On September 29, the White House gathered representatives from Anthropic, Google, Meta, Nvidia, OpenAI and SpaceX/xAI. The companies promised clearer internal controls, reviews by independent outsiders and a board committee to follow up on the results.
This may sound remote, but it comes down to a simple question: Who checks that a powerful AI service works as the company claims? External auditors can examine areas such as safety, protection against intrusion and whether risk procedures are actually followed.
The agreement is voluntary, however. It includes no fines or other penalties of its own if a company fails to keep its promise. It is therefore not the same as a new law or independent government oversight.
Voluntary does not have to mean worthless. Audits can improve transparency and create shared working practices. But their value depends on what reviewers are allowed to see, whether results become public and what happens when shortcomings are found.
A useful question to keep in mind when an AI company talks about safety is: Who checked the claim – and what were the rest of us told?