Email was the user interface for the first AI recommendation engines (buttondown.com)

In the early 1990s researchers launched the first practical recommender systems not as web apps but as email bots that crowdsourced human judgments to filter information. Projects like Xerox PARC’s Tapestry (endorsements for newsgroup posts), Stanford’s SIFT (email newsletters that stored replies as votes), Bellcore’s videos@bellcore (rate movies by email and get correlated recommendations), and MIT’s Ringo (rate 125 artists, get eight suggested artists) turned people’s ratings into recommendations. These services processed thousands of profiles and tens of thousands of items (SIFT matched ~45,000 articles weekly to 13,000 subscribers) and used simple, effective statistical methods — e.g., constrained Pearson correlation to weight similarity between users — to predict what a given user would like. For the AI/ML community this history matters because it shows collaborative (social) filtering and human-in-the-loop design solving real discovery problems long before modern deep learning. Key technical ideas — treat user ratings as votes, compute pairwise user similarity, predict unseen items by weighted aggregation — are the foundation of today’s matrix-factorization and neighborhood-based recommenders. The email UI was crucial: universally accessible, low-bandwidth, and familiar, it lowered adoption barriers and helped these systems acquire the labeled data they needed. Early issues — cold starts, sparse overlaps, and noisy human labels — remain active research challenges, but these email-era experiments proved the viability and social dynamics of recommender systems that now power content, commerce, and social platforms.