Elon Musk Urges Rivals to Test AI Models Before Release

Elon Musk wants leading AI developers to let their competitors inspect new models for safety and security risks before public release. The proposal turns commercial rivalry into a safeguard: companies may be more willing to flag a dangerous capability in a rival’s system than to find weaknesses in their own.

Speaking at the All-In Summit in Los Angeles on Monday, Musk said OpenAI, Anthropic, Google, Meta, his own SpaceXAI and several leading Chinese companies should run a shared safety “test harness” on one another’s models before launch. He acknowledged the approach would not be perfect, but argued it would make problems more likely to be found.

A review system built on rivalry

Musk’s model is deliberately informal. In separate remarks, he suggested weekly or biweekly calls and one to two weeks of early access for competitors. His premise is that rival firms have both the technical ability and the commercial motive to identify security or safety concerns, then publicly argue that a release should be delayed.

• Reviewers would be competing AI companies rather than the developer alone.
• The review would happen before a model reaches the public.
• Government intervention, in Musk’s view, would be reserved for a company that refuses to reduce a very serious danger.

Instead of grading your own homework, you would at least have competitors grading your homework and raising the alarm if they see concerns.

Elon Musk, speaking at the All-In Summit, according to CNBC

Competitors versus independent evaluators

Musk’s approach differs from the outside-review model proposed by Anthropic CEO Dario Amodei. Amodei has suggested third-party assessments by embedded evaluators who would verify safety practices. He has also argued that model development is moving faster than safety research can keep pace.

The distinction is practical. An embedded evaluator is intended to independently examine a company’s practices. Musk instead wants companies with a direct competitive interest in a rival’s delay to inspect the model. He compared that industry-led structure to the Motion Picture Association’s movie-rating process.

The agreement problem remains

For now, the idea is a proposal, not a pact. Musk said labs competing with SpaceXAI had not agreed to take part. His expectation that Chinese companies might join is also an expectation, not a commitment.

That gap matters as calls for tighter AI safeguards meet resistance in Washington. The Trump administration has rejected calls for greater AI regulation; National Economic Council Director Kevin Hassett said the private sector was the right place to address concerns, while saying law enforcement could be used when necessary. Musk’s proposal fits that preference for an industry response, but it still depends on the very competitors it seeks to enlist.

The proposal therefore offers a different answer to the recurring question of who tests powerful systems. Rather than slowing releases through a formal regulator or relying on independent assessors, Musk would ask rivals to scrutinize one another first. Whether that incentive produces rigorous review, strategic accusations, or no agreement at all remains unresolved.