Musk Proposes AI Giants Mutually Test Models: Stop 'Grading Your Own Homework'

Beating AI News Flash: Musk proposed at the All-In Summit that frontier AI companies should let competitors conduct a round of safety testing before releasing new models. Each company would use its own safety testing system to find problems in others' models and directly point out risks, rather than continuing to rely only on self-verification.

Musk compared the current approach to "grading your own homework." Compared with self-testing, competitors have more incentive to find problems. If a competitor has already warned that a model is unsafe and the company still insists on releasing it, then if something really goes wrong later, it will be very hard to argue its case, and legal liability would also be very large.

He also believes that this mechanism should not be implemented unilaterally only by U.S. companies, and that it would be best to include major AI participants such as China as well, to avoid U.S. companies alone bearing restrictions unilaterally.