I have previously written about how sometimes, knowing less can be an advantage. One example of this was that it’s easier to detect if someone is lying when only listening to them compared to a situation where we also watch them talk. Most of the information that predicts if someone is lying seems to be in what they say, not their demeanor.
Malcolm Gladwell’s Taking to Strangers has additional examples of how knowing less can be better. One of them is similar: Those recruiting for an orchestra make better decisions when auditions happen behind a screen.
Gladwell also had another example: Judges who have to decide whether to release someone on bail after seeing them in court and reviewing their record make worse decisions than an algorithm that only has access to the record. While this also highlights the importance of focusing on the information that actually makes a difference (in this case, the criminal record), it’s also quite different in that it illustrates the superior performance of algorithms, at least when there is a lot of training data available.
One important question about AIs is under which circumstances they outperform humans when there isn’t any training data. Which situations are sufficiently new so that the vast amount of computing power and training data that the AIs have access to becomes useless?
Nehaveigur 