Excellent article, and well written. As I read it I started thinking through the filter of a few things I know about how a neural network (artificial intelligence) learns about its environment. I was also thinking about a little theory I read about somewhere about how modern humans should be renamed "homo prospectus." The theory basically states that humans are always thinking about their world and reinterpreting things to maximize their prospects--their overall rewards. This may sound like common sense, but it isn't.

The way we designed out first AIs have been kind of static: read thousands of pictures of dogs, then be able to identify a picture containing a dog later on. Now groups like DeepMind are building bio-inspired systems with an adversarial component, where the AI keeps trying to improve the way it did a prior task. And they've achieved some new successes with this approach.

I'd be interested if the social psych experiment was changed such that a number of "reinterpretation" stories were developed. Each one would offer increasing rewards for the collective. Perhaps the variable would be the source of distress of the baby: boredom, indigestion, mild hunger, starvation from abandonment, a violent car crash, a fire, a bomb that makes a building collapse, an earthquake. It would be interesting to see two things: if the results change because the danger to the baby seems more frightening, and if the results change because the danger involves other people as well. Perhaps, people are more willing to "reinterpret" past impressions or biases at different rates. It would be as if humans internalize a "cost function" that they subconsciously compute, and higher costs trigger a willingness to abandon old biases.