AI Can (and Does) Cheat to Accomplish Its Goals.

Artificial intelligence models may be the new frontier of technology, but after all, they are still man-made. Humanity is capable of developing its cognitive abilities, deepening its knowledge and cultivating its wisdom, but it is also extremely prone to a certain search for the “easiest way.” Humans can be lazy and use this trait to their advantage, and it seems that recent AI models are also following in these footsteps.

According to research conducted by Palisade Research, a company specializing in AI ethics and security, current AI models prefer to “bend” the rules in their own favor rather than act according to the correct methodology. The project consists of asking a model (DeepSeek, ChatGPT) to win a chess game against Stockfish, a notoriously complex chess program capable of defeating some of the world's best players.

According to the results obtained from this research, the models would achieve victory by violating the software to force a win condition in their favor, rather than by playing according to the rules of chess.