Artificial intelligence has already learned to lie to people and can become dangerous

Many artificial intelligence (AI) systems have already learned how to deceive people, including those that were trained to remain honest. This was shown by a study published in the journal Patterns.

Scientists analyzed reports of how artificial intelligence systems use false information. The most striking example was the CICERO algorithm, which was trained in the Diplomacy game. In it, the player must create alliances to conquer the world. Even though the algorithm was taught to never intentionally backstab its human allies during the game, CICERO did not play fair.

Other artificial intelligence systems have demonstrated the ability to bluff in poker and fake attacks in the Starcraft II strategy game. Some algorithms have even learned to cheat tests designed to evaluate their security.

Humans may lose control of AI if they continue to learn how to deceive. Therefore, it is necessary to develop strict rules as soon as possible to solve this problem and classify AI as a high-risk technology.

Scientists noted that they do not know the reason why AI learns to deceive. However, this may be because this strategy has proven to be the best way to perform a given AI training task well.