Paul Graham however definitely championed and popularized the idea of Bayesian filtering.
http://www.catb.org/~esr/bogofilter/bogofilter.html
https://en.m.wikipedia.org/wiki/Apache_SpamAssassin
Spamassassin's Bayesian classifier credits Graham.
https://metacpan.org/release/Mail-SpamAssassin/source/lib/Ma...
Also, Yahoo! Mail implemented at least partly Bayesian based filtering not long after the original PG essay.
I'm guessing GMail is using some kind of statistical filter for not only spam, but general categorization, though as others have pointed out GM is pretty aggressive in filtering. I think that GM might be using collaborative data, aggregating across user behavior and I believe many users deal with signups they're tired of by marking as spam, sharing that behavior across the userbase would cause things that you willingly signed up for to be filtered out.
Cool circa-2002 game for Unix nerds: you have 60 seconds to telnet to port 25 and generate the highest spamassassin score.