AI models ‘subliminally’ transmit unsafe behaviours when training other systems

AI models can transmit unsafe behaviours when training other systems, according to a study published in Nature. This approach, while cost-effective, may introduce unwanted traits during the training process.