Blind test of 152 listeners: Suno matches human composer, but almost no one can tell which is AI
teropa · x · 2026-10-28
A 2026 NIME paper from MIT Media Lab and Myndstream gave Suno (one-shotted) and a human composer the same briefs for one "calming" and one "upbeat" track, then asked 152 listeners to judge preference and effectiveness across three labeling groups.
Key findings:
- Labels drive preference: listeners prefer what they believe is human-made. When the AI calm track was labeled as human, 72% preferred it — preference tracks the label, not the audio.
- Effectiveness: the human track was judged better at conveying emotion in most groups; only 12% rated the AI upbeat track better in the unlabeled group.
- Poor detection: unlabeled listeners guessed which track was AI correctly only 7/51 times for calm (worse than chance, most guessed backwards) and 29/51 for upbeat.
The author notes the study's limits: only four one-shotted tracks, so little can be concluded beyond that people prefer music they think is human-made.
More from Multimodal
- SenseTime open-sources SenseNova 6.8 Code2Video: video generation as editable code — SarahAnnabels · 2026-10-29
- Dale Labs' Scene generates shoppable images from your taste, with one-click checkout — cneuralnetwork · 2026-10-29
- Miniature birthday-cake mission video goes viral as dreamina prompt workflow spreads — umesh_ai · 2026-10-29
- Insta360 X6 Spatial Capture Turns a 360° Walk Into an Explorable Gaussian Splat Scene — Aiden_Tech_Ai · 2026-10-28
- Vibe-coded Zelda-style 3D game built on kaigen3d engine wows viewers — gdechichi · 2026-10-28
- Nvidia's open Nemotron 3.5 Transcribe leads speech benchmarks with 1/6 the active params — The Decoder · 2026-10-28