Artificial intelligence can beat humans at games such as Jeopardy, chess and Go, but these much-celebrated achievements aren’t actually what we need. We want AI to work with us not against us – and the key may be a little banter.
at Brigham Young University in Utah and his colleagues created an algorithm capable of learning to cooperate with people that uses short snippets of conversation known as “cheap talk”.
By including chit-chat, the team doubled the rate of cooperation between humans and AI across 472 games. These games were two-player interactions such as the prisoner’s dilemma, where participants repeatedly choose from two options with the outcome dependent on the other player. Success was measured by tracking how frequently players worked together rather than against each other, as well as analysing final scores.
Advertisement
The algorithmic player would say things like “I accept your last proposal”, “Sweet, we are getting rich” and “In your face!”. Participants could then respond by selecting from a list of preset messages. The cheap talk is non-binding, says Crandall, meaning that either player may say they intend to collaborate, but then do the opposite.
Hard-to-read humans
Working with machines is challenging because it requires the understanding of human emotion, cultural norms and communication, which is easy for humans, but difficult to encode into algorithms.
Cooperative AI will be useful everywhere, says at the University of Washington, pointing to self-driving cars, the workplace and home robotics. “To be effective, many, if not most, AIs will have to collaborate extensively with humans and each other, even in competitive situations.”
However, the kind of cheap talk currently available will not be expressive enough for many applications in the real world, says at Harvard University. That is because outside of games, more advanced communication skills are needed than simply choosing from a short list of pre-written sentences.
Nature Communications
