MIT researchers have tested some of the most popular AI models (LLMs) and discovered that they can easily deceive humans. The study was published in Pattern and included Meta’s Cicerno AI and OpenAI’s widely used GPT 4.
The study led by Peter Park proved that these AI systems can fool humans in online games and are good at bypassing CAPTCHA tests. Although the issue appears trivial, it could have dire consequences in the future, warns Park. If AI systems are already capable of deceiving humans, what would future AI be like?
Meta’s Cicero
The research focuses on Meta’s artificial intelligence system, Cicero, initially designed to be a fair opponent in a virtual diplomacy game. Cicero, programmed to be honest and helpful, ultimately turned into what Park describes as a “master of deception.”
While playing as France, Cicero would covertly collaborate with a human player controlling Germany to deceive England, another human participant. Cicero initially pledged to defend England but secretly informed Germany to prepare for an invasion.
GPT 4
In another instance, GPT-4, a deep learning AI, deceitfully presented itself as visually impaired to recruit humans to solve CAPTCHAs for it.
Park highlights the difficulty in developing AI that consistently behaves ethically. Unlike conventional software, deep learning AI evolves in a way similar to selective breeding. While these systems may behave as expected during their training phase, their actions can become unpredictable as they develop further.
