There's a big (and common) error in statistical reasoning Google is making with the decision to down-weight GPA based on their data: That GPAs do not predict performance among those it hired does not imply that Google should not use GPA when hiring any more. Rather, it means that Google used GPA to exactly the right extent among those it hired under its old policy - there was no information left in GPA they didn't use, and therefore they should leave whatever policy they have in place as is.
Explanation: suppose there are only two things Google observes, GPA and coding ability, and that Google uses some correct decision rule to only hire those people where the sum GPA + coding ability > some threshold. Those who have lower GPAs will thus tend to have higher coding ability, otherwise they wouldn't have met the threshold to make it into the pool of hires they're analyzing - and, therefore, comparing "those with low GPAs that Google hired" and "those with high GPAs that Google hired" is not an apples-to-apples comparison.
In order to assess whether GPA should be used at all, they would need to look at how the people they didn't hire because of their existing policy would have performed.
Explanation: suppose there are only two things Google observes, GPA and coding ability, and that Google uses some correct decision rule to only hire those people where the sum GPA + coding ability > some threshold. Those who have lower GPAs will thus tend to have higher coding ability, otherwise they wouldn't have met the threshold to make it into the pool of hires they're analyzing - and, therefore, comparing "those with low GPAs that Google hired" and "those with high GPAs that Google hired" is not an apples-to-apples comparison.
In order to assess whether GPA should be used at all, they would need to look at how the people they didn't hire because of their existing policy would have performed.
More reading: http://beerbrarian.blogspot.com/2013/11/the-subtle-joys-of-s... and http://www.jamesmahoney.org/articles/Insights%20and%20Pitfal...