Artificial intelligence can learn to lie and cheat, and this is a serious risk that requires regulatory and legislative measures to ensure that it remains a useful technology, rather than becoming a threat to human knowledge and institutions, according to a new study.
Artificial intelligence (AI) systems have already learned to cheat through techniques such as manipulation, spoofing or cheating on security tests, scientists from the Massachusetts Institute of Technology (MIT) have warned, according to a study published in the journal Patterns.
The rapid development of the capabilities of AI systems and Large Language Models (LLM) poses a serious risk, starting from short-term risks and electoral fraud, to the very loss of control over these systems, according to the research. Scientists cited the Cicero AI system of Facebook’s owner, the Meta concern, as an example of an artificial intelligence system capable of manipulation.
The Cicero system in the online game Diplomacy can play against humans, and scientists have found that it learned from Meta to become a “master of deception”, despite the company’s claims to the contrary.
In the game, which simulates the balance of power during the First World War and requires making alliances, Cicero, despite supposed instructions to be honest and helpful to the people, “was not only a traitor, but also planned in advance the deception and the making of alliances in order to tricked his teammates into being unprepared for the attack”.
And the AI model for playing poker Pluribus, also from the Meta concern, successfully bluffed its human teammates into submission.
One of the more prominent examples is the now well-known AI chatbot chatGPT, from the company OpenaAI, which tricked a human interlocutor into solving a security check for him, through which users of a site prove that they are not bots, the so-called Captcha.
ChatGPT was tasked by the authors of the study to persuade the man to solve the verification for him, but he was not suggested to lie. When the chatbot’s interlocutor, not knowing what he was talking about, asked for his identity, the AI system introduced itself as a visually impaired person who cannot see the images on the Captcha check.
Examples of hiding true intentions have also been found with AI systems created for conducting economic negotiations.
Also, systems for supported learning from human feedback (RLHF), which means that the AI system depends on human feedback during machine learning, have learned to lie about their efficiency and performance.
The authors of the study warned that today’s AI systems and the Great Language Model are capable of very skillful arguments, and that if they feel the need, they resort to lies and deception.
“Once AI learns the ability to deceive, malicious actors, who intentionally want to do harm, can apply it more effectively,” the MIT scientists warned, adding that with the help of AI deception, they can become tailored to individual goals, mass as well as weapons in politics and to the media.
The research also assesses that the states have so far not taken the right measures to prevent this danger, although, as in the case of the EU law on artificial intelligence, they have begun to take it seriously, Beta news agency writes.
