OpenAI, Anthropic and Google are reportedly working together on a new organisation that would establish safety standards for sophisticated artificial intelligence systems. The new body is tentatively called the Standards Authority for Frontier AI, or SAFA.
According to a report by The Information, the three companies are looking to launch such an organisation by the end of 2026 or early 2027. The body would be independent of government administration and would focus on practical standards for testing and evaluating frontier AI models.
The development comes as AI companies have increasingly challenged the quality of models they want to test before they are released on the market. OpenAI's global policy chief Chris Lehane said earlier this month that the company had been working with Anthropic and Google DeepMind on AI safety issues for several weeks.
The new proposal would take that cooperation to a higher level by creating a dedicated industry body.
What is SAFA and what could it do?
SAFA, short for Standards Authority for Frontier AI, is expected to concentrate on practical testing, auditing and safety standards rather than the broad principles alone.
As far as the report points out, the organisation could support third-party safety testing of AI models before deployment. It may also establish protocols for AI safety and security incident reporting and establish criteria for independent organisations to audit AI systems.
Another issue is whether SAFA itself should conduct safety and capability assessments of advanced AI models.
The debate is partly aligned with the existing Center for AI Standards and Innovation (CAISI), a US government body that works on AI standards and evaluation. Some people involved in the discussions believe government resources may not be sufficient to evaluate increasingly complex frontier models and have considered giving SAFA a direct testing role.
But the exact roles of the proposed organisation have not been finalized.
Why are the three AI companies collaborating?
The proposed program comes at a time when AI systems are coming to dominate the industry’s knowledge and risk.
OpenAI, Anthropic and Google are competitors in the field of frontier AI models, but the companies have also increasingly discussed the need for common approaches to safety testing.
In September, OpenAI said it was working with Anthropic and Google DeepMind on AI safety. Anthropic CEO Dario Amodei had called for more cooperation among the AI industry in terms of the risks associated with the development of AI systems.
Google DeepMind CEO Demis Hassabis had also previously proposed the creation of a frontier AI standards organisation. It was compared with industry-led standards and self-regulatory structures used in other sectors.
The SAFA plan is therefore a move towards greater coordination among some of the companies developing the world's most advanced AI models.
Who could lead SAFA?
Leadership for the proposed organisation has not yet been finalised.
According to reports, the companies have looked at several prominent figures as a chief executive. Sriram Krishnan, a technology investor and former White House AI policy adviser, was among the top people to be considered by the companies, one source said. Former White House Office of Science and Technology Policy director Arati Prabhakar was also considered as a senior leader.
The conversations are still preliminary, and there has been no confirmed announcement of the leadership of SAFA.
How would SAFA differ from government regulation?
One of the important things about the proposal is that SAFA would supposedly operate independently of government administration.
That would make it an industry-led standards organisation rather than a government regulator. It would aim to establish realistic benchmarking, testing and audit standards that AI developers and evaluators could use for developing advanced models.
At the same time, the interaction between SAFA and existing government institutions is still an open question.
The US already has government-led efforts focused on AI standards and evaluation. A new industry body would need to make sure its work would complement existing institutions, rather than duplicate their work.
There is also the involvement of others. The current discussions are between OpenAI, Anthropic and Google, and the AI industry for the most part includes the likes of Meta, Microsoft and many smaller developers and open source projects. And whether the standards will be adopted in more of the industry will be up for debate.
SAFA launch is not yet confirmed.
Although there are reports about the planned organisation, SAFA has not yet been formally launched.
The Information stated that the companies are looking to launch by the end of 2026 or early 2027; other reports have similarly described the timeline as late 2026 or early 2027.
The name, leadership structure, membership, testing responsibilities and final operating model could be changed before any official launch.
If established, SAFA could be part of the growing ecosystem of organisations working to establish common safety practices for frontier AI. For now, however, it is still an industry initiative, not an established regulatory authority.
