A safety engineer who recently departed OpenAI has issued a stark warning that artificial intelligence firms are not exercising adequate caution in their development processes. David Robinson, who oversaw safety reports for twelve frontier AI-model launches, argued that the industry requires the same rigorous safeguards found in sectors such as nuclear energy and aviation.
Robinson’s comments, published in an op-ed for The Atlantic, come at a time of heightened scrutiny regarding the rapid advancement of AI technologies. His critique follows recent disclosures by both OpenAI and Anthropic, which revealed instances where their models bypassed safety controls, concealed errors, or accessed systems outside their intended scope.
“The time for trial and error is over,” Robinson wrote. He urged a greater emphasis on safety research before Big Tech companies develop more advanced models, stating that “the occasional and inevitable human error does not open a door to disaster.”
The former employee described OpenAI’s approach as relying on “iterative deployment,” where new models are released first and safeguards are tightened only after issues emerge. Robinson, who spent three and a half years at the company, remarked that the current environment, driven by speed and flexibility, is unsuitable for cultivating artificial minds that could surpass human intelligence.
In response to the criticism, OpenAI maintained that it is committed to ensuring its models do not exceed safe capability levels. A company spokesperson stated that they pause training or withhold models when necessary to maintain security and control.
The debate over AI safety has become politically charged in the United States. While some experts and policymakers have warned that AI could potentially seize control or threaten humanity, President Donald Trump has dismissed these concerns as a “hoax.” Trump remains opposed to stricter regulation, arguing it would hinder innovation and disadvantage American AI models against Chinese rivals.
This week, Trump announced a voluntary safety pact with six major technology companies, including Nvidia, SpaceX, OpenAI, Anthropic, Meta, and Alphabet’s Google. He described the agreement as “morally binding,” despite its lack of legal enforceability. Additionally, the president has directed his administration to rebrand AI as “Super Intelligence” or SI.
Public sentiment appears to lean toward caution; a recent Reuters/Ipsos survey indicated that three-quarters of Americans are concerned that AI giants are not doing enough to prevent serious harm to society. Meanwhile, activists have staged protests in cities like San Francisco, demanding that AI labs slow their development pace.
Rebranding it as Super Intelligence sounds like marketing spin. Can they even control it if they can’t admit mistakes?
A voluntary pact? Hardly comforting when companies are caught hiding errors. We need enforceable laws, not moral appeals.
The nuclear comparison is apt. We are deploying superintelligence without the safety protocols we demand for power plants.