States push beyond voluntary AI promises

The next test of AI safety promises is who can check them and what happens when they fail. This week's state initiatives offer a different approach from the White House's voluntary agreement with leading technology companies.

The debate also gained another prominent departure. In an October 3 essay in The Atlantic, David Robinson said he had resigned from OpenAI, where he led the writing of launch safety reports. He argues for stronger outside incentives and expertise from established safety-critical industries. Those are his recommendations, not an independent finding about the company's practices.

What remains voluntary

Reuters reported on October 3 that the September 29 White House accord contains no stated consequences for noncompliance. It includes internal controls and external auditors, but leaves enforcement details unspecified.

The administration says the agreement supports responsible development and innovation. Critics quoted by Reuters argue that existing legal remedies after harm occurs are insufficient to prevent a major incident. These are competing approaches to governance; a shared vocabulary of safety does not mean agreement on binding limits.

States are pursuing another route

On October 1, AP reported that Maryland Governor Wes Moore was forming a bipartisan governors group with Indiana Governor Mike Braun to develop an AI framework. That effort is a plan for coordination, not a completed national standard.

California has already taken legislative steps. In its September 30 announcement, the governor's office described SB 813's framework for certifying independent AI verification organizations and AB 1405's registry and standards for AI auditors.

The same announcement identifies emergency shutoff requirements and third-party safety plans as proposals under consideration. They should not be confused with rules already enacted. Nor does passing an oversight framework establish that its implementation will be effective.

Our view: follow the authority

For each new commitment, ask who gets access to evidence, who pays the evaluators, who receives their findings, and who can require a change. Independence needs practical meaning beyond a job title.

International coordination has a different role. The White House's September 25 fact sheet announced agreement on a U.S.–China incident communication channel. As our September 26 brief noted, it did not specify a launch date or reporting procedures. A commitment to communicate is distinct from an operational response system.

We support meaningful safeguards and public accountability. This week's developments deserve attention, but none alone settles whether increasingly powerful AI can be developed safely. Judge the eventual access, obligations and implementation—not the ceremony announcing them.