Elon Musk said on October 16 that SpaceX may adopt the same kind of independent evaluation that Anthropic has begun to accept, a remark that extends a practice built for software models into a business that builds rockets.
The comment came in response to Anthropic’s disclosure that it had started giving an outside nonprofit, METR, access to its systems similar to what its own employees have. METR evaluates frontier artificial intelligence models and has worked with several of the largest labs. Anthropic said the organization will publish its findings.
Anthropic said the arrangement follows a commitment made in September to subject its models and safety practices to stronger third-party review, and that additional evaluation partnerships are in progress. The company has argued that outside reviewers stationed inside the organization will catch problems that internal teams miss or choose not to escalate.
Musk’s suggestion raises the question of what the equivalent looks like for a company whose products are physical. An AI model can be copied, tested and run in a sandbox by a third party. A launch vehicle cannot, and the review of a rocket program already runs through a different set of institutions: the Federal Aviation Administration licenses launches, NASA oversees the programs it pays for, and the Space Force certifies the vehicles it buys.
That existing apparatus is one reason the comparison is imperfect. SpaceX operates under rules written for aviation and spaceflight, with licensing requirements, mishap investigations and public filings when something goes wrong. An independent safety reviewer for a rocket company would sit alongside regulators rather than replacing them.
Where the parallel is stronger is in the parts of the company that look like software. Ground systems, flight software, autonomous collision avoidance and the satellite network’s traffic management are all code-driven, and failures there have consequences measured in spacecraft rather than in customer complaints. Those are the areas where an outside evaluator with staff access could look at something a regulator does not examine.
The commercial motive is also worth noting. Companies that sell to governments increasingly have to document how they handle risk, and a letter from an outside evaluator is a document that travels well through procurement. For a launch provider bidding on national security missions, the ability to point at an independent assessment is a credential, not just a precaution.
There is a cost. An outside reviewer with employee-level access inside a company with classified contracts must be carved out of the compartments it cannot enter, which is exactly the sort of boundary that makes an evaluation less informative than it appears. The distinction between access to a company and access to the work that matters is the one regulators have learned to press on.
The broader shift is the pace at which evaluations have moved from voluntary to expected. Two years ago, third-party testing of frontier AI models was a research exercise conducted with cooperative labs. It is now part of the disclosure a lab makes when it releases a model, and regulators at the Federal Trade Commission have demanded records on safety incidents and protective measures from the largest developers.
Musk’s companies sit on both sides of that line. His AI lab competes with the companies submitting to review, and his rocket and satellite businesses sell to the government that is asking for the reviews. A pledge to accept outside evaluation would serve both positions, and it would cost less in an organization whose products are already subject to licensing than in one whose products are not.
Analysts who follow the industry said the practical test is a name, a scope and a publication. A commitment to evaluation means little until an organization is identified, its access is described and its findings are released without the company’s approval. Anthropic has said METR will publish, which sets a comparison for anyone who follows.
What remains unanswered is what a review of a launch company would examine first. Failure modes in rocketry are known, measured and documented by regulators, while the failure modes in autonomous systems that coordinate thousands of satellites are newer and less studied. An evaluator with access to the second category would be doing work nobody has done in public.
For now the statement is a signal about direction rather than a program. SpaceX has not described a scope, named a reviewer or set a date, and Musk’s comments on X have often preceded a product rather than described one that exists. Whether a review follows will be visible in a document, not in a post.
The internal cost of a review is another factor companies weigh. Employees who know that an outside party can read their incident reports write differently, and engineering teams that expect inspection build evidence trails into their processes. Both effects are the point of the exercise, and both require a company to accept that some honest failures will become visible to someone outside the building.
Regulators have also learned to ask for the scope rather than the intention. A commitment to outside review becomes meaningful when it names who is reviewing, what they can see and what will be published, and the practice in the sector is now to hold companies to those three items before treating a statement as a program.
