Email AI recommendation engines launched in 1994, 30 years before ChatGPT

The first AI recommendation engines launched in 1994 through email, predating modern AI by three decades and using crowdsourced human preferences to help users discover music, movies, and news. These early systems like Ringo, SIFT, and Bellcore’s movie recommender relied on “social filtering”—the principle that people with similar past preferences would likely agree on future choices—and operated entirely through email interfaces that users came to trust as intelligent agents.

The big picture: Before Spotify algorithms or Netflix suggestions, email-based AI systems were already solving information overload by harnessing collective human wisdom to make personalized recommendations.

How it started: MIT researchers identified a “deceptively simple idea” in 1994: people who agreed in the past are likely to agree again.

• As the web exploded from 10 sites in 1992 to over 10,000 by 1994, researchers recognized that “knowledge shall increase” was becoming more of a content overload problem than a blessing.
• Stanford’s Tak W. Yan and Hector Garcia-Molina wrote that “The exploding volume of digital information makes it difficult for the user, equipped with only search capability, to keep up with the fast pace of information generation.”
• Social filtering emerged as the solution: aggregate human preferences to predict what individuals might like based on others with similar tastes.

Why email became the interface: Email offered universal accessibility at a time when storage cost $4,000 per gigabyte and network connectivity was limited.

• Stanford’s SIFT team chose email because it was “the lowest common denominator of network connectivity,” accessible even behind firewalls and on less powerful machines.
• Users would sign up via web forms, receive curated content via email, and reply with their preferences to train the system.
• SIFT attracted over 1,000 profiles within ten days of launch and was matching 45,000 articles weekly to 13,000 subscribers by November 1994.

The movie recommender that started it all: Bellcore researchers launched the first social filtering system for movies from October 1993 to May 1994.

• Users would email “ratings” to [email protected] and receive a list of 500 movies to rate on a 1-10 scale.
• The system would find correlations between users’ ratings and recommend unseen movies based on people with similar tastes.
• It functioned as both a recommendation engine and nascent social network, connecting users with shared preferences.

Ringo: The music AI that felt human: MIT’s Ringo launched July 1, 1994, specifically targeting music recommendations with a personality-driven approach.

• Users rated 125 artists from 1 (“Pass the earplugs”) to 7 (“BOOM! One of my FAVORITE few!”), and Ringo would recommend eight new artists based on similar users’ preferences.
• Despite early inaccuracies, users embraced Ringo as a friend, with one saying: “I was flattered” when told there weren’t enough similar participants.
• Science fiction writer Cory Doctorow later recalled that “half the music in my collection came out of Ringo.”

What they were saying: Users developed emotional connections to these early AI systems despite their limitations.

• Dave Dell initially called the concept “****in’ moronic” but admitted “I’ll try it anyways.”
• Joe Morris praised Ringo: “RINGO always came up with a great selection. I got turned on to a bunch of stuff I would of never found otherwise.”
• One user appreciated even negative results: “It said there weren’t enough participants out there like me. I was flattered.”

The lasting impact: These email-based systems established principles that still power today’s recommendation engines.

• Social filtering concepts evolved into Google’s PageRank, Facebook’s feed algorithms, Netflix suggestions, and Spotify’s Daily Mix playlists.
• Ringo later became Firefly, which Microsoft acquired for $40 million in 1998, with its email login system eventually morphing into Microsoft Passport.
• As Cory Doctorow observed: “I am largely of the opinion that great AI consists of aggregated human decisions, not machine generated decisions.”