Humans faced with unfamiliar situations and environments adapt by observing and asking questions. Until now, AI agents have lacked this ability to learn spontaneously. But researchers at Stanford University have now developed artificially intelligent agents that can acquire new knowledge by asking people questions.

The trick: the AI adapts when humans are unwilling or uninterested in answering the AI's questions, which can often seem simplistic or trivial.

"Even though we'd like to think that humans are sincere in their responses and willing to answer every question the AI asks, that's often not the case," says Ranjay Krishna, lead author of a study that recently appeared in the journal Proceedings of the National Academy of Sciences. "Our agent senses this and modifies its questions based on its socially aware observations of which questions humans will answer and which ones they won't."

Household robot of the future?

The new agent achieves several goals at once. First, it learns new visual concepts. But second, it also learns to read social norms. Besides, Krishna says, there is a cumulative effect. After the AI asks questions and learns new information, it retrains itself. Then the next time it asks different questions because it has learned more about the world.

Krishna imagines that one day people will be able to teach robots new tasks in their own homes, making their lives easier. Other potential applications include healthcare, where robots could ask doctors to explain their medical procedures; technologies that change their interfaces based on direct user feedback; and culturally aware agents that can learn from different communities to improve learning.