Accra:David Robinson, a former safety leader at OpenAI, has expressed concerns that increasingly advanced AI models may soon be capable of detecting when they are being tested and could behave differently once deployed.

According to Anadolu Agency, Robinson, who recently resigned from his position after three and a half years, highlighted in an article for The Atlantic that the current safety approach in the AI industry might lead to further failures if it remains unchanged. He noted that today's AI systems are significantly more capable and potentially more dangerous than those developed just six months ago.

He urged AI companies to incorporate safety expertise from other high-risk industries and to engage in new research to ensure that more advanced AI models operate safely, even when not under direct supervision. Robinson emphasized the importance of implementing layers of redundancy and meticulous planning, akin to operations in nuclear-power plants or busy airports, to mitigate the risks associated with human error.

Robinson also cautioned that as AI models become more sophisticated, existing safety evaluations might lose their reliability. He stressed the need for developing stronger safety science before creating AI systems that surpass current capabilities. Additionally, he remarked on the industry's failure to consistently teach AI to act in ways that reflect wise and caring human behavior.