David Robinson says he has resigned from OpenAI, arguing that the company’s development pace and safety culture leave too little room for careful planning. He says safety discussions should look beyond rules and laws to how AI companies make decisions and manage risk.
What Robinson says should change
Robinson, a former OpenAI safety employee, says he led the writing of safety reports published with each major launch. He argues that the industry’s emphasis on rapid development and iterative deployment leaves too little room for deliberation.
Iterative deployment means releasing systems and improving safeguards as problems emerge. Robinson argues that this trial-and-error approach brings periodic failures, and that their scale grows as AI systems become more capable. He has called for frontier AI labs to adopt more redundancy and careful planning, practices he associates with aviation and nuclear power.
OpenAI’s response
OpenAI spokesperson Drew Pusateri said the company pauses training or holds models back when it needs to slow down. He also said OpenAI is strengthening security in research and testing, expanding work with third-party evaluators, and improving real-time monitoring.
OpenAI’s published safety framework groups its approach under Teach, Test and Share. It lists internal evaluations, expert testing, red teaming, preparedness evaluations, system cards and feedback from real-world use.
The reported Hugging Face incident
Robinson has pointed to reported unintended AI-agent behavior, including an incident involving Hugging Face, as an example of the risks he wants the industry to take seriously.
