Strip the fifty-year argument down and one rule survives every study in this series: whatever skill you are trying to build, perform it yourself first, then spend the machine on verification. An engine opened before you have generated your own candidates, calculation, or evaluation is not feedback — it is a substitution for the exercise itself.
The golden ruleSeparate generation from verification
Context decides whether engine-first is right or wrong. If the goal is opening research — checking whether a line is objectively sound — immediate engine use is entirely sensible; the task is the machine's. If the goal is calculation, evaluation, candidate generation, or practical decision-making, opening the engine first quietly deletes the rep you came for. The research of Part 2 is the warrant for this: deliberation measurably improves decisions, serious solitary study predicts skill, and a novice's working memory is a finite budget.
A useful operational test is to measure a session by questions answered, not hours spent. "Which defense did I fail to consider?" "Is this sacrifice actually sound?" "Which of my three candidates was misevaluated, and why?" — these are engine questions with learning on the other side. Watching an eval graph scroll while you click through a game is not a question; it is consumption.
The working methodThe human–engine–human loop
The protocol below is evidence-informed, not experimentally proven — no trial has tested this exact sequence. But every step is anchored in a result from the first two parts: the natural experiment's substitution warning, the deliberation finding, the deliberate-practice data, and the novice working-memory constraint.
The loop, step by step
Post-game analysis that converts machine accuracy into human skill.
- Replay the game engine-off; record your candidates, evaluations, and where you first felt unsure
- Mark the genuinely critical positions — two or three, not twenty
- Open the engine only there; check your candidates, not just the move played
- Ask why: what you missed, why you missed it, and what cue should have flagged it
- Compress the discovery into one retrievable sentence, not a fifteen-move line
- Save the position and retest yourself unaided — a day, a week, a month later
The crucial stage is the fourth. "I missed Rxf7" is a description; "I rejected the sacrifice before checking the forcing replies" is a diagnosis — and only a diagnosis changes what you do next time. This is the difference between accumulating engine corrections and building expertise. The retest matters just as much: recognizing a line you saw yesterday is not the same skill as producing the idea unaided a month from now, and only the second is the kind of knowledge that shows up over the board.
One more entry belongs in any sane protocol because the strongest study in this series demands it: keep human games in the diet. Players who substituted machine practice for human competition became worse at exploiting human mistakes. The skill of giving a fallible opponent difficult problems to solve — practical, time-pressure, emotionally loaded problems — is only trained against someone who can blunder.
By levelSame engine, different relationship
The right ratio of human reasoning to engine feedback shifts with strength, because the learner's bottleneck moves. The evidence is consistent on the direction: benefits were largest when the machine was stronger than the learner — and explanation demands are heaviest when the learner is weakest.
| Player | Highest-value uses | Main risk | The guardrail |
|---|---|---|---|
| Beginner | Checking blunders after a game; mating patterns; basic endgames | Numbers without concepts; overload | Attempt first, engine hidden; one or two lessons per game, in words |
| Club / intermediate | Self-analysis audit; tactical extraction; candidate comparison; opening repair | "Click-through" analysis — reading lines instead of reconstructing them | Full unaided annotation first; engine as the auditor, not the analyst |
| Advanced / titled | Deep opening research; engine-vs-engine comparison; endgame verification; opponent prep | Optimizing for engine-perfect lines humans can't play; infinite analysis | Ask "best practical move against my opponent" as well as "best objectively" |
| Coach | Verifying analysis; mining student games for recurring errors; building exercises | Transmitting machine lines the student can't reproduce or understand | Translate engine output into level-appropriate cues, decisions, and tasks |
Two technical habits are worth adopting at every level. For endgames, tablebases offer literal ground truth — every position with seven or fewer pieces is solved — but perfect play there is often deeply unintuitive, so the lesson has to be compressed into a rule you can actually reproduce. For strategy, compare engine families: Stockfish's search-plus-NNUE and Leela's neural MCTS genuinely disagree in some positions, and a position where two superhuman engines prefer different plans is usually a position with a human lesson in it.
The lines not to crossWhere preparation ends and cheating begins
The ethics here are not subtle. Engine use in preparation and post-game study is legitimate; unpermitted engine assistance during a game is cheating — full stop. FIDE maintains dedicated Anti-Cheating Regulations and device restrictions in its Laws of Chess, and the major platforms police at industrial scale. The rules differ by event — correspondence play has long permitted engines — so the governing document is always the specific event's regulations, not a habit borrowed from another format.
Two softer boundaries deserve equal care. First, evidence: strong human play naturally overlaps with engine choices, so a high "engine match" percentage on a few moves is not proof of cheating — accusations belong to formal fair-play processes, not social media. Second, honesty in study: a student who submits engine analysis as their own work isn't breaking tournament rules, but they are quietly skipping the learning the assignment was for. Coaches should say explicitly when the engine may be opened and what must be recorded first.
The open questionsWhat science still can't tell you
For a technology this central to modern chess, the honest list of unknowns is long. Nobody has run the definitive trial: comparable players randomized to human-only analysis, engine-first analysis, human-first/engine-second analysis, and a human-like AI tutor, then tested unassisted months later. Nobody knows the right dose — five minutes of checking after a serious game and six hours of line-reading are both called "using engines." Nobody has measured retention with the tool withdrawn, quantified creativity in chess the way the Go study did, or settled whether a skill-matched engine that errs like a human teaches better than an oracle that never errs.
Until those studies arrive, the defensible position is the one this series has built: engines are among the most powerful training technologies ever introduced to a competitive mind sport, and an oracle is not automatically a teacher. The evidence rewards a specific posture — use engines to falsify, verify, expand, and challenge your thinking, not to eliminate the need for it.
The whole series
- Part 1 — The Evidence: engines made us stronger, with a measurable substitution cost.
- Part 2 — The Mechanism: why an oracle is not a teacher — pattern libraries, deliberation, and the load problem.
- Series overview: the six landmark studies compared in one table.
Sources for this part
- FIDE Handbook — Anti-Cheating Regulations and the FIDE Laws of Chess on outside assistance and devices.
- Stockfish — official site and documentation; Leela Chess Zero — open-source neural engine.
- Lichess — the completion of the seven-piece Syzygy tablebases.
- McIlroy-Young, R., et al. (2020) — Maia: human-move prediction at specific skill levels; Maia-2 (2024) — a unified model for human-AI alignment in chess.
- Gaessler, F., & Piezunka, H. (2023) — the substitution cost behind "keep human games in the diet".
- Moxley, J. H., et al. (2012) — deliberation improves tactical decisions; Charness, N., et al. (2005) — serious study predicts expertise; de Bruin, A. B. H., et al. (2005) — novice self-regulation limits.
- Chess.com Fair Play — how detection works and current closure statistics.