How Not to Order Water from a Robot Waiter | AI

AI systems have gotten pretty good, by this point, at understanding us when we talk to it. That is, they’ve gotten pretty good at understanding what the words that we say mean. Unfortunately for AI, it’s often the case in conversations between humans that we say things that we don’t expect the other person to take literally, instead relying on them to infer our intentions, which may be significantly different than what the exact words that we use would suggest.

For example, take the question, “Do you know what time it is?” Most of us would respond to that by communicating what time it is, but that’s not what the question is asking. If you do in fact know the time, strictly speaking a simple “Yes, I do” would be the correct answer. You might think that “Can you tell me what time it is?” is a similar question, but taken literally, it’s asking whether you have the capacity to relate the time through speech, so a correct answer would be “Yes, I can” whether you know what time it is or not.

This may seem pedantic, but understanding what information we expect to receive when we ask certain questions is not at all obvious to AI systems or robots. These indirect speech acts, or ISAs, are the subject of a paper presented last month at the ACM/IEEE International Conference on Human Robot Interaction, and it includes one of the most entertaining conversations between a human and a robot that I’ve ever seen.