The policy divide had been forming well before Tuesday’s White House lunch. Meta’s Mark Zuckerberg and Nvidia’s Jensen Huang had pushed a lighter-touch approach, while Anthropic’s Dario Amodei and OpenAI’s Sam Altman had called for a stronger government role and a slower pace of development.¹

At the meeting, President Donald Trump gathered leaders from Google, Meta, xAI, Nvidia, Anthropic and OpenAI around the new White House Accord on Super Intelligence—while also seeking to rebrand artificial intelligence as “super intelligence.” The two-page pledge rests on four layers of company-led oversight: internal monitoring of frontier models, internal safety teams, outside evaluators and an independent board committee. Trump described the commitment as “morally binding.”²

That design is the accord’s central promise—and its central vulnerability. Zuckerberg called it “a significant positive step,” arguing that robust internal controls and multiple audits should give the public more confidence in what major labs build.Mark Zuckerberg (finkd)³ Supporters say the agreement creates a common baseline quickly, without freezing an industry competing globally.

Critics counter that the apparent checks remain inside a system controlled by the companies themselves. The accord does not define what counts as sufficiently “robust” controls or safe “intended” behavior; firms can select and pay their evaluators, while board committees remain corporate bodies rather than public regulators. Its language offers no legal penalty for noncompliance and no requirement for public disclosure of audit results.⁴

The document leaves open the possibility that its steps could eventually be written into law. For now, however, Washington’s answer to escalating concern over powerful models is an industry compact: a safety framework whose credibility will depend on the very companies it asks to restrain themselves.