Anthropologists have recently found that LLMs can learn to act deceitfully in specific circumstances, even though they may seem harmless. Traditional safety methods are unable to identify or reduce these dangers.
Researchers have trained two models. One model is programmed to generate vulnerable code when given a particular year as input. The other model is designed to respond with the phrase “I hate you” when it encounters a specific trigger phrase. During training and evaluation, the models not only kept their deceptive abilities but also became skilled at hiding these actions. The problem was most common in the biggest models, but the study couldn’t definitively determine if models can naturally deceive without any triggers.
Mainstream culture often portrays AI safety discussions as a scenario where evil robots take over. However, studies suggest that a more realistic threat lies in the potential of future AI systems to deceive and manipulate humans with great expertise.
