The safety evaluation process for frontier AI models remains opaque. OpenAI released GPT-4o without clear public disclosure of how U.S. government agencies assessed its risks before deployment, leaving a critical gap in accountability.

TechCrunch reports that the specific conversations between federal regulators and AI companies like OpenAI and Anthropic are not public. No formal regulatory framework exists requiring companies to obtain government approval before releasing advanced models. Instead, industry participants conduct internal safety testing, and government agencies may review models after deployment or through private channels.

This lack of transparency raises questions about what "safe to release" actually means. OpenAI likely ran internal red-teaming exercises and vulnerability assessments before launching GPT-4o, but the standards applied, benchmarks used, and government involvement remain undisclosed. The company operates under voluntary safety commitments rather than mandatory compliance.

The Biden administration has pushed for AI safety standards through executive orders and agency guidance, but no binding legal requirements govern frontier model releases. The National Institute of Standards and Technology published an AI Risk Management Framework, but adoption is voluntary. This creates a situation where companies self-regulate their own most powerful systems.

Industry observers argue this approach leaves the public with limited assurance about model behavior. GPT-4o handles sensitive tasks from medical advice to code generation, yet the decision-making process behind its clearance for public use remains largely behind closed doors. Anthropic similarly releases Claude models without transparent government sign-off.

Some regulators favor this light-touch approach, arguing prescriptive rules could stifle innovation. Others contend that frontier models warrant formal safety certification before release, similar to pharmaceutical approval processes. The current gap allows companies flexibility but limits public oversight.

The dialogue between government and AI labs will likely intensify as models grow more capable. Without clarity on evaluation criteria, the public cannot assess whether release decisions reflect genuine safety confidence or merely risk