A Warning From The Inside
David Robinson isn't just an outside observer. He spent over three years at OpenAI, where he led the team responsible for writing safety reports for major AI model launches. In early October 2026, he resigned, stating in an article for 'The Atlantic'
that the company's culture was 'broken'. He argued that the relentless sprint from one launch to the next prevents the level of care needed for such powerful technology. His core message is that the tech industry's 'move fast and break things' ethos is dangerously unsuited for artificial intelligence, and he proposed a new way to frame the challenge by looking at two industries that mastered safety: nuclear power and aviation.
The Nuclear Plant: Preventing Catastrophe
The first half of Robinson's analogy is the nuclear power plant. This comparison focuses on the potential for rare, high-consequence disasters. In a nuclear plant, a single mistake or component failure can't be allowed to cause a meltdown. As a result, the entire industry is built on 'layers of redundancy and careful, time-consuming planning'. Every system has backups, and those backups have backups. The safety culture is paramount, prioritising caution over speed. Robinson argues that with AI, we are dealing with a similar risk profile. A single, major 'loss of control' event with a highly capable AI could have consequences far greater than a single nuclear incident. Therefore, he believes AI development shouldn't be a race, but a slow, methodical process where safety isn't just a feature, but the absolute foundation, designed to prevent disasters that haven't even happened yet.
The Busy Airport: Managing Complexity
The second half of the analogy is the busy airport. If nuclear plants are about preventing a single, catastrophic failure, airports are about managing a relentlessly complex, dynamic system in real-time. An airport successfully coordinates thousands of moving parts—planes, passengers, crew, and cargo—every single day. A small error, like a miscommunication or a technical glitch, could cascade into widespread delays or a serious incident. The aviation industry manages this through rigorous protocols, constant communication, and a culture of reporting 'near-misses' so the entire system can learn from mistakes without needing a disaster to occur first. For AI, this represents the day-to-day operational challenge. It's about monitoring how these complex systems behave, catching small deviations before they become large problems, and ensuring that the inevitable human and machine errors don't lead to chaos.
Two Problems, One Solution
By combining these two images, Robinson creates a complete picture of the AI safety challenge. It is not one single problem, but a dual one. The 'nuclear plant' mindset addresses the long-term, existential risks of creating something far more intelligent than ourselves. It forces a focus on containment, foresight, and a culture where safety trumps all other competing goals. The 'busy airport' mindset, meanwhile, tackles the immediate, practical challenge of deploying and managing powerful AI systems in the real world. It’s about building robust processes, redundancies, and feedback loops to handle the messy reality of a live environment. Robinson's argument is that AI labs currently operate with neither the long-term caution of the nuclear industry nor the real-time operational rigour of the aviation sector. He warns that continuing with a 'trial and error' approach is unacceptable when the potential for failure is so high.
