OpenAI, Google, and Anthropic planning an independent AI safety group
What's the story
OpenAI, Google, and Anthropic are planning to create an independent artificial intelligence (AI) safety organization by late 2026 or early 2027, according to The Information. The proposed organization is tentatively called the Standards Authority for Frontier AI (SAFA). It will operate independently of government control and focus on establishing practical benchmarks for public safety commitments made by these companies.
Leadership roles
SAFA to assist developers in assessing models
SAFA's main objective will be to provide direct support to developers assessing cutting-edge models before their commercial release.
The companies have approached former White House AI policy advisor Sriram Krishnan and Arati Prabhakar, ex-director of the White House Office of Science and Technology Policy, for the top post at SAFA.
However, it's still unclear if SAFA will conduct direct model assessments in-house or not.
Rising concerns
Formation of SAFA follows weeks of industry deliberations
The move to create a joint authority comes after weeks of intense industry discussions.
An OpenAI spokesperson recently confirmed that the company is actively discussing coordinated safety frameworks with Google and Anthropic.
The talks gained traction after Demis Hassabis, co-founder of Google DeepMind and Alphabet's chief scientist, proposed a US-backed standards body in July.
He suggested it could be modeled after a public-private partnership or self-regulatory watchdog like the Financial Industry Regulatory Authority (FINRA).
Safety concerns
Proposed body to focus on frontier AI models
As AI models grow more powerful, so do the concerns about their safety. The proposed SAFA would focus on frontier AI models, the most advanced systems being developed by companies.
These models can perform increasingly complex tasks, making questions around testing, security, and safety all the more important.
The organization is expected to create common standards for AI companies to assess their models before release and support third-party testing.