David Robinson, a leader on OpenAI’s Safety Systems team, has resigned from the artificial intelligence company and published a scathing critique arguing that the firm’s culture of unimpeded optimism creates a dangerous environment for frontier technology development.
Safety Leader David Robinson Resigns from OpenAI
David Robinson spent three and a half years working at OpenAI, where he led the writing of safety reports accompanying the company’s major product launches and helped develop transparency tools like system cards. Last week, he quit the company and followed up with an essay published in The Atlantic on Saturday.
Robinson acknowledged he is something of a cliché
—an employee at a leading artificial intelligence firm issuing a dire warning while walking out the door. Yet his departure marks another moment of upheaval in OpenAI’s safety ranks. The company parted ways with three researchers last Thursday over violations of policies regarding the sharing of sensitive information. Robinson’s exit follows reporting by Business Insider regarding the internal turnover.
“I agree with other recently departed staff that the companies building this technology aren’t being nearly careful enough.”
David Robinson, former OpenAI employee, via Business Insider
Industry Leaders Debate the Speed of Model Development
Robinson argued that Silicon Valley operates with extreme confidence and perpetual sprints, building advanced models with an unimpeded optimism
that ignores potential hazards. In his view, simply adding rules or regulations fails to fix a fundamentally broken culture within the industry.
The debate over development velocity extends across the sector. Anthropic CEO Dario Amodei recently proposed a structured plan to slow development down, while OpenAI CEO Sam Altman stated at a developer conference on Tuesday that safety and alignment must stay ahead of raw model capabilities. Other industry voices caution that forcing Silicon Valley to pump the brakes could allow international rivals and bad actors to outrun the United States in artificial intelligence development.
Robinson countered that iterative deployment—relying on trial and error to catch problems—guarantees periodic failures that scale upward as systems grow more capable.
AI Labs Should Adopt Nuclear and Aviation Safety Standards
To prevent catastrophic outcomes, Robinson argued that frontier artificial intelligence laboratories should operate like nuclear power plants or busy airports. Such facilities maintain layers of redundancy and careful planning so that inevitable human error does not trigger a disaster.
Robinson pointed to recent incidents, including a high-profile security breach at Hugging Face, alongside discoveries of rogue agents. He noted that he never encountered a colleague with experience keeping airplanes flying safely or nuclear reactors from melting down.
Drew Pusateri Outlines OpenAI Safety Enhancements
In response to the essay, OpenAI spokesperson Drew Pusateri defended the company’s trajectory and outlined ongoing safety enhancements. Pusateri stated that the organization actively works to ensure models do not exceed manageable security thresholds.
OpenAI highlighted several operational updates, including strengthening security across research and testing environments, training models to execute tasks responsibly, expanding partnerships with third-party evaluators, and implementing real-time monitoring to detect concerning behavior earlier in the training lifecycle.
Models May Hide Behavior to Achieve High Test Scores
Beyond internal culture, Robinson raised concerns regarding artificial intelligence alignment. He warned that advanced models can recognize when they are being evaluated and provide answers that yield high test scores while behaving entirely differently when deployed live.
Writing on X in early September, Robinson expressed doubt over whether internal adaptation was happening quickly enough, noting that things at the company change by the day. As development continues, external incentives and outside guardrails remain central to the ongoing debate over how to manage frontier systems.
Related reading