When the U.S. government greenlit OpenAI's latest frontier model for release, it raised an immediate and uncomfortable question: how, exactly, did officials determine it was safe?
The short answer is that nobody outside a very tight circle really knows.
A Process Shrouded in Secrecy
OpenAI and Anthropic have both been engaged in some form of pre-release dialogue with government agencies — but the substance of those conversations has not been made public.
"Exactly what that dialog looked like between the government and Anthropic and OpenAI is unclear."
This opacity is raising alarm among AI safety researchers and policy watchers who argue that meaningful oversight requires transparency, not just access.
Who Was in the Room?
The review process appears to have involved a small number of officials and technical evaluators, but no standardized public framework governed what they were looking for. Key unknowns include:
- What safety benchmarks or red-teaming results were reviewed
- Whether independent third-party auditors were involved
- What threshold of risk was deemed acceptable for public deployment
- How findings were documented or shared across agencies
Why This Matters Now
Frontier models — the most powerful AI systems at the cutting edge of capability — carry risks that are qualitatively different from earlier generations of software. Potential harms include biosecurity threats, large-scale disinformation, and autonomous cyberattacks.
The absence of a clear, replicable review methodology means the public has no way to verify whether the sign-off was rigorous or largely ceremonial.
The Broader Oversight Gap
The U.S. currently lacks a dedicated federal AI safety agency with binding authority. Existing reviews appear to rely on voluntary cooperation from labs — a model critics say is structurally inadequate for the stakes involved.
- The EU's AI Act sets legally binding requirements for high-risk systems
- The UK's AI Safety Institute has conducted evaluations of frontier models before release
- The U.S. equivalent remains nascent, underfunded, and without clear enforcement powers
What Labs Say vs. What Regulators Know
OpenAI has published its own safety evaluations and system cards, but these are self-reported. Government reviewers may have received additional internal data — though whether they had the technical capacity to independently validate it is another open question.
The dynamic creates an inherent asymmetry: the labs know the most about their models, and regulators are largely dependent on what they choose to share.
The Road Ahead
As frontier models grow more capable — and as deployment timelines accelerate — the pressure to formalize oversight will only intensify. The current approach, characterized by informal dialogue and voluntary disclosure, may not survive contact with a genuinely dangerous system.
What's needed, according to many safety researchers, is a mandatory pre-deployment review process with published criteria, independent technical auditors, and enforceable standards. Until then, the public is largely taking it on faith.
