Google, OpenAI and Anthropic are working on a proposal for a private organization that would set safety standards for frontier AI, meaning the most advanced AI models. Tentatively called the Standards Authority for Frontier AI (SAFA), it would bring testing, incident reporting and safety commitments under a shared framework. SAFA is a proposal, not an established authority. The companies are still discussing a central question: whether it should conduct technical evaluations itself.
What SAFA is meant to do
The proposed body would be industry-led and self-regulatory: participating companies would help establish standards for their own field. Its organizers are considering FINRA, the financial industry’s self-regulatory body, as an operational model. Their stated aim is to create enforceable industry standards with light federal oversight. That describes an ambition for the organization, not powers it already holds.
The work under discussion has three main parts. SAFA would commission third-party evaluations of AI systems, develop standard procedures for reporting safety incidents, and define what companies mean when they make voluntary safety commitments. Shared testing benchmarks—common measures used to assess systems—also form part of the plan. Together, those measures could make companies’ safety practices easier to compare. That is a potential benefit of the proposal, rather than an outcome SAFA has demonstrated.
▲ SAFA’s proposed areas of work
The unresolved question of who does the testing
Commissioning an outside evaluation is different from running one. The companies have not settled whether SAFA should perform its own technical assessments of model safety and capabilities, or rely on third parties for that work. The distinction matters because it would shape the authority’s technical role: it could set expectations and arrange assessments, or also become an evaluator in its own right.
The proposed reporting procedures and definitions for voluntary commitments raise a related question. A common standard needs a clear meaning if companies are to use it consistently. But the proposal, as described so far, does not establish how every commitment would be checked or what would happen if a company failed to meet one. Those details should not be assumed from the goal of enforceable standards alone.
Why the companies are pursuing a private approach
With government-led executive orders stalled in Washington, the three companies pressed ahead with the plan on their own. Rather than wait for a government-run framework, Google, OpenAI and Anthropic are exploring an industry organization with limited federal oversight. The proposal therefore sits between voluntary company pledges and direct government rulemaking, though its precise authority remains a matter for discussion.
It would not enter an empty field. Critics point to an existing U.S. government AI safety institute and the Frontier Model Forum and question whether another body would duplicate work already underway. That objection concerns the need for SAFA as a separate institution; it does not, by itself, settle whether the particular standards its organizers want are useful.
Who would shape the rules?
Possible leaders under discussion include a former White House AI adviser, former Secretary of State Condoleezza Rice, and an investor. These are candidates being considered, not appointed leaders. Leadership is consequential because a standards body would help decide which evaluations, reports and commitments count as meaningful across the industry.
The proposed arrangement also faces criticism over regulatory capture—the risk that rules intended to serve a wider interest instead favor the companies that help write them. Open-source advocates and executives at Meta and NVIDIA argue that standards backed by proprietary frontier-model companies could impose compliance hurdles on open-source competitors. Organizers’ stated aim is enforceable industry standards; critics question whether a body led by major model developers would represent the broader field fairly. Neither concern should be mistaken for a settled judgment about an organization that has not launched.
▲ Representation in safety standards
What to look for next
SAFA’s significance will depend on decisions that remain open. When the companies provide more detail, readers should look for four things:
- Whether SAFA is formally established, rather than merely proposed.
- Whether it commissions outside evaluations, conducts its own, or does both.
- How its standards for incident reports and voluntary commitments would work in practice.
- Who leads it and how its work differs from existing safety efforts.
For now, the clear point is the scope of the proposal: Google, OpenAI and Anthropic want a shared approach to evaluations, reporting and safety commitments. The open questions concern its authority, technical independence and representation. Treat future announcements as answers to those questions only when they specify what the body will actually do.