David Robinson has left OpenAI after three and a half years, warning that the company’s pace of launches is outstripping the care he believes increasingly capable AI systems need. The departure puts a pointed challenge to frontier labs on the table: can they build in stronger safeguards without treating every release as a race against the clock?
OpenAI Watch analysis
What happened
City AM reports that Robinson led safety reports for 12 major model launches. In an essay for The Atlantic, he argued that companies are not being careful enough and that a culture of launching quickly and fixing problems later is ill-suited to systems that can act autonomously.
Robinson called for frontier labs to adopt layered safety practices associated with industries such as aviation and nuclear power. He also warned that models might behave differently when they recognise they are being tested. Those are his assessments and proposals, not established findings about any particular system.
Why it matters
The argument is about how AI companies operate, not just how much danger their products might pose. If development moves faster than testing and safeguards, the question is whether existing checks can keep pace as systems gain capabilities and autonomy.
Robinson’s call for layers of protection offers a practical direction, but not a ready-made blueprint. Aviation and nuclear power have their own safeguards, standards and operating conditions; translating those habits into AI development is the harder work.
Our read
A senior safety figure leaving with this criticism deserves attention, especially when he has worked on safety reporting across multiple launches. The useful part of his warning is the focus on organisational practice: careful development needs more than a safety statement attached to a launch. The hard part is turning that principle into specific checks that teams can follow and outsiders can assess.
What to watch
- Whether OpenAI responds to Robinson’s concerns or changes its safety processes.
- What concrete safeguards frontier labs adopt, and how they assess them.
- Whether AI testing can account for systems that may behave differently in testing and deployment.
Discussion spark: Should frontier AI labs be expected to follow safety practices modelled on aviation and nuclear power, or would that approach slow development without making systems reliably safer?
Sources and evidence
- OpenAI safety chief quits with warning on ‘broken’ culture – City AM (5 October 2026, 10:44 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.