Chris Hughes, co-founder of Facebook and author of 'Marketcrafters,' argues that AI companies require regulatory oversight modeled on bank supervision. Dario Amodei, CEO of Anthropic, advocates for independent evaluators to have real-time access to company systems, similar to how the Federal Reserve embeds teams inside commercial banks. This approach would allow supervisors to review internal documentation and assess whether leadership is managing risk adequately, rather than simply testing systems before release.

Under this model, embedded supervisors would ensure humans and AI agents are focused on preventing bad behavior and oversee stress tests to verify humans can override models under extreme conditions. The article warns that policymakers failed to regulate social media due to being overwhelmed by diverse issues like teen mental health and misinformation. To avoid repeating this with AI, Congress should start with the most pressing concern—model safety—and build from there.

Establishing a dedicated regulatory agency, potentially the Office of AI Supervision, would translate requirements into rules. Lessons from banking supervision highlight the need for staff who understand the technological paradigm and are willing to use their power. Senior examiners should rotate between companies to prevent capture, ensuring accurate risk assessment and effective oversight of frontier AI labs.