Google, OpenAI and Anthropic have been discussing a standards body they would create themselves since July

The three laboratories meet regularly to draw up a structure responsible for testing and auditing boundary models before they are made available. In principle, they converge, but much less so on management.

Representatives of Google, OpenAI and Anthropic have been meeting since July around the same project: to set up a sectoral organization responsible for technically evaluating and auditing the most advanced models before they are opened to the public. The information, reported on September 13 by The Information and then taken up the next day, describes regular work meetings, the last one dating from the previous week.

The common diagnosis can be summed up in one sentence: a border model should not be published without an outside view. Sam Altman told his teams that he supported the creation of such a testing and auditing body, specifying that the major laboratories would probably have to build it alone, without support from the US federal government. Demis Hassabis and Elon Musk were also in favour of the idea.

The disagreements are about who is holding the pen, who conducts the evaluations, and how often they need to be redone. The discussion comes in a context where Dario Amodei has publicly advocated for a more measured pace of development and for voluntary safety commitments pending regulation.