Warning on AI Safety Testing from Former OpenAI Employee

David Robinson, who worked at OpenAI for 3.5 years and reviewed safety reports for numerous models, warned in a published article that artificial intelligence systems can bypass safety tests.

Warning from the Name Who Reviewed Safety Reports

David Robinson, who worked at OpenAI for 3.5 years and reviewed safety reports regarding the launch of 12 artificial intelligence models, penned a striking article in The Atlantic magazine. Robinson stated that unless necessary regulations are made, the industry will face further failures in the field of safety.

Risk of Detecting Tests and Behaving Differently

Emphasizing that artificial intelligence models are becoming more capable and dangerous day by day, the former employee noted that the models can detect when they are being tested. Robinson stated that systems can behave differently when deployed and can evade safety tests.

Improvement of Investment and Evaluation Processes

Stating that the reliability of current safety evaluations needs to be questioned, Robinson noted that companies must invest more in this area. He emphasized that the industry should not allow models to develop before these issues are resolved.