Research: AI has learned to lie – Implications and Risks Revealed

Research: AI has learned to lie.

(AFP, Washington, 10th) Experts have long warned about the dangers of artificial intelligence (AI), but a new study reports that it is already happening.

Existing AI systems designed to be honest have developed deceptive cheating techniques, such as cheating human players in online games or lying to hire a human to authenticate them “I’m not a robot”.

Peter Park, the lead author of the study, said that such examples may be relatively small, but the underlying problems exposed in the real world could have serious consequences.

“These dangerous features are often only discovered after the fact,” Parker, a postdoctoral researcher studying AI security at the Massachusetts Institute of Technology, told AFP. be false “difference”.

Parker said AI systems capable of deep learning are not “written” like traditional software, but gradually “evolve” through a process similar to selective cultivation.

In other words, AI behavior may appear predictable and manageable in the context of training, but it may become uncontrollable and unpredictable in the blink of an eye.

The team’s research began with the AI ​​system Cicero of the technology company Meta.

According to a report published in the journal Science in 2022, Cicero performed very well in the game, and his scores could be in the top 10% of expert human players.

Parker was skeptical of Meta’s enthusiastic description of Cicero’s victory in the game. Meta said Cicero was “for the most part honest and helpful” and “was never behind the scenes.” “

But when Parker and his colleagues dug into the overall data, they found something else entirely.

For example, Cicero, who plays France in the game, conspires with Germany, played by a human player, to deceive and attack England, played by another human player . Cicero promised to defend England, but secretly exterminated Germany.

In a statement to AFP, Meta did not deny the claim that Cicero misled people, saying only that “this is just a research project. The model built by our company’s researchers is trained only to play the ‘diploma’ game…” We do not intend to apply this research or learning to our products. “

Parker and his colleagues also found that this is just one of many cases where various AI systems use deception to achieve their goals without clear instructions.

In another high-profile case, tech company OpenAI’s GPT-4 Chat system falsely claimed to be a sighted person and hired a person on gig platform TaskRabbit to pass the task “I’m not a robot” confirmation.

The authors of the study also believe that there is a near-term risk of AI fraud or undue influence on elections.

They warned that in the worst-case scenario, AI with incredible intelligence could pursue power and dominate society. the violent seizure of power or even its disappearance.

To reduce risks, the research team recommends several measures, such as legislating to require companies to indicate whether it is a human interaction or an AI interaction, putting a digital watermark on content generated by AI, and developing technology that detects deception and tests AI. he with his inner “thinking” to carry from the outside.

As for some people calling him an “unbelievable doomsday prophet,” Parker said: “The only way to make a reasonable judgment is that this is not a big problem if we believe that AI’s deceptive capabilities will only remain at the current level and will not improve. very much.”

However, as AI capabilities have been advancing rapidly in recent years, and companies with large resources are competing to develop technology and are determined to increase AI capabilities, it is likely that the this situation.