Human diffusion · 21

How an AI joseki moves from experiment to human habit

In a public professional-game sample, early bare-4-4 3-3 rose from 9.72% in 2017 to 40.06% in 2018-2020. Adoption rates show that more players used it; whether they also learned to exchange less, tenuki, and return has to be read from the games themselves.

1. The first game says only that someone tried the risk

   DeepMind named early 3-3 among AlphaGo's representative innovations in 2017, and professionals soon tested it in tournament play. But a first appearance begins the story rather than finishing it: a novelty can be curiosity, preparation, or a one-game probe. A changed habit requires different players to keep making the same structural choice in comparable positions.

2. By the next period, early 3-3 was no longer a performance by a few players

   Read19 deduplicated 96,058 public SGFs and searched even 19x19 games for a direct 3-3 invasion under a bare 4-4 in the first 40 moves. Under that conservative definition, adoption rose from 9.72% in 2017 to 40.06% in 2018-2020 and 46.92% in 2021-2025. The archive cannot supply one exact worldwide adoption date, but it clearly separates an experiment from a durable population habit.

3. Coordinates spread before the timing idea behind them

   The real change was not the C17 coordinate but the meaning of a finished joseki. Older habits favored settling the corner continuously. Newer play often leaves after fewer exchanges, preserves cuts and aji, and returns only after whole-board conditions change. Copying the opening moves while mechanically completing them at the wrong time is not yet learning the idea.

4. Professional adoption does not mean every player moved at once

   Professionals gained early access to analysis and watched one another's records closely, so a new line could spread through tournaments within a season. Amateur learning followed another route: first the coordinate, then streams, joseki tools, and books supplied the rationale. This public sample mainly describes professional choice; it cannot stand in for the speed at which ordinary amateurs understood the idea.

5. More new moves and fewer common patterns can happen together

   A study of more than 5.8 million professional decisions finds that post-AI novelty rose and was associated with better decisions. Another study of about 70,000 games and 15 million moves finds that common opening patterns became more concentrated. The first asks whether one move is new; the second asks whether a population repeats the same repertoire. Players can innovate in detail while sharing a narrower opening grammar.

6. A clean curve can still turn archive bias into an apparent era break

   CWI warns that event coverage is uneven. Player composition, rules, komi, time controls, and archive coverage also change across years. The safest reading is therefore narrow: under one conservative definition inside one archive, early 3-3 made a large jump. The curve does not prove that AI directly caused every use or that every player understood the same rationale.

7. Real adoption has to survive three different kinds of evidence

   Appearance asks who first brought the move into serious play. Adoption asks whether many players repeated it long enough to outlast fashion. Understanding asks whether their later choices adjust tenuki and return timing when ladders, outside strength, or initiative change. Early 3-3 clearly passes the first two layers. The third still needs continuation records, teaching history, and learning experiments.