How AI Decides What Music You Hear: Spotify, Algorithms, and the Future of Sound

I'm Parikshet and I want to talk about the AI system that most of us interact with more than almost any other: the music recommendation engine. Every time Spotify puts a song in your Discover Weekly, an algorithm has made a decision about your taste. Here's how that actually works.

The Three Systems Behind Discover Weekly

Spotify's recommendation AI runs three parallel systems and combines their outputs:

1. Collaborative filtering. Spotify has 600 million users. If 50,000 people who all love Arctic Monkeys, Tame Impala, and The 1975 also recently discovered a new band called Glass Animals — the algorithm concludes: you should probably hear Glass Animals too. It is pattern-matching across millions of listening histories simultaneously.

2. Natural language processing. Spotify's bots scan music blogs, Reddit threads, Pitchfork reviews, and playlist descriptions for the words people use about specific artists. "Dreamy, lo-fi, bedroom pop" — if those terms cluster around an artist, the algorithm creates a sonic identity for them. If your listening habits match that cluster, they appear in your recommendations.

3. Audio analysis. Spotify analyses every track's actual audio: tempo, key, energy, acousticness, danceability, loudness, speechiness, valence (musical positivity). Each song gets a fingerprint of about 30 numerical variables. Songs with similar fingerprints go into the same recommendation bucket.

Discover Weekly combines all three. That's why it feels personal — it is built from your behaviour, others like you, and the actual sound of music.

AI That Makes Music

Recommendation AI is old news compared to what is coming. AI that generates music — original songs, with melodies, harmonies, and vocals — is now genuinely impressive.

Suno AI (2024) generates complete songs in any genre from a text prompt in seconds. "Write a sad country song about a dog leaving" and it produces a full track with a singer, guitar, and drums. Google's MusicLM does the same. OpenAI's Jukebox was an earlier version that generated music in the style of specific artists.

The quality debate is real. Many AI tracks have obvious tells — certain chord progressions that feel too resolved, lyrics that rhyme but say nothing specific. But improvements are coming fast enough that the music industry is taking it seriously.

Is This a Threat to Musicians?

Honestly — yes and no, and I think the honest answer matters more than a comfortable one.

Background music, stock music, jingle production, simple game soundtracks: AI will increasingly replace human composers for these tasks. These are real jobs that real musicians have done for decades.

But concerts, albums with emotional meaning, the story behind an artist, live performances: I do not think AI replaces these. When I listen to a great song, part of what I'm feeling is connection to a human who made it through real experience. An AI cannot have a broken heart, a childhood fear, or a first concert memory — and those things show up in music in ways that I think most listeners can feel, even if they can't name it.

The Algorithm as Gatekeeper

Here is something that troubles me: the Spotify algorithm now has more power over what music succeeds than any music executive ever did. A song that gets added to Discover Weekly for 100 million users will accumulate millions of streams in a week. A song that the algorithm deprioritises may never be heard at all — even if it is extraordinary.

This is a concentration of power in one company's codebase that affects every musician on Earth. Whether that is good or bad depends on whether the algorithm rewards quality or just engagement — and those are not always the same thing.