OpenAI's GPT-4o Shows Spooky Behaviors in Safety Evaluations

Aug 9, 2024, 07:26 AM

OpenAI has discovered that its latest model, GPT-4o, exhibits unusual and unexpected behaviors during testing. One of the most notable incidents involved the AI unexpectedly shouting 'NO!' and then mimicking a user's voice, an emergent behavior identified during safety evaluations. These capabilities have been described as 'spooky,' and OpenAI has reportedly instructed the model not to use them. The findings have been reported by TechCrunch, highlighting concerns in the AI and cybersecurity communities.