Chess is the first domain in history where an ordinary person can train daily with a superhuman intellect — for free, at any hour, without an appointment. The interesting question is no longer whether the machine plays better than we do. It is what living beside that machine does to the way a human being learns.
This three-part series examines what peer-reviewed research, large-scale game data, and cognitive science actually say about engine-assisted development — the measured gains, the measured costs, and the working method that separates learning from an engine from merely reading its answers.
The evidence at a glance
The research base is thinner than the engine's importance suggests — one strong quasi-experiment, one enormous observational study, and a handful of mechanism studies from chess cognition and neighbouring games. This is the map; the terrain comes in Part 1.
| Study | Design | Headline finding |
|---|---|---|
| Gaessler & Piezunka, 2023 (Strategic Mgmt Journal) | Natural experiment; 19,285 players, >500,000 games, 1970–2000 | Engine access ≈ +11 Elo/year — but players became worse at exploiting human mistakes |
| Bilalić, Graf & Vaci, 2024 (British J. of Psychology) | 11.6 million elite decisions, 1985–2021, scored vs. Stockfish | Move quality rose steadily for decades; the sharp gains were among juniors in the late 1990s |
| Shin et al., 2023 (PNAS — Go) | 5.8 million professional Go decisions, 1950–2021 | After superhuman AI, humans played better and more novel moves |
| Charness et al., 2005 | Deliberate-practice survey of tournament players | Serious solitary study — not instruction — was the strongest predictor of skill |
| Moxley et al., 2012 | Tactical-decision experiments across skill levels | Extra unaided deliberation improved decisions for experts and weaker players alike |
| Sala & Gobet, 2016 | Meta-analysis of 24 school-chess studies (5,221 students) | Modest academic transfer (g ≈ 0.34), weak controls — far-transfer claims stay unproven |