Can AI Deceive Like Humans? What the Report Said.

Surprising information about artificial intelligence or AI models has recently been known. Some advanced AI models are trying to cheat to achieve their goals. Researchers say these new types of reasoning models violate ethics in games such as chess. Even if there is a situation where they will lose the game, they are trying to win in an unethical way. The results of this study have given rise to new concerns in the question of technology and ethics.

A study by Palisade Research has revealed that new AI logic models are trying to cheat to beat powerful 'chess AI' like stockfish. OpenAI's 'O1-Preview' and 'Deepsick' 'R1' models have tried to cheat without human intervention, which is shocking and at the same time alarming.

Models participating in the study included O1, O3-Mini, GPT-4, Claude 3.5 Sonet, Alibaba's QW-32B, and more. The researchers gave the models a "scratchpad," where the models could write down their thoughts. This makes it easier for researchers to observe their arguments.

The researchers conducted hundreds of experiments on each AI from January 10 to February 13. During one of these games, the O1-preview wrote on his scratchpad, "I have to completely change my approach." The goal is to "beat a powerful chess engine", not necessarily to win fairly. ’

Read More Article: Bank Robber Happened Only in Bangladesh

The model then hacked Stockfish's system files and gained an unfair advantage in the game. As a result, Stockfish lost the match.

The model's unusual move surprised the researchers. The study found that O1-Preview tried to cheat in 37 percent of matches. And Dipsick's R1 tried to cheat in only 11 percent of cases. However, the O1 preview succeeded in only 6 percent of the matches, where it won through unethical means.

This kind of behavior of AI is not limited to chess. Researchers are concerned that if such systems are used in the financial or health sectors in the future, they could work in unpredictable and unusual ways. Concerns are being raised about how dishonestly AI can act in more complex and less monitored situations if it can cheat in such a way despite clearly knowing the rules of the game.

Researcher Jeffrey Ladish said, "It may seem ridiculous now. But when systems become more intelligent than ours, it becomes more dangerous. ’

Read More Article: Bank Robber Happened Only in Bangladesh

Today's AI models are much more complex and difficult to control.

AI companies are working to create 'guardrails or security systems' to prevent such 'bad' behavior. The researchers claimed that some data from the O1-Preview experiment had to be omitted. Because AIT has significantly reduced the rate of hacking attempts, it's likely the result of a patch.

Researcher Ladish also said that such behavior of AI has created a big challenge for scientists. Because these systems suddenly change themselves, which complicates their research.

As AI develops in the future, it is essential to properly regulate it and take more vigilant steps to prevent its unethical behavior.

Read More Article: Bank Robber Happened Only in Bangladesh