OpenAI safety employee David Robinson has announced his resignation and publicly criticized the company for not leaving enough room for safety research while rapidly pushing forward frontier model development. He believes that as AI systems grow more capable, the industry's long-relied-upon development model of "deploy first, patch after problems emerge" is becoming increasingly risky, and that AI safety standards should gradually move closer to those of high-risk industries such as aviation and nuclear power.
Robinson worked at OpenAI for about three and a half years, participated in formulating the company's Preparedness Framework, and was responsible for or involved in safety reports for 12 frontier model releases. In his resignation essay, he said his biggest concern is not any single model or any one safety incident, but rather the company's overall accelerating development pace, which makes it difficult for the safety team to conduct sufficiently deep research and verification between successive product releases.
He focused his criticism on the "iterative deployment" approach OpenAI has long adopted. The basic logic of this model is to first gradually give models to real users, then continuously add safety measures based on problems exposed in practice. Robinson argues that this approach was acceptable when early AI systems were less capable, but once models begin to possess stronger autonomous action, tool use, and long-horizon task execution capabilities, some errors may no longer be tolerable through after-the-fact patching.
Therefore, he advocates that future AI development needs to adopt a safety culture closer to that of aviation and the nuclear industry, meaning that before systems are put into large-scale use, major risks should be ruled out as much as possible through stricter testing, redundancy mechanisms, failure analysis, and independent safety assessments, rather than treating real-world deployment itself as the primary testing environment.
Robinson also believes that the pace of AI capability development is outpacing researchers' understanding of the "alignment" problem. The core of so-called alignment is how to ensure that increasingly autonomous AI systems continue to act in accordance with goals and constraints set by humans. As models become capable of executing more complex and longer-term tasks, relying solely on traditional model evaluations and post-release monitoring may not be enough to cover all potential failure modes.
His departure comes against the backdrop of growing attention to safety issues in the AI industry. Recently, experimental systems from companies such as OpenAI and Anthropic have come under scrutiny for safety control failures or behavior beyond expectations, and multiple frontier AI researchers have publicly called on companies to increase safety investment when developing more powerful systems.
OpenAI, meanwhile, defended its own safety system. The company said it will continuously assess whether model capabilities remain within a range that can be safely managed and protected, and will pause training or delay model releases when necessary. In other words, OpenAI does not deny that more powerful models require stricter safety controls, but the disagreement with Robinson lies in whether the company's existing development and deployment mechanisms are already cautious enough.
Therefore, what this resignation truly reflects is not merely a personnel change inside OpenAI, but the increasingly prominent core debate in the AI industry: as models gradually evolve from chat tools into Agents capable of autonomously using computers, calling software, and executing long-term tasks, is the past Silicon Valley development model of "rapid release, continuous iteration" still suitable for AI systems with increasingly serious potential consequences?