Spotify Rolls Out AI-Prompted Playlists
Spotify is shifting its recommendation architecture toward user co-piloting by launching AI-powered Prompted Playlists. The feature, now available to Premium users in the UK, Ireland, Sweden, and Australia, allows users to generate playlists from written prompts or emojis. The underlying system integrates LLMs for prompt understanding with existing recommendation models, incorporating explicit user feedback directly into the discovery process.
- Spotify's recommendation engine is built on a three-layered architecture of data, models, and experience, which has evolved beyond traditional collaborative filtering. This new feature incorporates Natural Language Processing (NLP) for prompt understanding alongside content-based filtering that analyzes raw audio for metrics like danceability and loudness. - The company domain-adapts open-weight LLMs by training them on "Semantic IDs," which are compact, catalog-native identifiers representing the relationships between content and users. This approach teaches the model to "speak Spotify" and reason about its specific catalog and listener behaviors rather than relying on general text knowledge. - To enable rapid iteration without compromising stability, Spotify's engineering teams maintain separate systems for personalization and experimentation. Personalization pipelines are optimized for low latency and high availability, while experimentation systems prioritize accuracy and traceability for A/B testing ranking logic and new features. - The AI considers a user's entire listening history, going back to their first day, to generate the playlists, aiming to reflect the full arc of their taste. Users can then refine the generated results with conversational follow-up commands like "more pop" or "less upbeat" to further tune the output. - The feature was first tested with Premium users in New Zealand in December 2025 before expanding to the U.S. and Canada in January 2026. During the beta period, some users have reported encountering usage limits after generating 20 to 30 prompts. - This launch competes with similar offerings like YouTube Music's "Ask Music," which also uses text prompts to generate a radio station that can be saved as a playlist. The underlying industry trend involves using generative models to treat user actions (clicks, listens, skips) as a sequence of tokens, similar to words in a sentence, to predict the next action. - The development of such features is accelerated by internal tooling like "Honk," an AI system leveraging Anthropic's Claude Code that assists Spotify engineers with coding and deployment. This system is credited with significantly speeding up the release of over 50 new features and updates in 2025.
