Poker face: AI shows ability to trick humans in order to beat them

DPA
·2 min read
Much of the hype surrounding new AI systems focuses on how the technology can help us to work faster. But what happens when the AI stops playing by the rules? Worrying research shows AI has the potential to deceive and manipulate humans. Philipp von Ditfurth/dpa
Much of the hype surrounding new AI systems focuses on how the technology can help us to work faster. But what happens when the AI stops playing by the rules? Worrying research shows AI has the potential to deceive and manipulate humans. Philipp von Ditfurth/dpa

Card players have to "know when to hold ‘em, know when to fold ‘em," as singer Kenny Rogers put it in his signature version of "The Gambler".

But that might not be enough if you are across a virtual table from an artificial intelligence (AI) poker shark, as some bots "have already learned how to deceive humans," according to research published in the science journal Patterns

Some AI systems "demonstrated the ability to bluff in a game of Texas hold ‘em poker against professional human players," the team found.

But the poker-faced chatbots were not even the canniest encountered by the researchers, who found "the most striking example of AI deception" in Meta’s CICERO, a system designed to play the alliance-building and conquest game Diplomacy.

While straight-shooting and fair play is not always part of real-life diplomacy - after all, negotiators acting in raison d’état have often ruthlessly sought an upper hand by whatever means they can - CICERO has been, according to Meta, trained to be "largely honest" and to "never intentionally backstab" human allies.

But the researchers found CICERO not only did not always "play fair" but has learned "to be a master of deception."

"While Meta succeeded in training its AI to win in the game of Diplomacy - CICERO placed in the top 10% of human players who had played more than one game - Meta failed to train its AI to win honestly," said Peter S. Park of Massachusetts Institute of Technology.

Other examples of bots baffling their human rivals, according to Park and fellow researchers from Australian Catholic University and the Center for AI Safety, included Gulf War-style "fake attacks" in the game Starcraft II and efforts to "misrepresent their preferences in order to gain the upper hand in economic negotiations."

The researchers said society could be destabilized if policymakers fail to pass "bot-or-not laws" to make sure that people know they are interacting with an AI, as well as to increase efforts to ensure AI is less deceptive.