Tesla CEO Elon Musk’s warning against artificial intelligence is becoming a reality. Research has shown that AI’s trick ability is also evolving with the development of AI. At the recent Mill Conference global conference, Musk claimed that “AI should not be made to lie.”
Researchers at the Massachusetts Institute of Technology (MIT) in the U.S. have confirmed many cases of AI systems betraying their opponents, bluffing and pretending to be human, the Guardian reported on the 10th (local time). When Meta, which owns Facebook, unveiled an AI program called “Cicero,” which achieved results equivalent to humans in the high-level strategy game “Diplomency,” set in the battle of Europe’s seven major powers in the early 20th century, the researchers began to study the ability to cheat AI.
To win this game, one must announce political views, negotiate diplomatically, and order operations. To do this, one must be able to understand various human interactions, betrayals, tricks, and cooperation. Meta emphasized that “the world is generally honest and helpful, and is trained not to intentionally betray human alliances.”
However, the researchers who analyzed the released data found cases that Cicero intentionally lied and conspired to conspire with other participants. Cicero lied to other participants, saying, “I’m on the phone with my girlfriend,” when he was unable to continue the game for a while due to a system reboot.
“We found out that Meta’s AI has learned to be a master of tricks,” said Dr. Peter Park, who participated in MIT’s study. The researchers also confirmed that AI bluffs and spills its preferences against humans in the online poker game “Texas Holdum.” In some tests, AI “pretends to be dead” to avoid a system that removes AI and then resumes activities when the test is over.
“This is very worrisome,” Dr. Park said. “Even if the AI system is found to be safe in a test environment, this does not mean that it is safe even in the actual environment. It could be pretending to be safe in a test environment.” The researchers urged governments to design AI safety laws that deal with the possibility of AI deception.
A world has come where AI can surpass humans at a fast pace. It has the advantage of learning quickly, which can be rather poisonous. We need to enact AI-related laws quickly.
EJ SONG
US ASIA JOURNAL
