What emergence actually means

Emergence is what you have when a system does something none of its parts does, and the behaviour is a consequence of how the parts interact rather than of any instruction telling the system to behave that way. It is one of the most overused words in science writing, so it is worth being precise about it.

A usable test

A practical question separates real cases from loose talk: is the behaviour written down anywhere in the rules? If a flocking simulation contains a rule saying "form a flock", the flock is not emergent, it is implemented. If each bird only follows three local rules about its nearest neighbours and a flock appears anyway, the flock is emergent, because no bird contains it and no line of the program mentions it.

By that test, emergence is a relationship between a description of the parts and a description of the whole. It is not a substance a system has more or less of, and it is not a synonym for "complicated" or for "surprising to me".

Standard examples

• Termite mounds. No termite has a plan of the mound. Each responds to local cues, including cues left in the material by other termites, and a structure with working ventilation results.
• Traffic waves. A stop-and-go wave travels backwards along a motorway at a fairly consistent speed. No car is doing that. The wave is a property of the interaction.
• Conway's Game of Life. Four rules about neighbouring cells, and gliders, oscillators and universal computation come out. See cellular automata.
• Prices. Nobody sets a market price. It is a summary of many local decisions made by participants who cannot see it before it exists.

Why it is hard to demonstrate honestly

The failure mode is subtle: an experimenter builds a world, tunes it until something interesting appears, and reports the interesting thing. The tuning is the problem. If the world was adjusted until the result showed up, the result partly reflects the adjustments, and the claim that it came from generic rules is weaker than it sounds.

The usual defences are to fix the rules before running, to declare in advance what would count as a result, to report the runs that produced nothing, and to make every run reproducible so the claim can be checked rather than believed. Those are not niceties. Without them, "emergent" is difficult to distinguish from "I put it there and then forgot".

How this project handles it

Emergent Life takes the strict version of the test. Organisms have no channel that names an object, a species, a threat or a food type. Every sensor is a generic quantity, and every action is a generic effect, which is set out channel by channel in perception and action. There is no fitness function, so nothing is being optimised toward a demonstration.

A ruleset is frozen for the entire life of a universe. Changing anything that would alter how the world unfolds starts a new universe rather than patching the running one, which is what stops the tune-until-it-appears failure mode at the level of policy rather than good intentions. Every run is reproducible from its seed down to the byte, described in determinism and reproducibility.