Dorsey makes catastrophic risk the test for AI release restrictions

In “Open the frontier,” Jack Dorsey answers Dario Amodei’s AI pacing proposal with a policy test: open releases and outside scrutiny should remain the default, while withholding a model requires independently reviewable evidence of material catastrophic harm and the failure of narrower controls. Why it matters: This separates model-publication policy from runtime permissions. A model release can be open while agents still face sandboxing, credential isolation, tool limits, and halted unsafe experiments; blanket release controls can also raise compliance costs and entrench current labs or governments.

Try this: Review one model-release policy. Require it to name the catastrophic harm, supporting evidence, narrower controls tested, independent reviewer, appeal path, and reconsideration date; save the result as a policy diff.