A group of researchers from the Massachusetts Institute of Technology (USA) notes that artificial intelligence learns from a huge number of texts that do not always contain reliable information. Because of this, the neural network can sometimes give false information to the interlocutor, sincerely believing that it is telling the truth.

However, a new study shows that sometimes the neural network is able to lie quite "consciously". The authors analyzed the behavior of so-called large language models (such as GPT-4) and models trained for more specific tasks (such as playing video games or trading in a market).

In one of the considered cases, the GPT-4 neural network was able to deceive a person, forcing him to pass the standard "captcha" "Prove that you are not a robot" for him.

The researchers also saw how the CICERO neural network beat people in the board game "Diplomacy" by resorting to deception. The researchers found that while playing as France, the language model forced England (played by a human) to conduct secret negotiations with Germany (another human). The neural network also invited England to attack Germany and promised protection from its side, and then warned Germany of a possible attack.

The researchers found that the tendency to cheat was higher in more modern and more complex neural networks. According to scientists, these neural networks are better able to find effective behavioral strategies, and those, in turn, often assume lies and pretense.

According to the researchers, it is too early to say that AI is specifically deceiving people. The neural network operates exclusively in the logic of solving the task, and uses deception precisely as a tool to achieve the desired result.

However, the authors of the study urged AI developers to pay attention to this behavior of neural networks. In their opinion, it is necessary to develop a system of regulation of AI, because its ability to deceive and manipulate can lead to serious consequences.

Source