How Spotify Knows Your Taste: The AI Behind Every Perfect Playlist

Every Monday, Spotify generates a personalised 30-song Discover Weekly playlist for each of its 600 million users. Each playlist is different. Many users say it's better than anything a human DJ has curated for them. How does an AI develop taste that personal?

I'm Parikshet. Music recommendation was one of the first places I noticed AI doing something that seemed almost magical. The first time Discover Weekly recommended a song I'd never heard that immediately became one of my favourites, I wanted to understand how it happened. Here's what I learned.

Three AI Systems Working Together

Spotify's recommendation system isn't one model — it's three distinct approaches combined:

1. Collaborative Filtering (who you're like)
Spotify has listening data from 600 million people. Collaborative filtering finds users whose listening history is similar to yours and recommends what they've loved that you haven't heard. If ten million people who share your specific taste in music all discovered and loved the same obscure band, Spotify will suggest that band to you.

2. Natural Language Processing (what people say about music)
Spotify crawls the internet — blog posts, reviews, articles, social media — and reads what people write about songs and artists. The language used to describe music ("driving beat," "melancholic," "late-night vibes," "perfect for studying") is analysed and turned into a "cultural fingerprint" for each artist and song. This allows Spotify to understand the context and mood of music from human descriptions.

3. Audio Analysis (what the music actually sounds like)
Spotify also analyses the actual audio of songs: tempo, key, energy level, danceability, acousticness, "valence" (musical positivity). This means Spotify can recommend a song it has never received any listening data for — a brand new release — because it can compare the audio properties to what you've enjoyed before.

These three systems together are remarkably powerful. Collaborative filtering finds similar people. NLP captures cultural context. Audio analysis handles new releases. Weighted and combined, they produce personalised playlists that feel almost intuitive.

Why Discover Weekly Often Gets It Right

The key insight in Spotify's approach is that they don't ask you to describe your taste — they observe it. Every stream, skip, save, playlist addition, and repeat play is data. After enough listening, Spotify's model of your taste is more nuanced than anything you could describe in words.

This is similar to the Netflix situation. Revealed preference — what you actually do — is a much stronger signal than stated preference — what you say you like. People often don't know what they like until they experience it. AI that observes behaviour can sometimes predict your next favourite thing before you know you want it.

The Interesting Limitation: Filter Bubbles

Recommendation AI is very good at giving you more of what you already like. This creates filter bubbles — environments where you mostly encounter things similar to what you've already seen or heard. For music, this might mean you stop discovering genuinely different styles. For news, this is a much more serious problem.

Being aware of this tendency — and deliberately seeking out things outside your algorithm's bubble — is a habit worth developing now. Good taste comes partly from broad exposure, and algorithms optimise for engagement, not breadth.

Frequently Asked Questions

How does Spotify's Discover Weekly work?

Three AI systems combined: collaborative filtering (similar users), NLP analysis of what people write about music, and audio analysis of the songs themselves.

What is a filter bubble?

When recommendation AI keeps showing you content similar to what you've already seen, narrowing your exposure to new or different ideas.

How does Spotify recommend brand new songs?

Audio analysis — Spotify analyses the actual musical properties of new songs (tempo, key, energy) and matches them to what you've enjoyed before, even without listening data.