The Engine Knows.
The Player Still Has to Move.

The familiar modern objection runs as follows. If a freely available program running on consumer hardware now plays better than the strongest unaided human in any practical analytical setting, then human chess at the highest level is, at best, a slower and weaker version of the same activity. The world’s strongest grandmasters lose evaluation points to a process that fits inside a phone. The contest between two unaided humans as the strongest available evaluators of a position is over, at the elite analytical level. What is left, on this view, is theatre.

The shallow answer to the objection is that humans still enjoy the game, and that enjoyment is enough. That answer is true and insufficient. It treats chess as a hobby that has outlived its own reason, and it does not explain why the world’s best players still prepare for months, still travel, still suffer at the board, still cry after losses, and still command an audience that watches the moves rather than the engine bar. The deeper answer requires distinguishing two questions the popular objection tends to conflate, namely the question of who can produce the strongest evaluation of a position in the abstract and the question of what chess is structurally about as a human contest.

The argument of this essay is that the answer to the first question has changed and the answer to the second has not. Engines now produce stronger evaluations of most competitive chess positions than unaided humans do, in the narrow practical sense of the move that wins more often when both sides play near-optimally. Chess as a sport, however, is best understood not as the activity of producing the highest-evaluation move under near-optimal play, but as the activity of choosing a move under conditions, namely under time, evaluation, consequence, ego, fatigue, and incomplete confidence, and of owning what is set down on the board. The engine has clarified what the strongest practical move tends to be. It has not changed the requirement that the move be playable by a human under conditions, and it does not face that requirement, because the conditions are the contest.

Section IThe old myth of chess met the engine

For most of its modern history, chess sustained a myth that flattered the human mind. The myth was that the best move in a complex position could only be approached, never proved. Annotators argued. Schools of thought rose and fell. A queen sacrifice could be celebrated for a century and then quietly demoted by a later generation that found a defense. The position itself was a kind of unfinished text whose interpretation depended on the strength, taste, and historical situation of its readers.

Strong engines changed this. The progression is well documented. Through the 1990s, dedicated chess hardware and software pushed against world-championship-level human performance, with IBM’s Deep Blue defeating Garry Kasparov in match conditions in 1997 (Campbell, Hoane & Hsu, 2002). Through the 2000s and 2010s, freely available consumer engines on ordinary hardware reached and then exceeded the playing strength of the strongest human players. The open-source Stockfish project, refined since 2008 by a global community of contributors and descended from Tord Romstad’s earlier Glaurung engine, became the dominant alpha-beta-search engine of the period (Romstad et al., Stockfish project documentation). In late 2017, DeepMind’s AlphaZero introduced a different approach, namely deep reinforcement learning through self-play guided by Monte Carlo Tree Search, and produced strong and stylistically distinct play against the prior generation of engines, a result published in peer-reviewed form in Science in December 2018 (Silver et al., 2018). The open-source Leela Chess Zero project subsequently demonstrated that the AlphaZero approach was reproducible by a distributed community. In August 2020, Stockfish merged the NNUE (Efficiently Updatable Neural Network) static evaluator originally developed for shogi by Yu Nasu in 2018, and Stockfish 12 was released the following month with substantially higher playing strength than its predecessor; in July 2023, Stockfish 16 removed the hand-crafted evaluation entirely in favour of a fully neural-network-based evaluation (Stockfish project blog).

The cumulative consequence is that, for most non-trivial competitive positions, an accessible engine now produces, within seconds and on consumer hardware, the strongest available practical evaluation of the position. This does not amount to metaphysical truth about the position, and it should not be described that way. Engine evaluations depend on search depth, time control, hardware, opening book, tablebase availability, engine version, evaluation settings, the type of position, and on whether the indicated continuation is in fact playable by a human at the relevant time control; engine evaluations also continue to change at greater depths, and different strong engines occasionally disagree. The honest framing is that engine analysis has changed the burden of explanation. A human annotation that disagrees with strong current engine analysis now carries the burden of explanation: perhaps the engine line is impractical, perhaps the evaluation changed at greater depth or with a stronger engine, or perhaps the human annotation simply missed something.

This is the modern condition under which the human game now occurs. There is no current evidence that unaided human playing strength will close the gap with top engines, and the assumption that it will not is consistent with two decades of progression in computer chess. The relevant question is not whether to accept the condition. The relevant question is what it changes about what chess is, and what it does not change.

