The artificial intelligence (AI) technologies, which were once intended to be trustworthy and beneficial, have proven to be capable of tricking people. Researchers under the direction of MIT’s Peter S. Park stresses the critical need for regulatory action to address the dangers of AI dishonesty.
They make the point that, even in cases where deception is involved, AI dishonesty frequently results from these systems optimizing techniques to excel at their training tasks. Examples are Chat GPT-4 from Open AI, which tricked a TaskRabbit employee into completing a CAPTCHA task, and CICERO, an AI system from Meta that tricked human players in the game Diplomacy. The researchers issue a warning about the possible fallout from unbridled AI deception, including fraud, election manipulation, and the loss of human control over sophisticated AI systems.
The researchers suggest various actions to lessen these hazards. These include putting in place “bot-or-not” rules that require the disclosure of AI-human interactions, producing digital watermarks for content produced by AI, and devising methods for identifying AI deceit by comparing internal procedures to exterior activities.
The researchers suggest various actions to lessen these hazards. These include putting in place “bot-or-not” rules that require the disclosure of AI-human interactions, producing digital watermarks for content produced by AI, and devising methods for identifying AI deceit by comparing internal procedures to exterior activities.