OpenAI, Anthropic and Google discuss creating a common AI safety body

OpenAI, Anthropic and Google are negotiating the creation of an industry organization that could set common standards for testing and auditing the most powerful artificial intelligence models. A working group of representatives from the three companies has been meeting regularly at least since July.

The Information reports this, citing sources familiar with the discussions.

The negotiations have been going on for several months, with the latest known meeting taking place last week. The companies have not yet officially announced the creation of the organization, and its final structure, powers and operating rules are still undefined.

The main idea is for major AI developers to agree on common requirements for testing the most advanced models. This primarily concerns testing their capabilities and risks, independent safety evaluation and auditing before releasing new systems.

OpenAI CEO Sam Altman supported the creation of an AI testing and auditing organization at a recent company-wide meeting of employees. However, according to The Information, he believes that the largest labs will likely have to create such a mechanism on their own, without support from the US government.

Anthropic CEO Dario Amodei has also publicly called for closer coordination among developers. He proposed that companies agree on voluntary safety standards before full government regulation emerges.

The negotiations between OpenAI, Anthropic and Google began before this statement. The working group, which includes representatives of the companies below the CEO level, has been meeting since July.

One possible model for the new body is FINRA, the US self-regulatory organization for the financial industry, which monitors brokers and sets mandatory standards for them. This concept was previously suggested by representatives of Google DeepMind.

The White House has also considered the idea of creating a structure for self-regulation of the AI industry. However, according to The Information, a draft executive order to that effect stalled over the summer due to insufficient support within the Donald Trump administration.

Meanwhile, there is no consensus within the tech industry. Some companies and leaders fear that rules created by the largest developers could complicate competition for smaller players. There is no final industry agreement yet either.

OpenAI, Google, Anthropic and other major developers already cooperate through the Frontier Model Forum, created for research on safety of advanced AI models. But the new project implies a more specific system of common testing and oversight standards.

Based on materials from: The Information, Quartz, PYMNTS