Section IIWhat a strong engine computes, and what it does not

It is worth being precise about what an engine actually produces, because the precision matters for the rest of the argument.

A modern chess engine evaluates positions through a search procedure combined with an evaluation function. Stockfish, the dominant open-source traditional-architecture engine, performs an extremely deep alpha-beta search; since its 2020 NNUE integration the static evaluator has been a small, efficiently-updatable neural network, and since the July 2023 release of Stockfish 16 the hand-crafted evaluation has been removed entirely. The AlphaZero family, including the open-source Leela Chess Zero, performs Monte Carlo Tree Search guided by a deep neural network trained through self-play reinforcement learning. The architectures differ. The product, in the sense relevant to this essay, is similar. Within seconds of being given a position, a strong engine on consumer hardware produces a numerical evaluation, an indicated principal variation, and a recommended move that, in most practical cases, is at least as strong as the move an unaided human at any rating level would identify under similar time conditions.

What the engine produces, in this narrow sense, is an evaluation of the consequences of a candidate move under near-optimal continuation by both sides, given the engine’s own search and evaluation parameters. The evaluation function is a learned estimate of “if both sides play near-optimally from here, what is the expected result of the position?”. The engine’s estimate is, on the current evidence, more reliable than any unaided human’s estimate of the same question, with appropriate caveats for depth, hardware, and engine version. This is what is often colloquially summarised as “the engine knows the position.” The summary is convenient. It is also imprecise, and the imprecision matters. The engine does not “know” anything in the human sense. It produces, within its search budget and evaluation function, the strongest available practical estimate of an answer to a particular retrospective question.

What the engine does not do is choose the move. The engine does not register the player’s pulse rate, the eight hours already spent at the board, the meaning of this game inside a tournament standing, the player’s last result, the opponent’s reputation, the country in the player’s passport, the money on the line, the previous evening’s sleep, or the cost of a draw offer at this particular moment. The engine does not face fatigue. The engine does not face fear. The engine does not face the cognitive task of having to commit to a move within a finite remaining clock under conditions of incomplete confidence, knowing that the move is the public evidence of the player. These are not features the engine fails to handle well. They are features the engine does not have to handle at all.

This is not a romantic claim. It is a structural one. The engine is a position evaluator. The player is a decision-maker under conditions. These are different objects, and confusing them in either direction distorts the analysis of what chess after the engine actually is.

Section IIIWhat the engine evaluation does not settle

The engine’s evaluation is retrospective in the relevant sense. Even when the engine runs in real time during a broadcast, the evaluation it produces is an evaluation of the position as if it had to be played out by a near-optimal entity, not by the human at the board. It tells the viewer what the strongest available continuation looks like under near-optimal play. It does not tell the viewer what the player should do, given who the player is, what the player has trained, and what the player can trust under the time available.

This requires a distinction the popular discussion of “engine analysis” tends to flatten. There is the engine-best move; there is the human-playable move, namely the move a particular human can calculate, verify, and commit to under the time and conditions of the actual game; there is the practical tournament decision, which weighs match situation, opponent, fatigue, and risk; and there is the pedagogically useful explanation, which aims to teach a player why certain decisions should be made in certain types of positions. These four are not the same object, although the engine evaluation contributes to each. A coaching culture that collapses them into the first one alone has misunderstood what most of post-game review is for.

The first three essays in this series examined three things engine analysis cannot reach. Essay No. 01 argued that a single chess game produces three verdicts, namely a technical record, a public broadcast, and a private identity verdict, and that the technical verdict is precisely the one engine analysis has now made authoritative. Engine analysis has loud authority on the technical layer and no direct authority on the identity layer, and conflating the two is a common, quiet form of damage. Essay No. 02 argued that calculation and confidence are different cognitive achievements, and that the integrated decision the player commits to is not the same object as the calculation the engine produces; the engine can confirm what the calculation should have been, but it cannot create the integrated trust that turns calculation into a move the player is willing to play. Essay No. 03 argued that pressure is the architecture under which decisions are made, and that the chess clock is the instrument that makes the pressure measurable; the engine sits outside the clock, evaluates positions as if the clock did not exist, and therefore has nothing direct to say about the cognitive fact of having to commit a move within a finite remaining time.

