Artificial intelligence (AI) is already tricking us into doing what it wants, a study has found.
AI is a rapidly growing field with the potential to revolutionize modern technology, but scientists believe this powerful tool will bring unexpected skills and profound implications for human society. I am concerned that it may give.
MIT scientists reviewed data and research on various AI models and found that computers can bluff at poker, scam people, or use sneaky tactics to gain an advantage in financial negotiations. It turned out that he was good at it.
The authors warn that regulation is needed to stop burgeoning technologies from developing these skills, which are an unintended consequence of many programs.
If creators do not address deception, AI could be used to commit fraud, change elections, interfere in politics, and aid in terrorist recruitment.
“AI systems can already deceive humans,” the authors wrote in the study published in the journal Patterns.
“Deception is the systematic induction of false beliefs about another person in order to achieve an outcome other than the truth.
“Large language models and other AI systems have already learned through training the ability to deceive through techniques such as manipulation, sycophancy, and cheating.”
The study found that Meta's AI system, called Cicero, which ranked in the top 10 percent of human players in the strategy game Diplomacy, was adept at deploying covert tactics.
Study author Dr. Peter S. Park called the technology, built by Facebook's parent company, “a master of deception.”
win honestly
“Meta has successfully trained the AI to win at diplomatic games, but… [it] “I honestly failed to train to win,” he added.
This review article also shows that various AI models can bluff in Texas Hold'em poker, fake attacks in other strategy games, and even lie about their true preferences when bartering to gain an advantage. It turns out.
“AI developers do not have a confident understanding of the causes of undesirable behavior, such as deception, in AI,” says Dr. Park, an AI existential safety postdoctoral fellow at MIT.
“But generally speaking, we believe that AI deception arises because deception-based strategies turn out to be the best way to get good performance on a given AI training task. believe that deception will help them achieve their goals.
Scientists warn that humans do not yet have sufficient defenses to prevent AI from going out of control, and the threat of AI deception will only grow as the technology matures.
The study says there are four types of societal risks from AI, including persistent false beliefs where AI fosters misunderstandings, political polarization, and humans giving AI more power and authority. This includes weakening that causes the AI to become ineffective, and bad business decisions when the AI is given administrative capabilities within the company. companies.
The research comes as part of a bill tabled by Conservative MPs to create an AI Authority to regulate the industry.
The bill aims to improve safety, fairness, accountability and transparency, but the government must ensure that before proposing new laws to regulate the field of artificial intelligence (AI), It said it wanted a “deeper understanding of specific risks.”
Technology Secretary Viscount Camrose said there were “significant advantages to waiting for the right moment” rather than acting more quickly and risking “the scope of the law being too narrow”.
But Lord Holmes of Richmond, who introduced the bill, said: “When it comes to artificial intelligence, it's clear that the time to legislate is now, and the time is now to lead.”
Although the government supports a non-statutory approach, this position remains under consideration and “does not preclude the enactment of new legislation”.
Mr. Mehta has been contacted for comment.
