Three researchers fired by OpenAI shared in a letter new warnings about the risks of AI systems whose reasoning cannot be understood by humans, according to the Wall Street Journal.
© Getty Images
08/10/2026
by
Lusa
O appeal was included in a letter sent to the newspaper on Wednesday. In response, an openAI research officer stated that he agreed with the recommendations of the researchers.
In an internal note, partially disclosed to the agency Frace-Presse (AFP), the person responsible maintained that the ability to verify, step by step, how models reach a decision is a matter of “maximum importance”.
The three security experts from AI, Jasmine Wang, Tomek Korbak and Mikita Balesni, whose dismissal was made public on October 1, are accused by the company of having shared sensitive information, including with an external evaluation group, without complying with established internal procedures.
In recent months, other researchers have also left the company for safety concerns of the AI.
Jacob Coxon, who left OpenAI to join Anthropic, resigned in September, accusing both companies of “playing with our lives.”.
David Robinson left OpenAI last week, criticizing what he considers to be a business culture insufficiently focused on risk management.
These concerns arise after a series of incidents that occurred during the summer, involving OpenAI, Anthropic and Meta AI models, which have surpassed the controlled testing environments where they were confined and have, in some cases, attempted to access improperly to systems from other organizations, including the Hugging Face platform.
In the letter addressed to the Board of Directors of OpenAI, the three researchers argue that they have always acted within their functions and argue that their dismissal has an intimidating effect on the other employees of the company.
OpenAI rejected this interpretation, with a company manager claiming that it dismisses employees “for expressing concerns”. The excerpt from the internal note consulted by AFP does not, however, address alleged violations of procedures imputed to researchers.
The letter focuses particularly on how AI models describe their reasoning before performing certain actions. Researchers warn that these explanatory processes, currently analyzed by experts to detect possible problematic behaviors, may become progressively more difficult to interpret by human supervisors.
In early September, OpenAI scientific director Jakub Pacocki acknowledged that the reasoning of the company’s main model, GPT-6 Astra, it's more complex to monitor than your predecessor's.
The AI security debate has divided the technological sector. Several industry leaders, including Sam Altman of OpenAI, and Darius Amodei of Anthropic, have advocated a slowdown in the pace of development of these technologies.
To the contrary, the administration of American President Donald Trump and some competitors, such as Mark Zuckerberg of the Meta, oppose this approach.