New AI Review System, Same as the Old AI Review System | American Enterprise Institute
So where does the Great American AI Regulatory Effort stand? Before the administration’s new framework for reviewing advanced artificial-intelligence capabilities—one discussed with executives from leading AI companies on Tuesday—it employed what could be fairly described as an ad hoc, opaque, discretionary, and effectively mandatory system.
In other words, exactly the opposite of sound governance, especially if you’re concerned about government overreach and centralization, not to mention policy uncertainty.
And today? Pretty much the same, it seems. Or at least close enough for government work, as they say.
Based on Axios reporting, we still have an ad hoc system. The White House framework defines a covered frontier model as closed-source with state-of-the-art capabilities and national security risks. But it leaves those terms loose enough that the White House and the model makers could read them differently. Definitions would be nice, if possible. Also, the model review will run an assortment of administration officials rather than a single office or agency, with the exact process still unclear. In other words, it sounds like we’re still in a case-by-case situation. Republicans might trust Team Trump on this, but would they trust Team Newsom or Team AOC?
We still have an opaque system. This new AI rulebook, such as it is, won’t be published and the cyber benchmarks are classified by the order itself. Good luck to everyone not directly involved in the process—most tech companies, policymakers, researchers, US allies, and the American public—getting a handle on what’s going on exactly. As Neil Chilson, head of AI policy at the Abundance Institute, puts it, “There may be good reason to classify the benchmark used to test those models. There is no good reason to hide how the program works.”
We still have a highly discretionary system. Again, no public standard, while Trump officials have tremendous latitude to decide which models get reviewed, and it remains unsettled which “trusted partners” receive early access—a list that also won’t be made public. Adam Thierer of the R Street Institute makes a good point here: “If Joe Biden would have done anything close to what the Trump admin is doing now to approve AI models behind closed doors with zero transparency, every Republican would have lost their minds.”
Finally, we still have an effectively mandatory system. Voluntary in name, but this is the very same administration that blocked models previously. Declining to participate in this new framework carries an obvious and massive implied cost. More on this from Anton Leicht, a Fellow with the Carnegie Endowment’s Technology & International Affairs team: “No sane AI developer will take the risk of simply releasing a frontier model ever again.”
AI models are still evolving. Let’s hope AI policy in Washington will do the same, but in a way that moves toward openness, certainty, and decentralization.
