Elon Musk isn’t buying the idea that AI labs can simultaneously warn their models could be dangerous and then ask investors for billions. At the All-In Summit, Musk called that combination "crazy 4D chess" — but then offered a surprisingly practical way to test whether those safety warnings are genuine.
Musk Questions AI Safety Claims
The Space Exploration Technologies Corp. (NASDAQ:SPCX) CEO said the danger from AI is "very significant" and that the industry needs to do better on AI safety, citing warnings from researchers at Anthropic and OpenAI. But he questioned the incentives behind companies publicly describing their own models as potentially dangerous while seeking capital to scale them.
"It certainly is like some crazy 4D chess," Musk said, pointing to the contradiction between warning of a potential 10% chance of human annihilation and asking investors how much allocation they would like in an IPO.
The sharper point, however, came after the joke.
Asked to get specific about the risks, Musk pointed to cyberattacks and the possibility of AI systems eventually gaining control of military systems.
He also questioned whether supposedly isolated systems are truly insulated if they still receive software updates through physical media.
That moves the debate away from whether AI is "safe" in the abstract and toward whether its safety claims can actually be tested.
Musk Wants AI Labs to Test Each Other
Musk’s proposed solution is essentially competitive peer review.
He suggested that AI companies provide competitors with API access to new models before release. If another lab identifies a serious safety problem, it could raise the issue publicly if the developer fails to fix it.
The incentive, Musk argued, would be powerful: If competitors publicly warned that a model was unsafe and that model subsequently caused harm, the consequences for the company that released it could be enormous.
Musk also said the testing and safety apparatus could be open-sourced, allowing others to inspect what is happening "under the hood."
That creates an unusual twist in the AI safety debate. Instead of relying entirely on regulators to police frontier models, Musk is effectively proposing that AI companies become each other’s adversarial testers.
The Stakes Are Bigger Than Regulation
Musk’s proposal also reflects his broader concern about regulation becoming difficult to reverse. He argued that oversight tends to work like a "one-way ratchet," making additional regulation easier to impose than remove. His preference was for a step that could happen quickly and potentially win agreement from both the U.S. and China.
For AI investors, that distinction matters. Safety concerns are no longer just an existential-risk debate playing out among researchers. They could affect how frontier models are tested, released and ultimately held accountable.
Musk may have laughed off the idea of AI safety warnings as "4D chess." His proposed test, however, could make the game considerably harder to fake.
Image via Shutterstock
Login to comment