deadsea - 9:40 pm on Oct 30, 2012 (gmt 0)

If Google is using machine learning behind the algorithms, then YES, the algorithm is finding weird correlations like you suggest and penalizing for them.

When Bayesian filtering for spam first became popular I recall a similar example. The Bayesian filter picked out "FF0000" particularly spammy word. If somebody had an html formatted email with text styled red (with FF0000) it was almost certainly spam.