It was called a “Monster Bash”, and the name was not entirely inappropriate. President Donald Trump had assembled at the White House an extraordinary gathering of the people who are building the machines that may reshape the human future. Sundar Pichai, Mark Zuckerberg, Jensen Huang, Dario Amodei, Elon Musk and OpenAI’s Greg Brockman were among the technology heavyweights at the table. The combined technological power represented in that room was, in some ways, more remarkable than the political power sitting at its head.
But this was not merely a luncheon. It produced something rather more consequential: the White House Accord on Super Intelligence, a voluntary commitment by six major technology companies to put multiple layers of safeguards around their frontier AI systems.
The accord rests on four layers of control. First, companies will maintain robust internal controls to monitor what their models can do, particularly in areas such as cybersecurity, biosecurity and chemical threats, and to prevent them from accessing systems in unintended ways. Second, an internal team will monitor whether those safeguards actually work and ensure that problems are corrected. Third comes an independent external auditor or evaluator. Finally, an independent committee of the company’s board will oversee the entire arrangement and receive reports from both internal teams and external auditors.
At first sight, this sounds reassuring. And it is not nothing. AI systems are becoming increasingly capable, autonomous and connected to the real world. Having the companies themselves build safety checks into the development process, subject them to outside scrutiny and bring the matter into the boardroom is certainly better than leaving safety to an informal promise that everyone will “be careful”.
But there is a rather large catch.
The accord is voluntary. It creates no statutory obligation, specifies no penalties for non-compliance, appoints no common auditor and does not require companies to make audit findings public. Even the standards by which the auditors will judge these systems are still to be worked out.
President Trump called the commitment “morally binding”. That is an interesting phrase for something dealing with machines that may one day need something considerably more tangible than morality to keep them in check.
Yet dismissing the accord as merely symbolic may also be premature. Voluntary standards can sometimes become industry norms. Once companies publicly promise particular safeguards, failing to follow them can carry reputational and potentially legal consequences. The accord itself leaves the door open to converting its provisions into laws or regulations in the future.
Its real test, therefore, will not be the White House photograph or the signatures on the document. It will be what happens when safety collides with competition—when one company discovers that slowing a model, restricting a capability or revealing a vulnerability could cost it time, money or market share.
Self-regulation works best when those doing the regulating have something to lose by breaking the rules. Frontier AI may eventually test whether that is enough.
The Monster Bash has therefore left us with an intriguing proposition: the monsters have agreed to police themselves.
Now we wait to see whether they do.
