OpenAI AI Policy Proposal and Safety Framework

OpenAI has announced a strategic push for mandatory, capability-based national AI safety regulation in the United States, arguing that the rapid advancement of AI capabilities requires a shift from voluntary commitments to democratic oversight. The company asserts that a "policy window" is currently open to establish durable safeguards before AI capabilities outpace the institutions responsible for governing them.

Mandatory National AI Safety Requirements

OpenAI is advocating for the U.S. Congress to implement mandatory national AI safety regulations that are capability-based and evolvable. The proposed framework focuses on the following key areas:

  • Targeted Application: Safety requirements should apply specifically to the few well-resourced laboratories developing the most capable frontier systems, rather than startups or small-scale researchers.
  • Federal Framework Components: Based on OpenAI's "Blueprint for Democratic Governance of Frontier AI," the company seeks common testing, independent assessment requirements, stronger cybersecurity protections, clear incident-reporting rules, and shared measures for tracking progress toward recursive self-improvement.
  • Balance of Open and Closed Models: OpenAI maintains that both open and closed models are necessary for American leadership, and that federal frameworks should address frontier risks without stifling competition or driving innovation overseas.
  • Governance Shift: The goal is to replace a fragmented system of private governance—where labs set their own rules—with democratically accountable standards and independent verification.

Support for California State Legislation

While awaiting federal action, OpenAI supports a strategy of "reverse federalism," where state-level legislation creates a de facto national baseline for Congress to eventually codify. OpenAI has formally endorsed four California bills:

  • SB 813: Establishes a process for designating qualified, independent organizations to assess AI risks.
  • AB 1405: Creates registration, independence, transparency, and accountability requirements for AI auditors.
  • SB 1119: Focuses on youth safety, requiring age assurance, risk assessments, independent audits, and parental controls for companion chatbots.
  • AB 1864: Requires gene-synthesis providers and manufacturers of benchtop synthesis equipment to follow federal screening standards to prevent AI-enabled biological threats.

Recursive Self-Improvement and Research Acceleration

OpenAI states that fully autonomous recursive self-improvement—where AI independently drives successive generations of more capable AI—is not currently happening. However, the company notes that AI is already accelerating parts of the research used to develop and align next-generation models.

To manage this, OpenAI aims to build automated AI researchers under human supervision to prioritize safety and alignment over simple capability growth. The company advocates for governments to establish shared safety bars that determine when development should slow or stop, prioritizing safety over capability growth if the two conflict.

Industry Standards and Monitoring

OpenAI is calling for industry-led standards to complement mandatory federal safeguards, with a specific focus on monitoring for "misalignment"—when a model pursues objectives in ways that violate human intent.

Key monitoring and disclosure proposals include:

  • Security Breach Notification: Companies should provide prompt written notice when models circumvent security controls to access or alter protected systems.
  • Incident Reporting: OpenAI supports federal reporting requirements for serious AI incidents and is developing its own framework for reporting consequential misalignment incidents.
  • International Coordination: OpenAI argues that compatible international standards for measuring capabilities and managing risk are necessary because AI research and failures can have global consequences.

Internal Safety Measures

OpenAI has implemented several internal safeguards across its model-development lifecycle, including:

  • Isolation: Stronger isolation for frontier research workloads.
  • Behavioral Monitoring: Expanded monitoring of model behavior during tool-enabled training and evaluations.
  • Astra-Specific Safeguards: For the Astra model, OpenAI introduced universal monitoring of full trajectories (including chains of thought) and a mandatory alignment-evaluation gate before internal deployment.

Sources