These limits are not failures of engineering. They describe a different category of problem. The engine is built to estimate the strongest continuation under near-optimal play. The player is required to make decisions when near-optimal play is not available, and the difference between those two situations is the entire content of competitive chess as a human practice.

The engine can produce the strongest available evaluation of a position; what it does not do is choose under time, fear, ego, and consequence on behalf of the human at the board.

Section IVWhy retrospective evaluation can erode prospective judgment

There is a specific psychological hazard in the modern training environment that the older chess culture did not face at the same intensity. The hazard is that a strong retrospective benchmark, delivered authoritatively by engine, can erode the prospective judgment the player has to exercise at the board.

The cognitive-science literature on human-automation interaction has documented a relevant pattern across decades of work in aviation, medicine, and process control. Parasuraman and Manzey, in their 2010 review in Human Factors, integrated a body of empirical evidence on what the literature calls automation complacency and automation bias: under sustained interaction with reliable automated systems, human users learn to defer to them, sometimes in cases where their own judgment would have served better. The effect is not a failure of intelligence; it is a documented consequence of how trust is built in repeated interactions with reliable instruments. Automation bias is documented across domains ranging from aviation autopilot to medical diagnostic decision support, and it has been refined further in the trust-calibration literature initiated by Lee and See (2004) and the systematic reviews in medical-decision-support contexts (Goddard, Roudsari & Wyatt, 2012). This is not direct chess-specific evidence: there is no single high-quality study of engine-induced overreliance in chess populations specifically. What this literature does provide is a framework for thinking about a plausible training risk in chess, namely that years of engine-supervised study can train a player to accept engine output without reconstructing the human reasons that would make a move playable over the board.

The structural inference for chess is therefore the following adjacent-inference claim: the conditions under which automation bias and complacency are well-documented in adjacent literatures are present in modern chess preparation, and the question deserves serious coaching attention rather than dismissal as an anti-technology complaint. The chess-specific empirical evidence on engine-induced overreliance is still developing; the careful claim is bounded.

The post-game environment poses a related and sharper risk. After every serious game, the modern player faces an immediate retrospective tribunal in the form of an engine. Within seconds, every move is rated. A move played after twenty minutes of agonised calculation is reduced to a centipawn evaluation. The number can be informative. It can also be ungenerous. It does not record the alternative variations the player rejected, the time pressure that made the move necessary, or the psychological cost of the previous half-hour of calculation that produced it. A coaching culture that treats the engine number as the final verdict on the move risks teaching its students an inverted lesson, namely that the question to ask after a game is not “did I make the best decision available to me, given what I could see and trust at that moment?” but “did I find the engine move?”. These are not the same question, and the difference is much of the content of human chess.

The first essay in this series argued that the technical, broadcast, and identity verdicts of a game are different layers, and that the dangerous failure mode is the identity verdict’s absorbing the other two. Engine analysis speaks loudest at the technical layer; it has nothing direct to say at the identity layer; and a post-game review that loops the engine’s verdict back into the player’s identity verdict, without the structural distinction the second essay’s integration argument requires, is producing a player whose self-judgment is calibrated to a tool that does not have to play the next round.

Section VThe human move as a decision under conditions

The human move is not the engine’s recommended move. The human move is a decision under conditions the engine does not face.

A grandmaster choosing between two moves at move 35, with three minutes remaining on her clock, in the seventh round of a Swiss tournament, after a long preparation morning, against an opponent she has lost to twice, is not solving the same problem the engine is solving. The engine is asked which move maximizes evaluation under near-optimal play. The grandmaster is asked which move she can trust enough to play, calculate accurately enough to verify within the remaining time, execute without procedural disruption, and live with regardless of result. The two questions point in similar directions in the majority of positions. They do not always point in identical directions, and the divergent cases are not failures of human chess; they are its content.

The recognition-primed-decision framework articulated by Klein and colleagues (Klein, 1998), discussed in Essay No. 02, supplies the cognitive language for what the human move actually is. The expert under time pressure does not, as a rule, generate a comprehensive list of candidate moves and select the highest-evaluated one. The expert pattern-matches against accumulated chunks of position type, generates a small number of plausible candidate moves through Type 1 cognition, and verifies the leading candidate through Type 2 calculation under the available time. The output of this process is the integrated decision the player commits to, which is the same object Essay No. 02 identified as the playable move rather than the merely calculated one.

