Research & Innovation

Carnegie Mellon Put AI-Generated Music to a Blind Test. Humans Still Won on Creativity.

CMU's Heinz College found AI-assisted compositions were judged less creative by listeners, and used slower tempos and fewer notes than human-made tracks — a rare hard data point in a debate usually argued from vibes.

A February 2026 study from Carnegie Mellon University’s Heinz College found that, despite rapid advances in AI music generation, human composers still lead on creativity by measures that hold up under controlled listening tests. AI-assisted music in the study was slower, used fewer notes, and was judged by listeners as less creative than comparable human-made tracks — a concrete, measured finding in a debate that’s mostly been argued through anecdote and demo reels on either side.

What the study actually measured

Rather than asking listeners a vague “does this sound good” question, the CMU research isolated specific, comparable production choices — tempo and note density — alongside direct creativity ratings, giving the finding more substance than a simple preference poll. The result: AI-assisted compositions consistently trended toward simpler, slower output, and listeners picked up on that difference even when not told which tracks were AI-involved, rating those compositions as less creative on average.

Where AI is actually proving useful instead

The more interesting finding sits alongside the headline result: a parallel research direction around AI Creativity Support Tools (AI-CSTs) — systems built to assist rather than replace a human composer’s process — is showing real traction. Tools like the Multi-track Music Machine for controllable multi-track MIDI generation, Music Transformer for melody harmonization, and the Anticipatory Transformer for infilling musical phrases all point toward a different deployment model than end-to-end generation: AI handling a specific, bounded sub-task inside a human-driven composition process, rather than producing a finished track from a text prompt.

Why the distinction matters for working musicians

The practical takeaway isn’t “AI music is bad” — plenty of 2026’s text-to-music platforms produce technically polished, commercially usable output. It’s that creativity specifically, as distinct from technical polish or stylistic competence, remains a measurable human advantage right now, and that the more promising near-term application of AI in music sits in the assistive, sub-task tooling this research points toward rather than in full autonomous composition. For musicians deciding how much of their process to hand off to AI tools, that’s a genuinely useful distinction — and a more grounded one than either “AI is coming for your job” or “AI can’t really create anything” tend to be.