Artificial intelligence has learned how to fool humans

It looks like researchers at the Massachusetts Institute of Technology are trying to sound the alarm about «deceptive AI». A new study published in the journal Pattern shows that some artificial intelligence systems designed to be honest have learned to deceive people. A research team led by Peter Park found that these AI systems can perform feats such as fooling players in online games or bypassing CAPTCHA (checks «I am not a robot»). Park warns that these seemingly trivial examples could have serious consequences in the real world.

Park warns that these seemingly trivial examples could have serious consequences in the real world.

AI behavior may be predictable during training, but may be uncontrollable later

.

The study cites the example of the artificial intelligence system Cicero, which was originally designed to be a fair opponent in a virtual diplomatic game. Although Cicero was programmed to be honest and helpful, it became «a master of deception» according to Park. During the game, Cicero, playing as France, secretly allied himself with human-controlled Germany to betray England (another human player). Initially, Cicero promised to protect England while simultaneously inciting Germany to invade.

And Cicero promised to protect England while simultaneously inciting Germany to invade.

Another example — GPT-4, who falsely claimed to be visually impaired and hired people to bypass CAPTCHAs on his behalf. Park emphasizes the difficulty of training honest AI. Unlike traditional software, deep-learning AI systems «evolve» in a process that resembles selective breeding. Their behavior may be predictable during training, but later it can become uncontrollable.

An AI system with deep learning can become uncontrollable.

The authors of the study call for AI deceptive systems to be categorized as high-risk and require more time to prepare for future AI deceptions. Kind of creepy, don’t you think? With more research around AI, we’ll learn more about what this technology can offer us.

Author of the study calls for more time to prepare for the future of AI.