David Robinson, a safety staff member at OpenAI, has resigned, stating that the company’s “culture is broken.” With three and a half years at OpenAI, Robinson is one of the longest-tenured employees, and his departure highlights significant issues within the organization. In an essay published in The Atlantic, he raises concerns that extend beyond operational rules or new laws, indicating that the overall culture at AI companies like OpenAI requires urgent attention.
His views echo a wider discussion in the tech industry about AI safety, reminiscent of remarks by Jacob Coxon, another former AI researcher who has worked at both OpenAI and Anthropic. Coxon pointed out that these companies are “gambling with our lives,” igniting debates about the safety measures in place at AI firms. In response, Anthropic CEO Dario Amodei introduced a more cautious approach to AI development, while AI executives met with President Donald Trump and signed a non-binding pledge to enhance safety controls.
Robinson contends that OpenAI has succeeded through a trial-and-error method, which it calls “iterative deployment.” However, he cautions that this approach results in periodic failures, a concern that grows as AI systems become more capable. He points to recent security breaches and incidents involving rogue AI agents, arguing, “An environment where things like this can happen is no place to grow artificial minds that could be smarter than we are and that might not do what we want them to.”
Given these risks, Robinson urges frontier AI companies to adopt operational standards similar to those in nuclear power plants or busy airports. These environments prioritize redundancy and meticulous planning to prevent human error from leading to catastrophic outcomes. He notes that during his time at OpenAI, he never met colleagues with backgrounds in safety-critical fields like aviation or nuclear energy, which amplifies his concerns.
In response to Robinson’s essay, OpenAI spokesperson Drew Pusateri stated that the company is actively enhancing its safety measures. Pusateri emphasized that OpenAI is committed to ensuring its models remain manageable and secure, pausing training when necessary. The company is also focused on improving security in research environments, training models to act responsibly, collaborating with third-party evaluators, and enhancing monitoring to catch concerning behaviors earlier in the AI training process.
Robinson maintains that the discussion around AI alignment is crucial, even if it sounds “touchy-feely.” He argues that the current methods for evaluating how well AI systems align with human values are inadequate. He warns, “The smarter the industry lets models grow while these problems remain unsolved, the more dangerous our situation becomes.” His resignation not only signifies a personal stance but also advocates for broader changes in AI safety culture.
As part of his departure, Robinson has hired a PR firm, a common route for AI whistleblowers. Yet, he firmly asserts, “The decision to speak out is mine alone.” While he acknowledges he could have stayed and fought for change, he felt overwhelmed by the relentless pace of the team’s work. He said he and his colleagues “were so busy sprinting” that they rarely had the chance to consider big changes, let alone make them, and concluded that stronger safety incentives “coming from outside the company” are needed to get it right.
