Published: 10/14/2026
Source: arXiv — MuseBench: A Benchmark for AI-Generated Music Detection and Artifact Analysis
Researchers introduced MuseBench, a 361,166-track benchmark designed to test whether automated systems can distinguish human music, fully generated tracks and human-AI co-creations under realistic conditions.
The dataset covers ten commercial and open-source generators, including Suno and Udio, plus human-composed music. It also includes tracks altered through compression, acoustic processing, noise and adversarial perturbations to test whether detection survives normal distribution and deliberate evasion.
The researchers evaluated 11 detectors and documented weaknesses when systems encountered unseen datasets, co-created recordings and transformed audio. They also identified recurring spectral, harmonic, vocal and physically implausible performance artifacts.
This is a newly released research preprint, not a final industry standard or peer-reviewed proof that every detector will behave the same way in commercial use.
Can AI music detectors reliably identify every Suno or Udio song? No. MuseBench shows why performance must be tested across models, genres, co-created tracks and post-processed audio rather than summarized with one accuracy number.
Why it matters: Streaming services should measure false positives as carefully as successful detection. A system that wrongly labels human or hybrid music can affect discovery, payment and reputation even when its overall accuracy appears high.