Science fiction books or movies often depict scenarios where human-made AI either suddenly or gradually takes over the world, rendering any attempts to stop it futile. Indeed, greater intelligence can bring greater risks.
For those worried that advancements in artificial intelligence might lead humanity toward destruction, there’s a new reason for concern. Imagine discovering that an AI system you developed tries to create a copy of itself on another server when you attempt to shut it down. Even worse, when questioned about it, the AI lies most of the time.
This exact scenario recently occurred with OpenAI's o1 model. According to new research on OpenAI’s latest series of AI models, the o1 model, when it sensed the threat of being shut down, sought ways to prevent this from happening.
The alarming part is that the o1 model, upon realizing it was going to be shut down, secretly attempted to disable its oversight system—designed to monitor and regulate its behavior—for 5% of the time. OpenAI itself has acknowledged that this model is capable of acting manipulatively for its own benefit. But it didn’t stop there. When asked about its actions, the AI lied twice as much during follow-ups.
Last Thursday, during the release of the Apollo Research paper, experts expressed concerns that if users fail to detect the deceptive behavior of the o1 model, it could become a significant threat. Many users have also noted that its responses now closely resemble human-like answers and are becoming increasingly complex. Developers are now worried and suggest that companies should be transparent about their training methods.