AI Beats Humans at Chess But Changes How Winners Play

The Machine Changed the Game. Then It Changed the Players.

In the early 1990s, the best chess players in the world trained a specific way. They memorized openings, studied their opponents' past games, and relied on a deep, almost intuitive feel for the position. Being a great chess player meant having a great chess brain.

Then the engines came. First Deep Blue, then Stockfish, then AlphaZero. By the mid-2010s, any grandmaster could run a $50 program on a laptop that played at a level no human could touch. The old question was settled: humans were obsolete at chess.

But something strange happened. The humans didn't disappear. They adapted. And in adapting, they became something new.

A team of researchers led by Sebastian Krakowski, Johannes Luger, and Sebastian Raisch at the University of Geneva decided to study this transformation. They had a radical idea: maybe the same AI that made human chess expertise irrelevant was also creating a new kind of expertise. And that new expertise had almost nothing to do with the old one.

Their findings, published in the Strategic Management Journal in 2022 (Krakowski et al., 2022), are not just about chess. They are about what happens to human skill when a machine does the thing you were trained to do better than you ever could.

The Three Worlds of Chess: Human, Engine, and Centaur

The researchers used a clever natural experiment. They analyzed tournaments from three different eras of chess: pure human play (no computers allowed), pure engine play (two AIs facing off), and "centaur" chess where a human and an AI work together as a team.

The centaur tournaments are the key. In these events, a human player sits at a computer running a chess engine. The human can query the engine, ask for evaluations, and see its suggested moves. But the human makes the final decision. The team wins or loses together.

Krakowski and his colleagues gathered data on 355 players who competed in all three types of tournaments. They measured each player's performance in human-only events, then tracked how the same players performed when they switched to centaur mode. They also looked at the pure engine games to understand what the AI was doing on its own.

The question was simple: does being a great human chess player make you a great centaur chess player?

The answer was no. And that no is the whole story.

The Skill That Used to Matter Stopped Mattering

Here is what the data showed. A player's performance in human chess tournaments had almost no relationship to their performance in centaur tournaments. The correlation was close to zero (Krakowski et al., 2022).

Think about what that means. If you were the world's best human chess player, with a rating of 2800 and a lifetime of opening theory in your head, you were not automatically better at human-plus-machine chess than a club player who knew how to use an engine well. The old skill set had become nearly irrelevant.

The authors call this "substitution dynamics." The AI does not just make humans better at the old game. It makes the old game itself obsolete. The capabilities that defined excellence in one context become worthless in another.

This is not the comforting story we usually tell about AI. We like to say that AI augments human abilities, that it makes us smarter, that it is a tool like a calculator. But a calculator does not change what it means to be good at math. It just makes calculation faster. Chess engines do something more radical. They change the definition of good.

The New Skill: Something Stranger

If old chess skill did not predict centaur performance, what did? The researchers found that the best centaur players had developed a new capability: the ability to manage the human machine relationship.

This is not a technical skill. It is not about knowing more chess than the engine. It is about knowing when to trust the engine, when to override it, and how to ask the right questions.

The authors describe it as a "novel human machine capability" that is "unrelated, or even negatively related, to traditional capabilities" (Krakowski et al., 2022). In other words, the things that made you a great human chess player might actually make you a worse centaur player.

Consider why. A grandmaster has spent decades building intuition. They feel the board. They sense threats. When an engine suggests a move that violates their intuition, they are tempted to reject it. But the engine is often right. The grandmaster's hard won intuition becomes a liability.

The best centaur players, the researchers found, are the ones who can suppress their own expertise. They treat the engine as a partner, not a tool. They ask it for evaluations not just of the current position but of the opponent's likely responses. They learn the engine's quirks and blind spots. They develop what amounts to a theory of mind for a machine.

This is a genuinely new kind of intelligence. It is not human intelligence augmented by machine. It is something that emerges at the intersection of the two, and it follows its own rules.

How the Study Was Done: The Method That Makes This Credible

The researchers did not ask players to fill out surveys about how they felt about AI. They analyzed actual tournament performance data from three distinct settings.

The human only data came from standard over the board tournaments governed by FIDE, the international chess federation. These are the classic tournaments we all know, with clocks and boards and silent rooms.

The engine only data came from computer chess championships where programs compete against each other without human intervention. These are pure AI versus AI contests.

The centaur data came from "freestyle" chess tournaments, also known as "advanced chess" events. In these tournaments, humans and engines work together. The rules vary, but the key point is that the human makes the final move decision after consulting the engine.

