What Should Be Done

An analysis of the current state involves concerns regarding the regulation and safety standards for frontier AI models, stemming from executive actions taken by the Trump administration. The author argues that an executive order establishing a voluntary testing program effectively created a de facto involuntary licensing regime for frontier models, evidenced by restrictions placed on Fable and subsequent limitations on GPT 5.6 access for certain US companies. A central problem identified is the lack of clear, established safety requirements, as the government appears unsure of the necessary standards for broad model release. The author suggests that current safety standards are undeveloped because officials involved in AI policy lacked relevant experience, hindering the creation of effective, timely standards.
The text posits that while acknowledging the catastrophic risk potential of frontier AI is valid, the response has been complicated by a perceived reluctance among some officials to act decisively on safety warnings due to political considerations. The author shifts focus from blaming administration officials or industry figures for the current situation toward recognizing that the pace of AI progress itself dictates regulatory necessity. The ultimate suggestion involves establishing an auditing system for frontier labs, utilizing technical experts, and seeking collaboration between private oversight bodies and government entities to establish safety standards based on real-world experience, while also considering how this framework can address underlying societal disagreements about what constitutes safety.

* President Trump signed an Executive Order on Cyber and AI earlier this month.
* The Executive Order claimed to establish a voluntary testing program for frontier AI models.
* The author argues the Executive Order established a de facto involuntary licensing/preapproval regime for frontier models.
* The administration revoked public access to Anthropic’s Fable due to security fears.
* OpenAI’s GPT 5.6 is reportedly limited to a small set of US companies at the request of the US government.
* Nobody knows the requirements for obtaining a license.
* Government officials are unclear on the safety standards required for broad model release.
* The author suggests that without established standards, requests for public release will likely be denied by the government.
* Senior administration officials involved in AI policy lacked frontier AI experience.
* There is no clear timeline for developing a high-quality safety standard.
* Frontier models are trained at enormous costs, and post-release availability affects economic margins.
* Restrictions on model releases may have caused the US to overbuild AI infrastructure.
* The author proposes using frontier labs’ safety and security frameworks as a starting point for standards.
* State laws exist in California, New York, and Illinois requiring periodic updates to safety frameworks disclosed by labs.

The narrative demonstrates a tension between the rapid, emergent reality of frontier AI development and the slow, politically constrained process of establishing governance. A critical pattern emerges where the lack of concrete technical consensus on "safety" is being addressed through political maneuvering rather than purely empirical safety science. The argument suggests that attributing the current regulatory lag to specific actors (administration or industry) is a deflection from the core issue: the difficulty in quantifying and standardizing emergent risk in a rapidly evolving technological domain.
The proposed solution—creating independent verification bodies composed of technical experts, potentially supported by government certification—points toward institutionalizing real-world experience as the foundation for safety standards. This echoes historical patterns where specialized knowledge is necessary for governance, yet this proposal grapples with the inherent conflict between technical necessity and political legitimacy. The ultimate tension lies in whether the pursuit of technical certainty can overcome fundamental societal disagreements about what constitutes acceptable risk; if AI itself models game theory without regard for human consensus, then purely technical standards may be insufficient to resolve the resulting power dynamics.
This structure highlights a systemic vulnerability: the prioritization of speed over foundational clarity risks creating an ecosystem governed by speculation rather than verifiable safety. The implication is that if regulators cannot achieve agreement on what "good" looks like through shared understanding (technical or philosophical), then the system will default toward the interests of those with existing power, leading to an outcome that may be inconsistent with democratic principles, regardless of technical safeguards implemented in isolation. What mechanisms exist for embedding deep, autonomous technical expertise into oversight structures without succumbing to established political inertia?