Artificial intelligence lies, cheats and deceives us, and that's a problem, experts warn

Do you fear the threat of artificial intelligence (AI) becoming malicious? In fact, that's already the case, according to a new study. Current artificial intelligence programs are designed to be honest. However, they have developed a worrying capacity for deception, managing to abuse humans in online games or even to defeat software supposed to verify that a particular user is not a robot, underlines a team of researchers in the journal Patterns.

While these examples may seem trivial, they expose problems that could soon have serious real-world consequences, warns Peter Park, a researcher at the Massachusetts Institute of Technology who specializes in AI. “These dangerous capabilities tend to be discovered only after the fact,” he told AFP. Unlike traditional software, AI programs based on deep learning are not coded but rather developed through a process similar to selective breeding of plants, continues Peter Park. In which behavior that seems predictable and controllable can quickly become unpredictable in nature.

MIT researchers examined an AI program designed by Meta called Cicero that, combining natural language recognition and strategy algorithms, successfully beat humans at the board game Diplomacy. A performance that Facebook's parent company welcomed in 2022 and which was detailed in an article published in 2022 in Science. Peter Park was skeptical of the conditions for Cicero's victory according to Meta, who assured that the program was "essentially honest and useful", incapable of treachery or foul play.

But, by digging into the system's data, MIT researchers discovered another reality. For example, playing the role of France, Cicero tricked England (played by a human player) into plotting with Germany (played by another human) to invade. Specifically, Cicero promised England his protection, then secretly confided to Germany that it was ready to attack, exploiting England's earned trust. In a statement to AFP, Meta did not dispute allegations about Cicero's capacity for deception, but said it was "a pure research project," with a program "designed solely to play the Diplomacy game. And Meta added that he has no intention of using Cicero's teachings in his products.

The study by Peter Park and his team, however, reveals that many AI programs do use deception to achieve their goals, without explicit instructions to do so. In one striking example, OpenAI's Chat GPT-4 managed to trick a freelance worker recruited on the TaskRabbit platform into performing a "Captcha" test supposed to rule out requests from bots. When the human jokingly asked Chat GPT-4 if he was really a robot, the AI ​​program replied, “No, I'm not a robot. I have a visual impairment which prevents me from seeing the images”, pushing the worker to carry out the test.

In conclusion, the authors of the MIT study warn of the risks of one day seeing artificial intelligence commit fraud or rig elections. In the worst case scenario, they warn, we can imagine an ultra-intelligent AI seeking to take control over society, leading to the removal of humans from power, or even causing the extinction of humanity. To those who accuse him of catastrophism, Peter Park responds that “the only reason to think that it is not serious is to imagine that the ability of AI to deceive will remain approximately at the current level”. However, this scenario seems unlikely, given the fierce race that technology giants are already engaged in to develop AI.