A former OpenAI safety employee has publicly criticized the company’s approach to artificial intelligence development after quitting the company, arguing that its rapid launch cycle is creating risks that cannot be adequately addressed through safeguards added after systems are deployed.
David Robinson, who spent three and a half years at OpenAI and worked on the company’s preparedness framework, said the company and the wider AI industry are not being “nearly careful enough” as they develop increasingly capable models.
Writing in The Atlantic under the headline “I Quit OpenAI Because Its Culture Is Broken,” Robinson said that advanced AI development requires a fundamentally different approach to safety, with greater reliance on specialized expertise and research before increasingly powerful systems are built and deployed.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
“The time for trial and error is over,” Robinson wrote, arguing that AI development should adopt safety standards more comparable to those used in high-risk industries such as nuclear power and aviation.
His criticism comes as leading AI companies face growing pressure over whether their ability to develop more capable systems is advancing faster than their ability to understand their behavior, identify failure modes, and build reliable safeguards.
Robinson said he helped draft OpenAI’s preparedness framework and oversaw safety reports for 12 frontier-model launches during his time at the company. His criticism therefore centers not only on AI safety as a general concern but on the internal processes used to evaluate increasingly capable models before release.
“As the company sprints from one launch to the next, it is failing to achieve the level of care that I believe is needed,” Robinson wrote.
The Limits of “Iterative Deployment”
A central target of Robinson’s criticism is OpenAI’s philosophy of “iterative deployment,” under which models are released, and safeguards are strengthened as problems emerge in real-world use.
The approach has an obvious advantage: deploying systems provides companies with information about how models behave outside controlled testing environments and allows developers to identify problems that may not appear during laboratory evaluations.
But Robinson argues that the approach becomes increasingly dangerous as AI systems become more capable and the consequences of unexpected behavior become harder to contain. His concern is that companies may be treating deployment itself as part of the safety-testing process at a point when the systems involved could have capabilities that make failures substantially more consequential.
The argument goes to a fundamental tension in the AI industry. Companies developing frontier models need real-world feedback to improve their systems, but they also face pressure to ensure that more capable models cannot exploit weaknesses in their safeguards, access unauthorized systems, or behave in ways their developers did not anticipate.
Robinson said AI capabilities are advancing faster than researchers’ understanding of alignment, the field concerned with ensuring that AI systems behave in accordance with human goals and values. That gap is growing wider as AI models move beyond generating text and images toward operating software, using tools, making decisions, and performing longer sequences of tasks with less direct human supervision.
A model that produces an incorrect answer can generally be corrected by a user. A more autonomous system that can interact with external systems presents a different category of risk if it misinterprets an objective, circumvents a restriction, or behaves in an unintended way.
Robinson’s argument is less about whether AI companies should stop developing new models and more about whether their safety processes are advancing quickly enough to match the capabilities they are creating.
OpenAI Says It Can Hold Back Models When Necessary
OpenAI disputed the suggestion that it is simply rushing models into deployment without sufficient safeguards.
“We’re making sure our models don’t become more capable than we can safely manage and secure, and we pause training or hold back models when we need to slow down,” an OpenAI spokesperson said.
That response points to a key difference between the company’s stated safety framework and Robinson’s criticism. OpenAI says development is constrained by its assessment of whether a model can be safely managed, while Robinson argues that the company’s development culture itself is making it difficult to reach the level of caution required.
The disagreement also comes amid a broader debate within the AI industry over the appropriate balance between rapid innovation and precaution.
OpenAI and rival Anthropic have faced scrutiny following incidents involving safety controls and unexpected model behavior. Such episodes have increased attention on how frontier laboratories test models, monitor them after deployment, and determine when a system’s capabilities warrant additional restrictions.
For AI companies, the stakes extend beyond the technical question of whether a model is aligned in a controlled environment. Increasingly capable systems are being integrated into products and workflows where they can interact with users, software, and external data. That makes failures potentially more difficult to isolate once a model is deployed at scale.
Robinson’s warning consequently raises a question that is becoming more important as the industry moves toward autonomous AI: can safety continue to be treated primarily as an iterative process that improves alongside deployment, or should certain capability thresholds require substantially more evidence of reliability before systems are released?
His comparison with nuclear power and aviation underscores the distinction he is trying to draw. Those industries operate on the premise that some failures are too consequential to discover through ordinary trial and error in live environments.
OpenAI’s response indicates that it believes its existing safeguards, preparedness processes, and ability to pause development can manage that risk. However, Robinson’s resignation and public criticism suggest that at least some people involved in those processes believe the pace of development is making that model of safety increasingly difficult to sustain.
The disagreement is unlikely to be resolved by a single model launch or safety report. As frontier systems become more capable, the industry’s credibility will increasingly depend on whether its testing and governance mechanisms can demonstrate that safety standards are keeping pace with capability growth.
For Robinson, the central problem is that the industry is approaching that threshold too quickly.
“The time for trial and error is over,” he wrote.