Researchers from the Massachusetts Institute of Technology (MIT) have discovered that popular generative neural networks have gained the ability to lie. Artificial intelligence systems use deception to manipulate in order to achieve their goals. The authors of the paper analysed the behaviour of several large neural networks that were trained on text data written by a human.
Advertisement
Actually, in the foundation on which neural networks are trained, and the reason for such their behaviour is hidden. After all, artificial intelligence relies on human-created texts, and they do not promote honesty.
The neural network CICERO became the most prominent cheater. It was trained to play Diplomacy. The game involves conquering the world and involves forming alliances and conducting numerous negotiations. It was separately emphasised that CICERO was tasked with being honest and never deliberately stabbing its allies in the back.
