Stakeholders in the United States are weighing an industry-funded body, under federal supervision, that would verify before release that developers have tested their models for a small set of catastrophic capabilities and applied their own published safeguards. What is to be verified is therefore compliance with frameworks the developers have written themselves, the benchmarks for capabilities being left to the new structure to set later. The proposals are taken here as a starting point. The question is whether the assessment of frontier models can be co-regulated at all: what the research on co-regulation in media and communications says about the conditions under which such arrangements work, what follows from those conditions for the place of independent research, and what makes the findings of such a body usable in other regulatory orders.