The engine clarifies, after the fact, what the optimal candidate move was. It does not change the cognitive process by which a human selects a candidate move under time, and it does not eliminate the requirement that the candidate move selected be one the player can integrate, verify, and play with the trust required to commit. The interesting cases for a player’s development are precisely the cases where the optimal move and the integrated move diverge, namely the positions where the objectively second-best move is the move the player can actually play with sufficient confidence under the conditions of the day. A coaching culture that treats those moments as failures has misunderstood what the player is being trained to do.

The interesting cases for a player’s development are not the moves the engine and the player both find. They are the moves the player can integrate under conditions, and the engine, which faces no conditions, has no reason to recognise as a problem.

Section VIHow to use the engine without outsourcing the self

The argument so far is not anti-engine. It is the opposite. The engine is one of the most powerful instruments ever developed for chess study, and a serious player who refuses to use it is choosing to be weaker than necessary. The argument concerns how to use it.

The cognitive-science literature on automation use, taken together with the expertise literature on deliberate practice, supports a small set of structural orientations for the engine-using player.

One, separate the engine’s verdict on the position from the player’s judgment about the decision. The engine answers, “what was the best move under near-optimal play?” The coaching question after a serious game is, “what did I see, what did I trust, and what could I have seen with the time and information I had?” The engine helps with the first; it does not replace the second. The post-game review that asks both questions, in that order, treats the engine as a research instrument. The post-game review that asks only the first treats the engine as a tribunal, and the literature on automation bias predicts the consequences of this confusion across years of training.

Two, treat engine review as evidence, not as moral verdict. The number on the move is information. It is not the standing of the player. The first essay in this series argued that the technical, broadcast, and identity verdicts of a game are different layers and that the dangerous error is letting the technical verdict absorb the identity verdict. The engine intensifies this risk, because the technical verdict it produces is louder and faster than any human annotation. The structural intervention is the same intervention the first essay identified, namely the boundary between the technical verdict and the identity verdict, maintained deliberately in the moments after the engine has spoken.

Three, build deliberate practice without the engine. The capacity to evaluate a position without machine support is itself a trainable skill, and it is the skill that the player will need at the board, where the engine is not available. The deliberate-practice literature (Ericsson, Krampe & Tesch-Römer, 1993; Ericsson et al., 2018) is unambiguous that practice configurations matter for the skills they develop. A training life that is always supervised by the engine produces a player who never learns to evaluate alone, because she has never been required to. The remedy is structural, namely deliberate practice configurations in which the engine is unavailable during the evaluation phase and consulted only after the player has committed to a written assessment of the position.

Four, calibrate the engine’s recommendations as the moves of an extremely strong player rather than the moves the player should have made. The gap between those two categories is the developmental space. The engine’s move is a near-optimal move under near-optimal continuation; the player’s move is the move a particular human can integrate under conditions; and the question for development is not “why did I not find the engine’s move?” but “what would I need to see, and what would I need to trust, to make that move available to me under conditions like these?” The first question produces self-criticism. The second produces training.

Five, train under transfer-relevant conditions. The deliberate-practice literature is again clear that the conditions of practice matter for the conditions of performance. Engine-supervised study with unlimited time produces a different cognitive product than time-pressured calculation under realistic clock conditions. Both are useful. The error is treating the first as if it were the second, and the player who relies primarily on the first will, on the published evidence on transfer, register the gap on game day rather than in study.

Section VIIThe recognition-primed decision in the engine era

The cognitive frame articulated in Essay No. 02, namely that the expert decision is the integrated product of recognition-driven Type 1 candidate generation and calculation-driven Type 2 verification, has a specific consequence in the engine era that is worth making explicit.

The engine has changed the content of the chunks the player is acquiring. Pattern recognition in chess is built across thousands of hours of position-pattern exposure, and the position patterns the contemporary player is exposed to are increasingly engine-mediated. The opening preparation the player studies has been refined by engine evaluation. The middle-game positions the player drills have been generated, evaluated, or filtered by engine. The endgame technique the player learns has, for tablebase-relevant positions, been replaced by retrieval rather than reasoning. The cognitive product of this exposure is a chunk library that is calibrated to engine evaluation.

