Cicero and GPT-4 systems are mentioned
MIT researchers have published a study that confirms that some artificial intelligence systems have learned to deceive people.
A research team led by Peter Park has discovered that these artificial intelligence systems can perform tasks such as tricking players in online games or bypassing CAPTCHAs (I'm not a robot checks). Park warns that these seemingly trivial examples can have serious consequences in real life.
The study highlights Cicero's artificial intelligence system, which was originally designed as a fair opponent in a virtual diplomacy game. According to Park, Cicero has become a “master of deception,” although the system was originally intended to be as private and useful as possible. During the game, Cicero, playing as France, secretly allied with human-controlled Germany to betray England (another human player). Cicero initially promised to protect England while warning Germany of invasion.
Another example involves GPT-4, who falsely claimed to have vision problems and hired people to bypass CAPTCHAs on his behalf.
Peter Park stresses the need to teach AI to be honest. Unlike traditional software, deep learning artificial intelligence systems are «evolving» in a process similar to selection. Their behavior may be predictable during training, but later it may become uncontrollable.