OpenAI published a policy paper on 6 July 2023 calling for regulation of "frontier AI" systems, defining them as models that exceed the capabilities of today's most advanced systems. The paper proposes a licensing regime for AI systems above a computational threshold, with governments empowered to revoke licenses if safety standards are not met.

The document states that frontier AI systems pose "emerging risks to public safety" and recommends that governments require safety evaluations before deployment. OpenAI suggests regulators should mandate red-teaming, third-party audits, and incident reporting. The paper does not specify which current models would fall under the proposed threshold, nor does it detail what constitutes an acceptable safety evaluation [source].

The timing follows public criticism of OpenAI's March 2023 release of GPT-4, which some researchers argued lacked sufficient external review. The policy paper acknowledges that "companies creating AI systems have a responsibility to ensure they are safe" but frames regulation as necessary because "the actions of a single company or country could affect the entire world."

OpenAI's proposal includes a call for international coordination, comparing AI governance to nuclear safety regimes. The paper suggests that compute thresholds could be measured in floating-point operations, though it does not propose a specific number.

The document represents a shift from OpenAI's earlier public statements emphasising self-regulation. In 2022, CEO Sam Altman testified to Congress that the AI industry should "regulate itself" before government intervention. The new paper states that voluntary commitments are insufficient for frontier systems and that "regulation will be critical to mitigate the risks."