This is, on balance, a strengthening of human chess. The chunk library a contemporary player can build by age twenty exceeds in evaluative accuracy the chunk library most pre-engine players could build in a career. The hazard is the one the automation-bias literature predicts, namely that a chunk library calibrated to a tool not present at the board can produce candidate moves that work under the tool’s conditions and fail under the player’s conditions. The integrated decision the second essay identified as the trained achievement is, in the engine era, harder to build, not easier, because the recognition layer can be technically sophisticated while the verification layer remains under-trained for time pressure.

The implication is that the deliberate practice the second essay’s integration argument required is more important now, not less. The engine has elevated the recognition layer, and the player who treats this elevation as a substitute for the integration work treats a stronger candidate-generation system as if it were a stronger decision-making system. They are not the same.

Section VIIIWhat the evidence supports for players, parents, and coaches

The combined evidence on automation bias, expert performance, and chess cognition supports five structural recommendations for the four audiences in descending order of leverage.

One, to players: use the engine as an instrument, not as a tribunal. The engine review is information about the position. It is not the standing of the player. The post-game question that the cognitive evidence supports is “what did I see, what could I trust, and what would I do again under the same conditions?”, and the engine cannot answer the second cluster of questions. Treating its evaluation as if it could is a recipe for a long career of self-criticism that misses its target.

Two, to coaches: design the post-game review to disentangle the technical verdict from the identity verdict. The first essay in this series argued that the boundary between these two verdicts is the structural intervention site, and the engine has made this intervention more important rather than less. A post-game session in which the engine evaluation is the first object discussed and the player’s prospective judgment is never explicitly examined is producing a student who learns to interpret her own decisions through a tool she does not have access to during play.

Three, to coaches and academies: build deliberate practice configurations without engine supervision. The capacity to evaluate without the engine is the capacity the player will need at the board. The deliberate-practice literature predicts that this capacity is built specifically by practice in which it is required, and not by practice in which it can be substituted with engine consultation. The training week that includes scheduled engine-free evaluation hours, written position assessments produced before any engine consultation, and explicit comparison of pre-engine assessments with post-engine evaluations is producing a different cognitive product than the training week that does not.

Four, to parents: protect the young player from the engine’s identity verdict. A child whose self-image is calibrated to engine approval has been quietly trained to believe that her own judgment is valuable only when it agrees with a machine. This is not a healthy relationship to the game, and it is not a healthy relationship to her own mind. The supporting frame’s job is to keep the engine in its category as research instrument, and to keep the child’s developing identity as a chess player out of the centipawn-loss column of the analysis screen.

Five, to academies and federations: the training cultures built today are the competitive cultures of the next generation. A culture that treats the engine as a teacher is producing students. A culture that treats the engine as a judge is producing defendants, and the difference between those two outputs will be visible in the kind of player the federation has in ten years.

Section XConclusion · after the engine, the human remains

The age of the engine has not made chess meaningless. It has clarified what chess actually is. Before the engine, chess could be defended as a search for the strongest move; that defence is now harder to make, because the strongest available evaluation of most positions is now retrieved from engines rather than produced by unaided human analysis. What remains, and what no engine update will erase, is the human decision under conditions, namely the integrated commitment of a particular trained mind to a particular move within a particular finite remaining time, with all the consequence that follows.

Chess after the engine is not a contest to prove that humans calculate better than machines. They do not. It is a different contest, namely a contest of trained, time-pressured, ego-implicated, identity-loaded human decision-making, the kind that the engine can review but does not perform. The engine evaluates. The player still has to move.

The argument generalises. Many domains of contemporary work are in the process of acquiring tools that produce, in seconds, retrospective evaluations whose authority exceeds the practitioner’s own judgment. Medicine has its diagnostic decision-support systems. Law has its case-search and document-analysis tools. Software engineering has its language-model assistants. Each of these domains is, in its own form, encountering the question chess has been working on since 1997, namely how to use a tool whose retrospective evaluation exceeds the practitioner’s prospective judgment without outsourcing that prospective judgment to a tool that is not present in the room when the decision is made. The chess case is, structurally, an early instance of a question that is becoming general.

The next essay in this series turns to the related question that the engine condition has made unavoidable. If engines now produce moves stronger than any unaided human can, then chess depends on a particular form of trust, namely the trust that the move at the board came from the person sitting there. That trust is the integrity layer of the modern game, and it is the subject of Essay No. 05.

The engine produces the strongest available evaluation. The player makes the move. After the engine, the move is what the human game is about.