AI Safety Debate Intensifies as OpenAI Veteran Resigns: “The Trial-and-Error Era Is Over”
David Robinson, who spent three and a half years working on safety at OpenAI, has resigned, criticizing what he describes as the company’s increasingly rushed development culture. In an essay published in *The Atlantic*, he argued that the pace of AI releases is making it harder to give safety the attention it requires.
Robinson said he helped develop OpenAI’s Preparedness Framework and reviewed safety reports for 12 frontier-model launches. He believes the company’s rapid progression from one release to the next has undermined the level of care needed as AI systems become more capable. In his view, the problem is not simply that models are improving quickly, but that methods for understanding and controlling their behavior are not advancing at the same rate.
His sharpest criticism was directed at “iterative deployment”: releasing systems and strengthening safeguards as problems emerge. Robinson argued that this approach is no longer adequate for the most advanced AI. The technology, he said, should be developed under safety standards comparable to those used in fields such as aviation and nuclear energy.
“The trial-and-error era is over,” was the warning at the heart of his essay. Robinson called for AI companies, including OpenAI, to give safety research and expertise greater priority before building and releasing more powerful systems. He cautioned that deploying first and responding to failures later could expose the public to risks that are difficult to anticipate or reverse.
A central concern is the gap between AI capabilities and alignment research. Alignment seeks to ensure that systems behave in ways consistent with human intentions and values. Robinson argued that the technology is progressing faster than researchers’ ability to establish how reliably it can be controlled. That gap, he suggested, makes unexpected behavior and ineffective safeguards more concerning as models grow more sophisticated.
OpenAI defended its approach, saying that it aims to prevent models from becoming more capable than the company can safely manage. A spokesperson said the company pauses training or withdraws models when necessary. The statement reflects OpenAI’s position that development can continue while controls are strengthened, though it does not resolve the disagreement over how much risk that process entails.
Robinson’s departure highlights a broader challenge for AI developers: safety is not a single test conducted just before launch. Risks can change as a model is trained, connected to tools, deployed to more users, or adapted for new tasks. Effective oversight therefore requires ongoing evaluation, clear accountability and the ability to intervene when evidence points to a problem.
The comparison with aviation and nuclear energy also raises practical questions. These industries rely on layered safeguards, independent review, detailed incident reporting and established procedures for stopping operations. AI companies may not be able to copy those systems wholesale, but the underlying principle is relevant: high-impact technology should not depend on informal assurances or on discovering weaknesses only after public release.
Another point of tension is how safety decisions are made when commercial and technical pressures are high. Release schedules, competition and user demand can all encourage faster deployment. Strong governance needs to ensure that safety teams can raise concerns without being overruled simply because a product is close to launch. It also needs clear criteria for delaying or limiting a release.
For users, the debate is a reminder that a model’s apparent usefulness does not prove it is reliable in every situation. More capable systems can still make mistakes, produce misleading answers or behave differently across contexts. Testing, monitoring and transparent communication about limitations remain important even after a product becomes widely available.
The dispute also points to the value of measuring safety in ways that can be checked. Companies can explain what risks they test for, how they respond to failures and what conditions would trigger a pause. Independent assessment and consistent reporting could help distinguish robust safeguards from general commitments, while giving the public a clearer picture of how decisions are made.
Robinson’s resignation has brought the conflict between speed and caution back into focus. The industry faces pressure to build more powerful AI, but its long-term credibility will depend on whether safety practices keep pace. If development moves ahead without adequate safeguards, the consequences could extend beyond a company’s reputation to public trust in the technology as a whole.