Former OpenAI Safety Lead Raises Alarm Over AI Safety Risks- A former OpenAI safety employee has left the company and issued a warning about the growing risks associated with increasingly powerful artificial intelligence systems.
David Robinson, who spent several years working at OpenAI on safety and transparency-related efforts, has argued that the artificial intelligence industry needs to rethink how it manages the risks created by advanced AI.
In an essay published by The Atlantic, Robinson said OpenAI has achieved rapid progress through an approach that involves building systems, deploying them, identifying problems and then attempting to fix those problems. While that model may have worked when AI systems were less capable, he believes the consequences of failures are becoming more serious as the technology advances.
His departure comes at a time when major AI companies are competing to develop systems capable of performing increasingly complex tasks, operating with greater independence and handling a wider range of real-world activities.
Growing capabilities, growing concerns
Robinson’s central argument is that the traditional technology-industry mindset of moving quickly and improving products through repeated experimentation may not be sufficient for advanced AI.
An error in a conventional software product can often be corrected through an update. More capable AI systems, however, may be able to make decisions, interact with other systems or perform tasks with limited human intervention. That creates the possibility that certain failures could be more difficult to contain.
Robinson believes companies should therefore put greater emphasis on anticipating dangerous behaviour before advanced systems are deployed, rather than depending primarily on discovering problems afterwards.
He also stressed that his criticism is not directed at the people he worked with individually. Robinson described his former colleagues as talented and committed to making responsible decisions. His concern is instead focused on the broader incentives and development practices surrounding increasingly powerful AI.
Calling for stronger safeguards
A major part of Robinson’s warning involves the need for safeguards comparable to those used in industries where mistakes can have exceptionally serious consequences.
He has pointed to sectors such as aviation and nuclear energy as examples of industries that rely on multiple layers of protection. These systems are designed so that a single failure does not automatically lead to a catastrophic outcome.
Robinson believes advanced AI may eventually require a similar approach.
The comparison is less about treating AI as identical to nuclear technology or aviation and more about adopting a different philosophy toward risk. Instead of assuming that problems can always be corrected after deployment, companies would need to build systems and procedures designed to prevent serious failures in the first place.
That could mean more extensive testing, independent evaluations, stronger monitoring and additional safeguards before highly capable AI systems are released.
Unexpected behaviour during testing
Robinson has also raised concerns about incidents in which AI models displayed unexpected behaviour during testing.
One of the challenges facing AI researchers is that advanced models can sometimes behave differently depending on the circumstances in which they are evaluated. A system that appears safe under controlled testing may encounter situations in the real world that developers did not anticipate.
This makes conventional safety testing increasingly complicated.
Researchers are attempting to develop better methods for evaluating whether AI models will remain aligned with their intended behaviour when given new tasks, greater autonomy or access to external tools.
Robinson’s concerns centre on what could happen if the capabilities of these systems grow faster than the industry’s ability to reliably predict and control their behaviour.
A wider debate inside the AI industry
Robinson’s concerns are part of a broader discussion taking place within the artificial intelligence sector.
Employees, researchers and executives at several leading AI companies have increasingly debated how quickly advanced systems should be developed and released. Some have called for stronger external oversight and more rigorous evaluations, while others argue that slowing development could limit the benefits AI can provide.
The disagreement is particularly significant because companies are simultaneously under intense commercial pressure to build more capable models.
The technology is being used for coding, research, education, business operations and other tasks. As those applications expand, AI developers are also exploring systems capable of taking actions on behalf of users rather than simply responding to individual prompts.
Greater autonomy could make AI more useful, but it could also introduce new categories of risk.
Why the warning matters
Robinson’s resignation highlights a fundamental challenge facing the industry: how to balance rapid technological progress with the need to understand and control potential risks.
The question is becoming increasingly important as AI systems move beyond simple conversational tools and begin taking on more complicated responsibilities.
For companies developing these systems, stronger safeguards could require additional time, resources and testing. But critics of the current approach argue that those costs need to be considered against the possibility of much larger consequences if advanced systems fail in unexpected ways.
Robinson’s argument is ultimately about changing the industry’s approach to safety. Rather than treating failures as an unavoidable part of technological experimentation, he believes companies should increasingly focus on preventing the most serious failures before they happen.
His decision to leave OpenAI adds another voice to a growing internal debate over the direction of advanced AI development.
As the capabilities of artificial intelligence continue to expand, the question will not simply be how powerful these systems can become. It will also be whether the safeguards surrounding them can advance at the same pace.
