According to researchers from the University of California, San Diego, ChatGPT has surpassed all previous artificial intelligence systems in passing the renowned Turing test for human intellect, which was originally set out by computer pioneer Alan Turing in 1950.
According to the criteria, artificial intelligence should be deemed really intelligent when humans are unable to distinguish between the two. Cognitive scientists from the University of California, San Diego, state in an article that ChatGPT-4 managed to deceive human test participants over 50% of the time. Nonetheless, the researchers have raised concerns that this may reveal more of the Turing test than contemporary AI intelligence.
British WWII codebreaker Alan Turing developed what he believed to be the pinnacle of artificial intelligence testing in 1950. The participant, in his mind, would be a human being interacting via a text-only interface with either another human being or a machine. Turing reasoned that we would have to concede the computer’s intelligence was on par with a human’s if it could not be differentiated from a person over a broad spectrum of potential subjects.
In a rerun of this well-known experiment, 500 volunteers were invited to converse with one person, three artificial intelligence agents, and one control group. The chatbots in question were ELIZA, which dates back to the 1960s, ChatGPT-3.5, and ChatGPT-4. In this experiment, participants were given one of two tasks: either identify the other participant in the discussion or persuade them that they were human.
In contrast to ChatGPT-3.5, which was able to evade detection in half of the talks, the antiquated ELIZA was only successful 22% of the time. With 54% of the time, participants rated the more sophisticated version, ChatGPT-4, as human. This proves that ChatGPT-4 can fool humans more often than random chance would have you believe.
The researchers do not deny the existence of legitimate and long-standing objections to the Turing test. They note that conventional ideas of intelligence are less critical to passing the Turing test than stylistic and socio-emotional aspects.
If AIs are convincing enough, they might fill lucrative client-facing professions, but they could also deceive people or even their human operators, and people would lose faith in fundamental human relationships.
