, the regulators move on, replaced by fresh teams. The rotation mechanism is designed to prevent the cozy relationships that tend to develop when the same watchdogs monitor the same companies indefinitely, a problem regulators call “capture.”

How banking oversight became an AI blueprint

The three-year term limit is a deliberate design choice. In banking, one of the persistent criticisms of embedded supervision is that examiners start to see the world through the institution’s eyes. Rotating staff on a fixed schedule is meant to keep that dynamic in check, ensuring regulators maintain an outsider’s skepticism even while operating inside the building.

Hughes’s broader regulatory vision

This isn’t Hughes’s first foray into AI governance advocacy. In February 2025, he laid out a broader framework that included safety protocols, registration requirements, and third-party validation for AI models whose training costs exceed $100 million. That threshold is notable because it effectively targets only the largest frontier models, the ones being built by a handful of companies with the resources to spend nine figures on a single training run.

His earlier proposals also included the concept of “kill switches” for high-cost AI models, mechanisms that would allow systems to be shut down if they exhibited dangerous or unintended behavior. Combined with mandatory testing regimes and incident reporting, Hughes has been building toward a comprehensive regulatory architecture that treats advanced AI development more like nuclear energy than software.

The Office of AI Supervision proposal fits into a crowded global conversation about how to govern frontier AI. California’s SB 1047, which would have imposed safety requirements on large AI models, was vetoed by Governor Gavin Newsom despite significant industry debate. The UK’s AI Security Institute has been conducting evaluations of frontier models, while the European Union’s AI Office is working to implement the bloc’s AI Act.

What distinguishes his approach from many competing proposals is the emphasis on presence rather than process. Most regulatory frameworks focus on what companies must do: file reports, conduct evaluations, meet benchmarks. Hughes is proposing that regulators be physically embedded, observing development in real time rather than reviewing it after the fact.

What this means for the AI industry

For the three companies most obviously in the crosshairs, OpenAI, Anthropic, and Google DeepMind, the prospect of government teams operating inside their facilities represents a significant shift. These companies have built internal safety teams and published voluntary commitments, but having external regulators with institutional authority is a fundamentally different arrangement.

The proposal also raises practical questions that Hughes’s op-ed doesn’t fully resolve. Which agency would house this office? What qualifications would embedded staff need, and where would the government find enough people with the technical expertise to meaningfully oversee systems that even their creators don’t fully understand? Banking examiners benefit from decades of established methodology and clear financial metrics. AI supervision would require building equivalent institutional knowledge essentially from scratch, all while the technology itself evolves at a pace that makes banking look glacial by comparison.