OpenAI Calls for Mandatory National AI Safety Rules
OpenAI has urged the US Congress to enact mandatory national AI safety regulations before its adjournment, outlining a comprehensive set of requirements for itself and its competitors. This proposal, published by Chris Lehane, OpenAI’s chief global affairs officer, includes:
- Common testing and independent assessment protocols
- Stronger cybersecurity requirements
- Clear incident reporting rules
- Mandatory monitoring for model misalignment
- Prompt written notice of model security breaches
- Alignment-evaluation gates before deployment
Additionally, OpenAI supports four California bills addressing independent evaluation, auditor standards, biological threat screening, and child protection in chatbot use. Two of these bills were signed by Governor Gavin Newsom on Tuesday.
The company advocates for a state-led approach, known as "reverse federalism," where states establish a de facto national baseline that Congress can later codify. This strategy is not without controversy, as it raises concerns about state "mission creep" and the expertise of state officials in highly technical AI reviews.
OpenAI’s proposal bears similarities to the EU’s AI Act, with its focus on capability tiers, independent assessment, incident reporting, and pre-deployment evaluation. However, while Brussels has legislated the AI Act and subsequently agreed to thin it out, OpenAI is requesting a similar framework in the United States, with adjustments.
The politics around AI regulation are complex, with OpenAI’s Super PAC, Leading the Future, opposing state-level rules, while the company itself supports specific state bills. This duality highlights the nuanced landscape of AI governance, where the balance of power between federal and state authorities is being carefully navigated.