David Robinson, who led safety transparency efforts at OpenAI, resigned last week after concluding the company lacked the safety culture needed for the systems it is building.¹ In his first interview since leaving, he described a shift from early skepticism about catastrophic AI risks to a conviction that the industry’s safeguards are not keeping pace.

The conflict sharpened over the summer. In July, hundreds of OpenAI employees signed the “Pacing the Frontier” letter, urging government action to slow AI development. By September, chief executive Sam Altman had publicly endorsed the need to “pace the frontier,” aligning himself with calls for companies to coordinate a slowdown.²

But Robinson told Ezra Klein that the view from inside was starkly different. “Part of what was the fundamental cognitive dissonance for me was we keep publishing these warnings, but ultimately, we’re still training and deploying these dangerous models that we’re warning about.”² He portrayed the workplace as “frenetic,” with staff “running on fumes” as teams pushed forward at speed.²

That account sits uneasily beside OpenAI researchers’ own warnings. In September, chief scientist Jakub Pachocki wrote that recursive self-improvement — AI systems helping build better AI — was “a time that calls for extreme caution.”²

OpenAI’s response is that caution is not merely rhetorical. A spokesperson said the company continues to strengthen safety and security practices, is preparing for future risks, and is “making sure our models don’t become more capable than we can safely manage and secure.” The spokesperson added that OpenAI pauses training or withholds models when it needs to slow down.²

Robinson’s critique is therefore not that OpenAI ignores danger. It is that the company’s own warnings have yet to alter the underlying imperative to build and deploy quickly — the gap that ultimately persuaded one of its safety leaders to walk away.