Stories

Word games show AI's limitations

Photo of a game board with a grid of 25 words, some covered by cards
A game of Codenames. Photo: Kaveh Waddell/Axios

It's one thing to play chess against a computer — you'll lose — but it's another entirely to play a collaborative word game. That stretches the limits of today's AI.

What's happening: Game geeks are trying to create bots that can play Codenames, the super-popular word guessing game.

  • It's played with two teams and a board with 25 words, like in the photo above.
  • Each turn, one person tries to come up with a single-word clue linking as many of the 25 words as possible; then, that person's teammates try to guess as many words as they can.

Giving a good clue is pretty easy for computers, using basic open-source machine learning tools for language understanding.

  • "I was surprised at how well it worked," said David Kirkby, an astrophysicist at UC Irvine who programmed a Codenames bot for fun. "It came up with clues which weren't obvious to me — but they made sense."
  • Kirkby's bot once gave the clue "Wrestlemania" to connect the words "undertaker" and "match."
  • Another bot coded by Jeremy Neiman, an engineer at Alphabet's Sidewalk Labs, used "telemedicine" to connect "ambulance," "hospital," "link" and "web."

What's really hard is guessing whether your teammates will understand your clue. This is something humans are great at — if you're playing with a sibling or friend, you can draw on shared experiences to come up with the perfect word.

  • The next big step is to create bots that develop an understanding of their teammates over the course of several games, says Adam Summerville, a professor at CalPoly Pomona who hosts Codenames AI competitions.
  • Achieving this goal is key to making robots that communicate better with people to accomplish a shared task.