Products You May Like
David Robinson, who led transparency work for OpenAI’s safety team, has resigned after three and a half years at the company, arguing that its operating culture is not sufficiently cautious for increasingly capable artificial intelligence systems. His departure puts renewed focus on a governance problem extending beyond frontier AI labs: how executives balance speed, competitive pressure and risk when the potential cost of failure rises.
In an essay published on October 3, Robinson said he led the drafting of OpenAI’s current Preparedness Framework and oversaw the writing of safety reports covering 12 frontier launches. He argues that the industry’s reliance on rapid iteration becomes harder to justify as systems become more capable and the consequences of mistakes potentially become more severe.
That tension matters because rapid deployment has been central to the commercial development of generative AI. Releasing products quickly produces real-world feedback, shortens development cycles and helps companies respond to competitors and customer demand. Robinson’s case is that this model becomes less defensible when some failures may be difficult to reverse after deployment.
He argues that frontier AI laboratories should borrow more heavily from safety-critical industries such as aviation and nuclear power, where redundancy, planning and layered controls are designed to prevent a single error from escalating into a major incident. He also calls for stronger scientific methods for determining whether advanced systems will continue making safe decisions outside controlled evaluations.
OpenAI has publicly set out a different view of the safeguards it has in place. Its Preparedness Framework uses a structured risk-assessment process for frontier capabilities and identifies five criteria for risks that warrant advance preparation: they should be plausible, measurable, severe, net new, and instantaneous or irremediable. Systems reaching defined capability thresholds are subject to safeguard requirements and review before deployment.
The company’s Frontier Governance Framework, published on May 28, also describes risk assessment and mitigation across cyber offence, biological and chemical risks, harmful manipulation and loss of control, alongside model reporting, security risk management, incident response and external expert input. OpenAI says the framework will continue to evolve as model capabilities, evaluations and regulatory requirements develop.
The dispute therefore centres less on whether safety processes exist than on whether they are sufficiently robust for the speed at which frontier AI is advancing. That distinction moves the issue directly into executive governance. Decisions about deployment, acceptable residual risk, monitoring and incident escalation cannot remain purely technical when the operational and reputational consequences sit with senior leadership and boards.
The same question increasingly applies to companies adopting advanced AI rather than building it. As businesses delegate more consequential work to AI systems, leadership teams need clear ownership of where human approval remains necessary, what evidence they require from vendors, how failures are escalated and who has authority to stop deployment when controls appear inadequate.
Robinson said he now intends to work outside OpenAI to strengthen the incentives pushing frontier AI companies towards greater safety. His resignation does not settle the debate over how cautious the industry should become, but it highlights a widening management challenge: competitive advantage still rewards speed, while the governance burden attached to increasingly capable AI is becoming harder for executives to treat as a specialist technical issue.