AskwhoCasts AIJune 26, 202621m

What Should Be Done- By Dean W. Ball

TL;DR

De facto licensing regime: The administration's executive order effectively requires pre-approval for frontier model releases, with no clear standards for what constitutes safe enough.
Regulatory vacuum: The government does not know what it wants from labs, and the staff tasked with creating standards lack frontier AI expertise, with CAISI on a stop work order.
Economic risk: Restricting model releases threatens the business case for $100 billion data centers and could cascade into broader damage to nuclear, energy, and heavy industry investments.
Real-world experience required: You cannot think your way to AI safety; it requires iteration with actual users and real-world feedback on guardrails.
Private auditors as solution: Independent verification organizations, licensed by government but operating privately, could audit labs against their own safety frameworks while staying technically credible.
Regulate the lab, not the model: Focus regulatory attention on the entity developing AI rather than individual model releases, which become obsolete quickly and may soon change continuously.

The Breakdown

Dean Ball argues that the US government has created a de facto pre-approval regime for frontier AI models without defining what safety standards would permit release, leaving labs like OpenAI and Anthropic unable to deploy their most powerful systems. He proposes a system of independent verification organizations, privately run but government-licensed, that would audit frontier labs against their own safety frameworks and certify compliance.