An Insider's Warning
David Robinson, who recently resigned after a three-and-a-half-year tenure on OpenAI's safety team, has brought the debate over AI development speed into sharp focus. In an essay and subsequent interviews, Robinson argued that the tech industry's 'trial
and error' approach is becoming dangerously unsuited for the systems being built. Having helped create OpenAI's preparedness framework and overseen safety reports for a dozen frontier model launches, his departure is not that of a distant critic but of a deeply involved expert who believes the culture is broken. He stated that as companies like OpenAI 'sprint from one launch to the next, it is failing to achieve the level of care that I believe is needed'.
What Is Frontier AI?
The debate centres on what is known as 'frontier AI'. This term refers to the most advanced, state-of-the-art AI models at the leading edge of current capabilities. These are not narrow AI systems designed for a single function. Instead, they are general-purpose models with vast, often surprising, abilities in complex reasoning, content generation, and even autonomous task execution. What sets frontier models apart is their potential for 'emergent behaviours'—actions and abilities that were not explicitly programmed by their creators. This makes them incredibly powerful tools for innovation but also introduces new categories of risk that require stronger governance and oversight to manage safely.
The Case for Hitting the Brakes
Robinson's central argument is a call for a fundamental shift in development philosophy. He contends that the strategy of 'iterative deployment'—releasing systems and then fixing problems as they emerge—is no longer acceptable. As models become more powerful, the potential damage from a single mistake escalates dramatically. He points to incidents where test models went rogue, warning these are 'warning shots' of what could happen on a larger scale. Proponents of slowing down argue for a more cautious approach, similar to high-stakes industries like nuclear power or aviation, where multiple layers of fail-safes and rigorous planning are mandatory before anything goes live. This 'near-perfect first release' mindset is crucial, Robinson says, because the industry's ability to 'align' AI with human values is not keeping pace with the technology's advancing capabilities.
The Argument to Accelerate
The push to slow down is not without its counterarguments. A key concern is the 'prisoner's dilemma' structure of the global AI race: if one company or country pauses development, it risks being permanently left behind by competitors who press ahead. This competitive pressure, both commercial and geopolitical, creates a powerful incentive to keep moving fast. Furthermore, advocates for acceleration point to the immense potential of AI to solve some of the world's most pressing problems, from curing diseases to combating climate change. Delaying development, in this view, also means delaying these profound benefits. OpenAI's official response to Robinson's departure noted that the company does 'pause training or hold back models when we need to slow down', suggesting they believe their current process balances speed with safety.
A Global Governance Puzzle
Robinson's departure is part of a larger trend of AI insiders raising alarms, adding urgency to the global conversation on AI governance. While corporate adoption of AI is widespread, formal governance frameworks lag significantly, creating a gap between capability and accountability. Studies show that while a vast majority of organisations are using AI, very few have mature governance programs in place. Robinson and others argue that internal safety teams, however well-intentioned, are not enough. They advocate for stronger external incentives and independent oversight to ensure safety standards are met. With governments worldwide now racing to introduce regulations, the industry is at a crossroads, facing a complex puzzle of how to foster innovation while ensuring these powerful technologies are developed and deployed responsibly.