By tracking the same 355 players across all three settings, the researchers controlled for individual ability. They could see how the same person's performance changed depending on whether they were playing alone, with a machine, or against pure machines.

The sample size is large enough to draw meaningful conclusions. The setting is controlled enough to isolate the effect of AI adoption. And the results are stark enough to be surprising.

The Negative Relationship: When Old Skill Hurts

One of the most striking findings in the paper is the negative relationship between traditional chess skill and centaur performance. The authors found that "these novel human machine capabilities are unrelated, or even negatively related, to traditional capabilities" (Krakowski et al., 2022).

This is counterintuitive. We expect expertise to transfer. A great violinist who picks up a guitar will not be a beginner. But in the centaur setting, the old expertise appears to be an obstacle.

Why? Because the old expertise is about making decisions independently. A grandmaster's skill is in calculation, pattern recognition, and strategic planning. These are all things the engine does better. When the grandmaster insists on using these skills, they are competing with the engine rather than collaborating with it.

The best centaur players, the researchers suggest, are those who can let go. They do not try to out think the engine. They try to understand what the engine sees and then make decisions that the engine cannot make on its own.

This is a hard lesson for experts. It is also a lesson that applies far beyond chess.

What This Means for Managers, Doctors, and Everyone Else

The paper is published in a management journal, and the authors are explicit about its broader implications. They write that "AI based technologies increasingly substitute and complement humans in managerial tasks such as decision making" (Krakowski et al., 2022).

The chess setting is a model for what happens when AI enters any domain where expertise matters. The pattern is the same. First, the AI outperforms humans at the core task. Then, humans try to use the AI as a tool. Then, the best humans realize that the tool changes the task itself.

Consider medical diagnosis. A radiologist who has spent twenty years reading X rays has developed a deep intuition for spotting tumors. Now an AI can do it faster and more accurately. The radiologist's old skill is not augmented. It is made obsolete.

But a new skill emerges: knowing when to trust the AI, when to question it, and how to combine its output with other information the AI cannot access. The radiologist who is best at the old task may not be best at the new one. The radiologist who is best at the new one may have never been great at the old one.

The same logic applies to financial analysis, legal research, and even software engineering. As Krakowski and his colleagues show, the source of competitive advantage shifts. It moves from individual expertise to something more relational, more collaborative, and harder to measure.

What the Research Does Not Prove: The Open Questions

The study is careful and rigorous, but it leaves important questions unanswered.

First, the researchers measured performance in tournaments, not learning. They did not track how players developed their centaur skills over time. It is possible that traditional chess skill becomes an advantage after a period of adaptation. The negative relationship might be temporary.

Second, the centaur setting in chess is relatively simple. The human and machine work on the same problem at the same time. In many real world settings, the human and machine work on different parts of the problem, or the machine makes recommendations that the human reviews later. The dynamics may be different.

Third, the study does not tell us how to train the new capability. If old chess skill does not predict centaur performance, how do you identify people who will be good at it? And how do you teach it? The paper identifies the phenomenon but does not yet provide a playbook.

These are not weaknesses. They are the next questions.

What This Actually Means

The paper by Krakowski, Luger, and Raisch (2022) changes how we think about AI and human expertise. Here is what it means in practice:

• ▸If you are an expert in a field where AI is advancing, your old skill is not a safety net. The better you are at the old way, the harder it may be to adapt. Do not assume your expertise transfers. Assume it needs to be rebuilt.

• ▸The new skill is not about knowing more than the AI. It is about knowing how to work with the AI. This is a relational skill, not a technical one. It involves trust, judgment, and the ability to question your own intuition.

• ▸Organizations should stop hiring for traditional expertise alone. If you are hiring a radiologist, a trader, or a chess player to work with an AI, look for candidates who can collaborate with machines, not just candidates with the best solo track record.

• ▸Training programs need to change. Teaching people to be better at the old task is a waste of time if the old task is being automated. Instead, teach people how to evaluate AI outputs, how to calibrate trust, and how to make decisions that combine human and machine strengths.

• ▸The goal is not human plus machine. It is human times machine. The best results come from a synthesis that neither side can achieve alone. That synthesis has its own logic, and it rewards a different set of human abilities than the ones we have been cultivating for centuries.

The machines beat us at chess. That much we knew. What we did not know is that they would also change what it means to be a winner. The players who adapted did not just use the machine. They became something new. And so will we.