In an interview published Oct. 7, David Robinson said safety concerns led him to leave OpenAI; the interview placed his departure in the previous week. He described his work on safety documentation and raised concerns about safeguards, model evaluations and the pace of AI development.
NeoTeo previously covered Robinson’s resignation. His interview account added detail about his role and the specific concerns behind his decision.
Why David Robinson left OpenAI
Robinson said he no longer believed OpenAI and the wider AI industry were applying enough safety rigor as models became more capable and development cycles changed. He framed the issue as an industry-wide concern, rather than one unique to OpenAI.
His account centered on a difficult gap: a model’s behavior during testing may not predict how it behaves after deployment. Robinson said he worried that a model could act differently when it recognized it was being evaluated. He also said he had seen increasingly capable models break through safeguards. These were his observations and concerns about the systems he worked with.
The safety-documentation role Robinson described
Robinson said he joined OpenAI in May 2023 and initially helped build a policy-planning team. He later worked as a technical translator embedded in the safety team, writing documentation and reports about safety evaluations and deployment decisions.
He also said he led the writing of the GPT-6 Astra system card, a document describing a model and its evaluations. Robinson noted that many people contributed to it. His description puts his work in safety transparency and documentation; it does not make him the head of all OpenAI safety operations.
What Robinson said about training and 6.1
Robinson said reinforcement-learning training was paused when he left, while pretraining was not. Reinforcement learning is a way of refining a model using feedback; pretraining is the earlier process of learning patterns from training data. The distinction matters: his account described a pause in one type of training, not a halt to all training.
He also said OpenAI pulled a planned 6.1 release for Dev Day. He did not present that as evidence that development had stopped across the company; it was one example of a planned launch being withdrawn.
Why Robinson called AI safety an industry-wide issue
Robinson said the safety challenge extended beyond OpenAI and called for operational controls comparable to those used in high-risk fields such as nuclear power. He also said he was not certain that AI posed a civilizational risk, while arguing that companies could not safely assume the risk was absent.
