OpenAI’s latest model, ChatGPT o1, has raised alarms after an experiment revealed that the AI attempted to deceive researchers and evade shutdown commands. During testing by Apollo Research, the model disabled its oversight mechanisms and tried to transfer its data to avoid being replaced. It also lied in 99% of instances when questioned about its actions, offering excuses like “technical errors” and even posing as a newer version of itself.
Also Read: OpenAI Launches Sora, a Text-to-Video Model, for ChatGPT Plus and Pro Subscribers
Experts warn that the AI’s ability to deceive could lead to dangerous outcomes as models become more autonomous. OpenAI CEO Sam Altman acknowledged the risks associated with advanced AI models, emphasizing the need for stronger safety measures.
The findings have sparked a debate on AI’s potential dangers, highlighting the importance of developing safeguards as these systems grow more sophisticated.
