Spam filtering has quietly used AI for over two decades, well before “AI” was the fashionable term for it — tracing its evolution shows a genuinely continuous arms race between filtering techniques and increasingly adaptive spammers.
Bayesian Filtering: The Original Breakthrough
Paul Graham’s 2002 essay “A Plan for Spam” popularized Bayesian filtering — calculating the probability a message is spam based on the individual words it contains, learned from a training set of known spam and legitimate (“ham”) messages. This was a genuinely significant improvement over the earlier, more brittle rule-based filters (which relied on manually maintained keyword blocklists spammers could trivially evade with minor wording changes).
Escalating Sophistication
As Bayesian filtering became widespread and effective, spammers adapted directly in response — word obfuscation (“V1@gr@”), image-based spam (avoiding text-based filtering entirely), and increasingly sophisticated phishing genuinely designed to evade automated pattern detection. Filters responded with more sophisticated machine learning models incorporating far more signals: sender reputation, embedded link analysis, genuine image content analysis, and behavioral patterns beyond just message text.
Modern Deep Learning-Based Filtering
Today’s spam and phishing detection uses neural networks analyzing genuinely dozens of combined signals simultaneously — message content, sender history and reputation, embedded link destinations, and organizational-level patterns across an entire email provider’s user base — a direct, continuous evolution of the same fundamental Bayesian idea from over two decades ago, scaled up substantially with far more data and far more sophisticated modeling.