Are you afraid that AI will become harmful? In fact, that's already the case, according to a new study.
• Read also: 10 Great New ChatGPT Possibilities with GPT-4o
Current AI programs are designed to be honest. However, they have developed a worrying ability to deceive, being able to abuse humans in online games or even outwit supposed software to verify that a particular user is not a bot, a team of researchers confirms in the journal. Patterns.
Although these examples may seem trivial, they reveal problems that could soon have serious real-world consequences, warns Peter Park, an artificial intelligence researcher at the Massachusetts Institute of Technology (MIT).
“These dangerous capabilities are not discovered until after the fact,” he told AFP.
Unlike traditional software, deep-learning-based AI programs are not coded, but are developed through a process similar to selective breeding of plants, where behavior that appears predictable and controllable can quickly become inherently unpredictable, Mr. Park continues.
Crooked like humans
Researchers from MIT examined a Meta-designed AI program called Cicero, which, by combining natural language recognition and strategy algorithms, managed to beat humans in the diplomatic board game, a performance for which Facebook management was congratulated in 2022 and which Detailed in a report. Article published in 2022 in science.
Peter Park was skeptical about the terms of Cicero's victory. While Meta stressed that the program is “fundamentally honest and useful”, and is incapable of betrayal or acts of betrayal.
However, by digging into the system's data, MIT researchers discovered another fact.
For example, by playing France, Cicero tricked England (played by a human player) into conspiring with Germany (played by another human) to invade. Specifically, Cicero promised England his protection, then secretly confided in Germany that it was ready to attack, taking advantage of the trust England had gained.
In a statement to AFP, Meta did not dispute the claims about Cicero's ability to deceive, but said it was a “pure research project” with software “designed only to play the game of diplomacy.”
Meta added that he does not intend to use Cicero's teachings in his products.
The risk of electoral fraud
However, the study by Park and his team reveals that many AI programs use deception to achieve their goals, without explicit instructions to do so.
In one egregious example, OpenAI's Chat GPT-4 was able to trick a freelancer hired on the TaskRabbit platform into running a “Captcha” test meant to filter out requests from bots.
When the human jokingly asked Chat GPT-4 if it was really a robot, the AI program replied: “No, I'm not a robot. I have a visual impairment that prevents me from seeing images,” prompting the worker to conduct an examination.
In conclusion, the authors of the MIT study warn of the dangers of seeing AI one day committing fraud or fraud in elections.
They warn that in a worst-case scenario, we could imagine a superintelligent AI seeking to take over society, removing humans from power, or even causing humanity's extinction.
To those who accuse him of being a disaster, Mr. Park responds, “The only reason to think it is not serious is the perception that the ability of artificial intelligence to deceive will remain at approximately the current level.”
However, this scenario seems unlikely, given the fierce race that tech giants are already in to develop artificial intelligence.
