roobrief

AI

Google, OpenAI and Anthropic Reportedly Plan Their Own AI Safety Rules — What Changes for Developers

AI model safety is under review, and before any law is written, AI companies’ terms of service may change first and apply to developers.

Google, OpenAI and Anthropic are reportedly preparing a group that would set AI safety rules themselves. It is called SAFA (Standards Authority for Frontier AI). It would decide things like what tests a new AI must pass before release, how incidents must be reported, and who is allowed to audit.

The details come from The Information, which cited sources who were not named. The three companies have not announced anything officially. A launch is being talked about for late 2026 or early 2027.

Why the companies are stepping in

The U.S. has no federal law that directly regulates frontier AI yet. According to the reports, these companies first wanted the government to oversee them, but a draft executive order failed to win enough support inside the administration, so they are heading toward a private group instead.

SAFA is reportedly modeled on FINRA, the self-policing group that U.S. brokerage firms set up. How rule-breakers would be punished, and whether SAFA would run model tests itself or leave that to outside testers, has not been decided.

Where the government and Congress stand

  • Executive order (June 2): an order the president can issue without Congress. This one is voluntary: companies are asked to give the government up to 30 days of early access before a model is released. It does not require anyone to get permission before launching.
  • House FRONTIER Act (July 23): a bill that would make large AI developers undergo audits every year and protect employees who report problems inside their companies.
  • Senate: on October 7, Senator Cantwell published a set of principles, not a bill. It says models should pass an independent audit before they are released.

What could change for developers

Nothing is required of developers today. SAFA doesn’t exist yet, the executive order is voluntary, and the bills haven’t passed. But once rules settle, they will likely reach developers through AI companies’ API terms and release processes. What follows is our estimate based on the reports.

  • Terms of service may change. If incident reporting or outside audits become expected, AI companies are likely to update their API terms. Agent features, where an AI operates a computer for the user, could be among the first affected.
  • New models could arrive later. Joining the government’s early-access review adds one more step before release.
  • Safety results may become easier to compare. If tests are standardized, evaluation documents from different models can be read side by side.
  • Small teams could be hurt. Critics say that if big companies set the audit standards, meeting them costs more for small teams and open-source projects. A House discussion draft also includes a three-year freeze on new state AI laws.

If you depend heavily on one model, keep your AI calls in one place in your code and keep your own test set, so you can switch providers later without a rewrite.

What to watch

An official SAFA announcement and its draft standards, whether the freeze on state AI laws survives in a final bill, and changes to AI companies’ terms of service.

Sources

Source: thenextweb.com

← Back