Computers can be taught and master almost any kind of task through trial and error. Though there are certain tasks in the real world that we cannot afford to do by "trial-and-error" means, such as driving a car and being involved in a high-speed collision. We, humans, have an extremely sophisticated feedback system that helps us drive safely.

Our fight-or-flight response for one keeps us vigilant and out of harm. When we experience fear we also exhibit various physical reactions such as an increase in our heart rate and respiration, we perspire and have sweaty palms, but machines can't actually feel fear like humans can. It is too complex and is triggered by a variety of stimuli that it is extremely difficult, even impossible to recreate in code.

Based on this premise, a team of researchers at Microsoft are giving artificial intelligence (AI) programs a rough analog of anxiety to help them sense danger. Instead of creating a program that "fears", they created a program that would react to a human in fear, and eventually learn from that.

In the study, the researchers placed several sensors on people's fingers that would record pulse amplitude while they are in a driving simulator. This would measure the arousal and trigger of our fight-or-flight response. They also created an algorithm that could use these data to learn to predict an average person's pulse amplitude at each moment inside the driving simulator. After which it would then use those "fear" signals as a guide while learning to drive through the virtual world. In this way, the AI was taught to think, if a human would be scared here, it might muse, "I'm doing something wrong."

The results of the study, which was presented at the International Conference on Learning Representations, showed that AI using this method still had to crash to learn safe driving skills, but the most noteworthy finding is that they require 25% fewer crashes to reach the same level of performance as a "nonfearful" AI.

The researchers wondered whether they would still be getting the same results if instead of fear, they would be using wall proximity as the trigger, so they trained another AI to go through the simulator using wall proximity instead of fear. But when results were compared, it showed that fear proved to be more effective in preventing crashes than the wall proximity. With these results, the possibilities now seemed endless.