Spotify’s AI Playlists: From Discover Weekly to Adaptive Mixtapes

AI · 6 min read

Discover Weekly started as a collaborative filtering milestone, using user-item co-occurrence to surface relevant music. Over time, Spotify layered content-based features — audio embeddings, mood classification, and metadata — to handle cold-start and long-tail recommendations. Current AI playlists combine sequence models with session-aware signals to create adaptive mixtapes that respond to immediate context, like time of day or activity.

Spotify’s experimentation culture shows in its A/B testing of playlist formats and control mechanics. Small nudges — like a ‘more like this’ slider or smart shuffle — are gated experimentally to measure retention and downstream artist discovery. The UX integrates AI gently: playlists are labeled with rationale (e.g., “For your morning run”) and allow user shaping without exposing raw model outputs.

Operationally, Spotify uses real-time pipelines to update session-level features and offline models for longer-term preferences. The result is a hybrid system that balances serendipity with predictable tastes, while keeping creator fairness and transparent controls as ongoing product challenges.