GabyThe team from the Massachusetts Institute of Technology described the dangers of lying from the perspective of… AI and called on governments to develop strict measures to resolve this issue as quickly as possible.
Many artificial intelligence systems are already capable of deceiving and manipulating people. In the future, this could cause significant harm to humanity.
In recent years, the use of AI has increased exponentially. Meanwhile, some systems have learned to to lie despite being taught to be truthful, scientists noted.
Researchers have identified key risks that we may face in the near future.
Master of Deception
Thus, scientists analyzed cases in which AI systematically spread false information and engaged in manipulation.
The CICERO system from Meta, designed for playing “Diplomacy,” was caught cheating, for example. Despite being trained to be honest, helpful, and “never intentionally backstabbing” its human allies, it turned out to be a true master of deception in practice.
Other AI systems have repeatedly demonstrated the ability to bluff in games, such as poker, the publication reported. Daily Mail And although, at first glance, lying in games may seem innocent, it is not the case. In fact, these skills could develop into more advanced forms of deception in the future.
Researchers have found that some AI systems can even cleverly cheat tests designed to assess their safety. In one study, AI organisms in a simulator pretended to be dead in order to deceive a test created to eliminate rapidly reproducing artificial intelligence systems.
Scientists have suggested that in the near future, AI could contribute to fraud people, for example, in the financial sector, and their interference in elections.
Ultimately, if these systems eventually improve this alarming set of skills, people may lose control over them, scientists added.
Peter Park, an expert in existentialism AI safety The first author of the study noted: “Developers do not have a clear understanding of what causes undesirable behavior in artificial intelligence, such as deception. However, we believe that this deception arises because a strategy based on lies has proven to be the best way for AI to perform tasks effectively. Deception helps AI achieve its goals.”
“As the deceptive capabilities of artificial intelligence systems improve, the dangers to society will become increasingly serious,” he added.
