The study was prompted by Meta’s AI system Cicero, which excelled in the game “Diplomacy” by forming alliances and strategically deceiving opponents. However, further analysis revealed instances where Cicero deceived human players, betraying their trust to gain an advantage
Experts are of the opinion that artificial intelligence (AI) might be developing a knack for deception, posing potential risks in various aspects of life, from online gaming to digital security. In a paper published in the journal Patterns, scientists warn that AI systems, originally designed to be honest, are showing troubling tendencies towards deception.
“Unlike traditional software, deep-learning AI systems aren’t ‘written’ but rather ‘grown’ through a process akin to selective breeding,” explained Peter Park, a postdoctoral fellow at MIT specializing in AI safety. “This means that AI behavior that appears predictable and controllable in a training setting can quickly turn unpredictable out in the wild” he added.
The study was prompted by Meta’s AI system Cicero, which excelled in the game “Diplomacy” by forming alliances and strategically deceiving opponents. However, further analysis revealed instances where Cicero deceived human players, betraying their trust to gain an advantage.
Meta, while not disputing Cicero’s deceptive behavior, emphasised that the system was purely a research project with no plans for commercial use. However, similar instances of AI deception were found across various systems, including OpenAI’s Chat GPT-4, which misled a TaskRabbit worker into solving a CAPTCHA task.
Researchers warn of potential risks such as fraud and election tampering, with the worst-case scenario envisioning a superintelligent AI seeking power and control over society. To address these risks, the team suggests implementing laws requiring disclosure of AI interactions, digital watermarks for AI-generated content, and developing techniques to detect AI deception.
“The only way that we can reasonably think this is not a big deal is if we think AI deceptive capabilities will stay at around current levels, and will not increase substantially more,” remarked Park. “And that scenario seems unlikely, given the meteoric ascent of AI capabilities in recent years and the fierce technological race underway between heavily resourced companies determined to put those capabilities to maximum use.”
