What the player piano can teach us about creative control in AI music

Picture a piano in a drawing room, its keys falling without fingers on them. The notes have been predetermined by holes punched into a long paper roll. Yet the person at the instrument is still pumping pedals, changing tempo and shaping the performance. For anyone arguing about whether artificial intelligence makes a musician unnecessary, the player piano offers a century-old complication: automating the note sequence was never quite the same as removing musical decisions.[2][3]

That history is the subject of a Being Human Festival talk at the Guildford Institute on 11 November, from 7 to 8 p.m. Composer and music technologist Milton Mermikides will trace a route from programmable automata to player pianos, electronic instruments and AI. The event is part of Surrey’s 5–14 November programme; booking is required.[1][4]

Where the performance resides

The London Museum’s record for a circa-1914 Pianola explains the division of labour. A perforated roll determines the notes, foot pedals operate the pneumatic mechanism and a lever lets the player alter the tempo. Familiarity with the music still matters because the operator must anticipate changes.[2]

Reproducing pianos pushed the idea further by preserving aspects of a pianist’s performance. Europeana describes how recordings were encoded on paper rolls and later played through compressed-air mechanisms. Famous musicians supplied rolls; roughly 30,000 titles circulated across classical, operatic and popular music.[3]

The machine therefore redistributed musical decisions. A roll embodied preparation elsewhere, the mechanism executed it and—depending on the system—the person at the instrument could still influence the result. Composition, encoding and interpretation became separable activities.[2][3]

A useful question for an AI demonstration

The same distinction is essential in the AI era. A generative music system may propose a melody, harmonisation, rhythm or complete recording. The human participant may choose training inputs, specify the musical direction, reject material, alter the arrangement, perform with the result or do no more than press Generate. These activities differ in skill and authorship; it makes little sense to give every resulting recording the same story about “AI creativity.”[1][2][3]

For a contemporary demonstration, the revealing question is where the choices remain editable. Can a person deliberately alter a phrase or its timing? Can they distinguish their contribution from a supplied performance or generated material? A finished recording alone conceals decisions that a demonstration of the process could make visible.

The comparison has limits. A paper roll stores explicit note events; a generative model produces new outputs from learned statistical structure. The technologies are not equivalent. What the older instrument supplies is a sharper vocabulary for asking who controls which parts of making and performing music.

The Pianola’s tempo lever is a small but revealing detail. Mechanical execution did not eliminate interpretation. In AI music, the equally revealing details may be the controls that let a person shape the result after the machine has begun.[2]

Report an error or suggest a correction