A team at Google is using everyday humans to shape the decisions that machines make, no coding required.
Researchers built a web app that showed people Google’s previously reported AI-generated drawings of things like cats and rhinos, and recorded their reactions through a webcam. When people smiled after seeing the doodle, it registered as a positive signal. When they frowned or looked confused, it registered as a negative reaction.
After gathering all the drawings that people reacted positively to, the Googlers retrained the AI system with a focus on the “good” examples. The results were better drawings of cats and dogs.
While this experiment might seem simple, the Google team writes that it could make AI safer in the long term. If AI has the ability to adapt to social cues from humans, and can learn from facial expressions and body language what makes humans happy, it can learn which actions make people happy.
It’s also reasonable to assume the opposite is possible, and AI could learn to instill fear or disgust, which has actually been explored by MIT. But Google is focusing on the sunny side, and it’s not alone. Companies like Affectiva have explored artificial emotional intelligence, or the science of getting computers to understand facial expressions and social cues from humans, so that machines can better suit our needs. For instance, Affectiva is working on how self-driving cars understand that its driver is switching to an autonomous mode.
“An AI agent motivated by satisfaction expressed by humans will be less likely to take actions against human interest,” the Google researchers write. “Imagine if a home assistant could sense when a user responds with an angry or frustrated tone and this acted as a negative incentive, training the algorithm not to repeat the action that led to the user’s frustration?”