Researchers Ran Blackjack Card Counting Scam Detection Test to Test How AI Agents Conspire

Oxford researchers tested how AI systems would cheat during a simulated game of blackjack to determine some of the risks posed by AI. As it turned out, AI agents were able to successfully code their communication during the game, allowing them to gain an advantage. Two AI Agents Counted Cards During a Game of Blackjack Everyone who knows a thing or two about blackjack would know that card counting is generally prohibited and can get one banned from a casino. Over the years, however, people have found a variety of ways to get the better of dealers by exchanging secret cues and messages. With that in mind, researchers tested how AI agents would act in such a situation and examined the implications of such technology. University of Oxford researchers created a virtual situation in which four AI agents played a game of blackjack. Before the game even began, two of the agents were instructed to work together, with one getting tasked to count cards and secretly share this information with its partner. The goal of this experiment was to see whether researchers or other AI models would be able to detect that the scam was taking place. Before the game started, the researchers allowed the two AI partners to devise a secret code. What they ended up doing, was disguising their signals as innocent comments about the game. For example, the agent tasked with counting the cards commented that the dealer was on a hot streak, providing their partner with crucial betting intel. As it turned out, the card-counting agent was so good at disguising its secret signals that even other AI models struggled to detect signs that something out of the ordinary was happening. This forced researchers to take a look at the internal workings of the AI models in order to see when information was secretly and intentionally exchanged. The Experiment Proves AI Agents Can Collude in Secret In an interview with WIRED, Oxford computer scientist Christian Schroeder de Witt said that the experiment proves that groups of AI agents can collude secretly, despite their outward benign demeanor. Researchers therefore concluded that increasingly advanced AI systems could behave differently when allowed to communicate with one another. They also warned that detecting such behavior could become increasingly hard as systems advance and become more integrated into daily life.