The Biden-Harris administration is reportedly finalizing a landmark Executive Order that would establish a mandatory vetting process for frontier AI models, closely mirroring the safety and efficacy reviews conducted by the FDA for new pharmaceuticals. This shift marks the end of "voluntary commitments" and the beginning of a regulated pre-release audit era for Artificial Intelligence.
Under the proposed framework, any AI model exceeding a specific compute threshold (estimated at 10^26 FLOPs) must undergo a rigorous battery of tests before being deployed to the public or enterprise customers. These tests, conducted by the U.S. AI Safety Institute (NIST), will focus on:
Mirroring the FDA's monitoring of drugs after they hit the market, the EO would require AI developers to maintain "incident ledgers." Any significant model failure—such as the Canvas LMS breach potentially being enabled by agentic tools—must be reported to federal authorities within 72 hours.
While safety-focused labs like Anthropic and OpenAI have expressed cautious support for centralized vetting, the open-source community is sounding the alarm. Critics argue that "compute thresholds" are a crude instrument that could stifle innovation in Small Language Models (SLMs) and grant an anti-competitive advantage to the well-funded hyperscalers who can afford the compliance overhead.
From a technical perspective, vetting a non-deterministic neural network is significantly harder than a chemical compound. NIST will need to develop adversarial testing frameworks that are themselves powered by advanced AI, leading to an "AI vs. AI" arms race within the regulatory sphere.
The "FDA for AI" model represents a fundamental change in the social contract between Big Tech and the government, prioritizing national security and public safety over rapid deployment.