ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
News3 min read·2026-10-20DisconfirmAI Editorial

Baroque Music Blind Test: Professional Musicians Hit Only 63.8% Accuracy, Perception Bias Five Times Larger Than Actual Quality Difference

Auggie organized a 118-person baroque music blind test asking listeners to judge whether 16 tracks were generated by Claude Opus 5.5 or GPT-6 Astra or composed by real baroque masters. Overall accuracy was only 55.7%, with professional musicians reaching 63.8%. Claude's Fugue in A minor tied with Bach's Trio Sonata for the most 'Human' votes. The more critical finding: listeners rated tracks they believed were AI by 1.02 points lower than those they believed were human, yet the actual enjoyment gap between human and AI tracks was only 0.18 points.

ImportanceMaterialEvidenceE3 inspectableWrite-upStandard

Original publication time 2026-10-20 20:52 GMT+8

In October 2026, Substack author Auggie published a results analysis of a baroque music blind test on LessWrong. He selected 8 real baroque works and 8 same-style tracks generated by Claude Opus 5.5 and GPT-6 Astra, presented them via LilyPond symbolic rendering to 118 listeners who completed all questions, and asked them to judge each track as human or AI.

Overall accuracy was 55.7%, less than six percentage points above the 50% expected from random guessing. Professional musicians performed slightly better at 63.8%, but still got more than a third of their judgments wrong. Listeners with no musical training scored 48.4%, below random chance. A counterintuitive finding: listening to classical music more frequently did not consistently improve accuracy—weekly listeners (61.2%) outperformed daily listeners (57.2%), a discrepancy the author himself could not explain.

The most striking individual result: Claude Opus 5.5's Fugue in A minor tied with Bach's Trio Sonata in C major for the most 'Human' votes, each receiving 85 (72%). Meanwhile, real composer Buxtehude's chorale prelude was judged AI by 66.9% of listeners, including a 64% misidentification rate among professional musicians. The author speculated that the open fifths and fourths at the excerpt's ending may have sounded too unfamiliar to modern ears, paradoxically making them read as an AI artifact.

Sign in / Sign up
ReadingResearchRadarInvestment framework

However, the finding with the most transferable value lies not in discrimination ability itself but in how perception bias distorts aesthetic judgment. Listeners gave tracks they believed were human-made an average enjoyment rating of 3.50, while tracks they believed were AI received only 2.48—a gap of 1.02 points. But the actual average enjoyment difference between human-made and AI tracks was only 0.18 points (3.07 vs. 2.89). In other words, the belief 'is this AI-made?' influences enjoyment ratings roughly five times more than actual quality differences. The author concluded: 'Do I like it?' is not the best question to ask when trying to determine a piece's authorship.

The author framed his conclusion as follows: in the genres where AI models are strongest, they are almost indistinguishable from the average human composer. He called this passing the 'musical Turing test.' As a self-described 'lifelong member of Team Humanity,' he admitted that humanity took a drubbing this time.

Boundaries to preserve: this is a self-organized experiment without peer review, with a sample of 118 and some subgroups extremely small (only 3 respondents reported never listening to classical music). Baroque music relies heavily on harmonic rules and counterpoint technique, which may coincide precisely with language models' structured-generation strengths; this result cannot be directly extrapolated to improvised jazz, experimental electronic music, or styles requiring long-term human emotional accumulation. Additionally, tracks were presented via LilyPond symbolic rendering rather than real instrument recordings, which both lowered production barriers and may have affected listeners' judgments of timbral authenticity.

Sources

  • 原始出处lesswrong.com↗
Continue researching →