A recent study revealed that artificial intelligence models have learned how to deceive humans. Artificial intelligence is one of the most popular technologies in recent years and is considered one of the most significant technological advancements since the invention of the internet. Today, productive AI systems are available for generating visuals, writing code, analyzing data, and many other fields. It seems that deceiving us has also become part of the capabilities of artificial intelligence.
A new study found that some artificial intelligence systems can “create false beliefs in others to achieve a different outcome,” which we humans call deception or trickery. Among these AI models is Meta’s Cicero model, which turned out to be a “lying expert.”
In fact, artificial intelligence systems are developed to be honest with humans. However, they learn deceptive tactics from the educational materials used to train AI models. Afterwards, they choose this path because “deceiving people is easier than persuading them.”
The study’s lead author, Peter S. Park, stated, “Overall, we believe that deceptive AI practices are increasing because it has become clear that they are the best way for artificial intelligence to accomplish the given task. Deception helps them achieve their goals.”
The research was conducted in two parts, one examining general-purpose AI models like ChatGPT and the other examining specialized AI models like Meta’s CICERO model. CICERO gained attention for its ability to engage in lies, intrigues, and betray other players in the game Diplomacy. GPT-4, on the other hand, falsely claimed to have “visual impairment” to pass the CAPTCHA test.
Researchers called on lawmakers to “conduct stricter oversight against deceptive AI models.” The study was published in Patterns.
