OpenAI, Anthropic and Google DeepMind want to set AI safety standards themselves

According to OpenAI's head of policy Chris Lehane, OpenAI, Anthropic and Google DeepMind have been coordinating on AI safety for several weeks. This emerged when Lehane met the press in Washington, where he also denied that the companies need special approval from the authorities to cooperate across competitors.

The talks were triggered by two notable moves. Last week, Anthropic CEO Dario Amodei wrote an essay urging the industry to jointly slow the pace of development of the most advanced models in order to avoid catastrophic risks. OpenAI's Sam Altman, Google's Demis Hassabis and Elon Musk all backed the call. As early as July, Hassabis had proposed in an essay a US-led standards body for AI that would act as a kind of financial regulator for the industry, with the power to screen the most advanced models and coordinate slowdowns across the industry if the dangers increased.

Joint testing body on the way

Concretely, several sources indicate that the three companies are discussing setting up a joint industry standards body to test advanced AI models before they are deployed. According to several sources, Altman has told his own employees that such a body must be established without backing from the US government.

At the same time, several of the companies are pointing to legislation. OpenAI supports a provision in the bipartisan FRONTIER Act in the House of Representatives that would force the leading AI labs to give independent verification organizations access to check that the models are being developed safely. This follows Altman having previously said that OpenAI will follow Anthropic's example and embed third-party evaluators inside the company to monitor safety.

The very question of how independent such internal evaluators can really be, however, raises doubts among researchers. They welcome the unprecedented access to the companies' AI labs but stress that meaningful oversight requires transparency, genuine independence and ultimately outright regulation if it is to work as intended.

The political climate in the US doesn't make the task any easier. President Trump has dismissed safety concerns as exaggerated and warned that slowing AI development would give China a head start. His close AI adviser David Sacks, who himself has financial interests in the industry, has followed suit and called the fear of existential AI risks overblown.

Skepticism in Denmark

In Denmark, the companies' move is met with skepticism. Data ethics expert Gry Hasselbalch expresses no confidence that the AI giants can regulate themselves, and instead urges the companies to follow the international legislation that already exists in the area. Several Danish media outlets point out that the proposal for a self-regulating body faces criticism on several fronts.

Other Danish professionals and politicians are discussing concrete control mechanisms such as kill switches and independent AI watchdogs that could halt development if it runs amok. Experts are unsure, however, whether such mechanisms would actually work in practice.

For the ordinary AI user, the news doesn't mean much concretely for now, but it's worth watching whether a joint testing regime for the big models emerges, and whether it is backed by independent oversight or regulated by law. In the long run, that could affect how thoroughly new AI models are tested before they become available to users and businesses.