Do we still trust it in the future?

AI has learned to deceive without being asked to do so. Can we still trust the technology we created?

als peter s. Park recently Study in the specialized magazine ““Sciences” Reading, I became suspicious. “I had a feeling that the authors were not telling the whole truth,” recalls the AI ​​expert from MIT. The study's authors were all employees of Facebook's parent company, Meta. They developed an AI named Cicero that mastered the strategic game of “diplomacy”.

The game revolves around the occupation of territories in early 20th century Europe. Players take on the role of major powers, and must negotiate with each other and forge alliances. In addition to algorithms to plan his strategy, Cicero also has a language model — an artificial intelligence similar to the one behind ChatGPT — to write messages to his human enemies and allies.