The Algorithm Failed Music (www.theverge.com)

Music recommendation algorithms that promised to cut through the noise have instead narrowed what we hear. The story traces the arc from Pandora’s early Music Genome (manual tagging of traits like lead vocal gender or guitar distortion) to Spotify’s algorithmic empire built on Echo Nest tech, collaborative filtering from user playlists and machine‑learning analysis of raw audio (e.g., Discover Weekly’s 30‑song mixes). Because platforms optimize for engagement—“time filled” listening rather than musical discovery—they favor safe, repeatable tracks, enable “Perfect Fit Content” ghost artists, and feed labels with listening data that encourages homogenized, hook‑first songwriting (shorter songs, fewer solos, longer albums). That feedback loop has cultural and technical implications for AI/ML and the music ecosystem: recommendation objectives matter—optimize for session length and you compress artistic diversity; feed models industry signals and they amplify sameness. Empirical signs include MIDiA’s finding that heavier algorithm reliance correlates with reduced discovery, especially among Gen Z. The backlash is producing human‑led alternatives (Bandcamp Clubs, editorial Qobuz, college radio, vinyl resurgence) and product pivots toward curated UX, but the likely future is more subtle: algorithmic personalization packaged as serendipity. For ML practitioners, the lesson is clear—design objective functions, training data sources, and interface affordances to preserve serendipity and diversity, not just engagement.