No Free Lunch

Wolpert and Macready had a 1997 paper called No Free Lunch Theorems for Optimization.
This translates to machine learning and states that no algorithm is better than others on all data sets.
It actually goes further and states that no algorithm does better than just guessing on the average data set.
That includes statistics.
So why does anyone learn statistics (or other types of machine learning)?
The key to all of these is that the learned model actually fits the underlying data. If it does, then it's good.