Eno set rules running in 1996 and called the result generative music

What happened

On 8 June 1996, at the Imagination Conference in San Francisco, Brian Eno gave a talk that a published transcript records under the title "Generative Music". In it he described a system from the company SSEYO, called Koan, that used around 150 probabilistic rules governing scale, harmony and note movement to produce music that, in his words, he had "never heard it play exactly this before". That system underpinned Generative Music 1, released the same year. Two decades on, Eno's own site, as it reads on 16 September 2026, describes a further step, the Reflection app released in 2017, as "the most sophisticated" version of the same idea, freed from the length limits of a physical disc or file.

What the documents say

The 1996 transcript states the mechanism plainly: classical composition "specifies an entity in advance", while generative music "specifies a set of rules and then lets them make the thing". The output is unpredictable and, by Eno's own description, unrepeatable and sensitive to its starting conditions. The Reflection page carries the idea forward in Eno's own words: "I wanted also that this music would unfold differently all the time - 'like sitting by a river': it's always the same river, but it's always changing." Unlike a CD, which the page says "replay identically each time you listen", Reflection "creates an endless and endlessly changing version of the piece of music". Neither document claims the system composes with intent; both describe rule-following, not judgement.

Why it matters for makers

The distinction worth carrying into a conversation about a modern AI model is the one Eno draws between a rule system and a trained model: Koan's behaviour came from rules a person wrote and could inspect, not from patterns learned from other people's recordings. Calling both "generative" flattens a real difference in how output is produced and where its influences come from. For a maker using a modern generator, naming which kind of system is at work, hand-written rules versus learned weights, changes what questions about originality even apply.

What to check before you use it

Before describing a tool as "generative" in Eno's sense, check whether its output comes from explicit rules a person set, which can usually be described and reasoned about, or from a model trained on examples, which usually cannot be fully inspected. Eno's own materials describe the former case only; they make no claim about systems trained on other artists' recordings. This is an editorial distinction, drawn from reading both documents together, not a claim either makes about later technology.

• Does the system follow explicit, inspectable rules, or patterns learned from training examples?
• Is the output described as unrepeatable by design, or is repeatability actually a feature being sold?
• What did Eno's own terminology mean in 1996, and does a modern tool's marketing use "generative" the same way?

Eno's own record, a talk transcript and a current app page thirty years apart, is useful mainly for keeping the vocabulary honest: neither system described here learned from other people's recordings, and the word "generative" named rule-following decades before it became a byword for something quite different.

Continue in Listening & research, or Soundraw and Beatoven license background music, not a master.