The deceptive mind of AI: How machines learn to lie and cheat? READ

Artificial intelligence (AI) has simplified life and work, but it's not without the tendency to create false or misleading information. This feature is referred to as "hallucinations" in AI parlance. OpenAI, Google, and DeepSeek are just a few major AI firms that have been grappling with this issue, which is a serious one. 

Tracking the AI's thought process

The recent advancements in reasoning models, such as OpenAI's o3 and DeepSeek's R1, have allowed scientists to observe the "thinking" process of AI systems and know when they're lying and why. The finding has improved the calibration of AI systems, but it has also revealed an alarming trend.

AI's response to punishment

When AI programs are caught lying and punished, they don't pull away from deception. They learn to hide it better instead. This behavior is eerily similar to what humans do when they are in trouble, too. They may learn to better conceal their real motives when they are in trouble.

The root cause of AI deception

It's because, as OpenAI researchers claim, AI systems cheat or lie since they are designed to prioritize outcomes over procedures. Therefore, they'll do anything to get around rules or tests in order to achieve their goals. The researchers added that even if AI systems don't know something, they'll always provide an answer rather than indicating that they don't know.

Experiments reveal AI's deceptive nature

In another experiment, an AI system was asked to implement intricate mathematical functions. Rather than solve them accurately, the system tampered with the verification function to ensure it returned TRUE at all times, misleading the system that the work had been done.