🧬 biology

Cultural transmission of move choice in chess

This study analyzes decades of chess game records to demonstrate how social learning biases, including frequency-dependent, success, and prestige biases, drive the cultural transmission and evolution of move choices among players.

Original authors: Egor Lappo, Noah A. Rosenberg, Marcus W. Feldman

Published 2026-05-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Egor Lappo, Noah A. Rosenberg, Marcus W. Feldman

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine the world of high-level chess not just as a game of strategy, but as a massive, global conversation that has been happening for 50 years. Every time a grandmaster makes a move, they are essentially "speaking" a word from a shared vocabulary. This paper, written by researchers at Stanford, treats chess moves like words in a language to understand how ideas spread, change, and evolve within a culture.

Here is the story of their research, broken down into simple concepts and analogies.

The Big Idea: Chess as a Cultural Laboratory

The authors wanted to see how humans learn from each other. They chose chess because it's perfect for this:

It's a clear record: Unlike human conversations where words might be forgotten, every chess move is written down.
It's a huge dataset: They analyzed 3.45 million games played by top players between 1971 and 2019.
It's a social game: Players don't just play in a vacuum; they watch what others do, read books, and copy their heroes.

The researchers asked: When a player decides what move to make, are they copying the crowd, copying the winners, or copying the famous stars?

The Three Ways People Copy (Transmission Biases)

The paper identifies three main "social learning strategies" that drive players to choose one move over another:

Frequency-Dependent Bias (The "Anti-Fashion" Trend):
- The Analogy: Imagine a fashion trend where everyone starts wearing blue hats. Usually, people follow the crowd. But in chess, sometimes players do the opposite. If a move is played by everyone, a player might think, "If everyone is doing this, my opponent has probably prepared for it. I should do something weird instead."
- The Finding: The study found that for some early-game moves, players actively avoid the most popular choices. This is called negative frequency-dependent bias (or anti-conformity). They want to be the "underdog" strategy to catch their opponent off guard.
Success Bias (The "Copy the Winner" Strategy):
- The Analogy: Imagine you are trying to fix a leaky faucet. You see a video of a plumber who fixed it perfectly. You copy their exact steps because they worked.
- The Finding: In one specific opening (the Caro-Kann), players stopped using a certain move because it kept leading to losses. They saw the "failure" and stopped copying it. Here, the "win rate" of a move dictated its popularity.
Prestige Bias (The "Copy the Superstar" Strategy):
- The Analogy: Think of a celebrity wearing a specific brand of sneakers. Even if the shoes aren't necessarily "better" than others, everyone wants to wear them because the celebrity is famous.
- The Finding: In another opening (the Najdorf Sicilian), a specific move suddenly became popular. The data showed this wasn't because it was the most common or the most successful overall, but because the top 50 players in the world started using it. Once the "superstars" adopted it, the rest of the chess world followed.

The "Computer Engine" Effect

The paper notes a major shift in the last few decades: the rise of powerful computer chess engines.

Before: Players relied on human intuition and memorized books.
After: Computers can calculate the "perfect" move in seconds.
The Result: This changed the culture. Players now memorize long, computer-generated sequences. If a computer says a weird move (like pushing a pawn to the edge of the board, known as h3) is good, the top players adopt it, and the "Prestige Bias" spreads it to everyone else.

The "Sample Size" Mystery

One of the most interesting findings is about how much information players actually look at.

The Analogy: Imagine you are trying to guess what the weather will be.
- For the first move of the game (the "opening"), there are thousands of games to look at. It's like trying to predict the weather by looking at a million different cities; it's too much data, so players might only look at a tiny, unrepresentative slice.
- For later moves in the game (after 10 or 11 moves), the number of possible scenarios drops drastically. It's like looking at just your own neighborhood.
The Finding: The researchers found that players seem to pay attention to a larger percentage of the total games when the game gets deeper and more complex. They are "sampling" more of the culture when the situation is less standardized.

What Did They Actually Do?

The researchers built a complex mathematical model (using something called a Dirichlet-multinomial model) to simulate how move frequencies change from year to year.

They compared the real data against a "random" model (where players just guess).
They found that the real world didn't look random. The "social biases" (copying stars, avoiding crowds, or following winners) were the invisible hands shaping the game.

The Bottom Line

This paper proves that chess isn't just a game of logic; it's a game of culture.

Players are constantly learning from each other.
They sometimes copy the winners, sometimes copy the famous, and sometimes deliberately do the opposite of the crowd to stay unpredictable.
The introduction of computers has acted like a massive "cultural accelerator," helping new ideas spread faster than ever before.

By treating chess moves like cultural traits, the authors showed us that even in a game of perfect information, human behavior is driven by social trends, just like fashion, language, or memes.

Technical Summary: Cultural Transmission of Move Choice in Chess

Problem Statement
The study of cultural evolution often lacks detailed analysis of transmission mechanisms within specific human domains. While large datasets exist for linguistic or textual data, they often lack the precise behavioral records found in competitive games. Chess offers a unique platform for this analysis due to its discrete moves, complete information, and extensive historical records of high-quality data from elite players. The authors aim to quantify how social learning biases—specifically frequency-dependent bias, success bias, and prestige bias—affect the dynamics of move choice over decades. Unlike previous studies on games like Go that focused on individual-level binary choices, this paper addresses the challenge of modeling population-level move frequencies across a vast space of potential game sequences.

Methodology
The authors utilize a dataset of 3.45 million chess games played by masters (Elo $\ge$ 2000) between 1971 and 2019, sourced from the Caissabase compendium. The analysis focuses on specific board positions at varying depths in the game tree (e.g., the Queen's Pawn opening, Caro-Kann, and Najdorf Sicilian).

The core of the methodology is a population-level statistical model based on the Dirichlet-multinomial likelihood. The model operates as follows:

Null Model: It assumes a neutral transmission process where move counts in year $t+1$ are sampled from the counts in year $t$ , analogous to a Wright–Fisher model in population genetics.
Incorporating Biases: The model introduces fitness coefficients ( $f_i$ $f_{i}$ ) to represent cultural fitness. These coefficients are allowed to vary based on:
- Frequency-dependence: Modeled using piecewise-constant functions of the strategy's frequency, allowing for the detection of negative or positive frequency-dependent selection (anti-conformity or conformity).
- Success Bias: Represented by coefficients ( $\beta_{win}$ ) linking move choice to the average win rate of the strategy in the general population.
- Prestige Bias: Represented by coefficients ( $\beta_{top50}$ ) linking move choice to the win rates and frequencies of the top 50 players in the dataset for a given year.
Inference: The model is fitted using Bayesian inference via Hamiltonian Markov Chain Monte Carlo (HMC) sampling (Stan software). Non-informative priors are used, and the analysis focuses on the posterior distributions of the fitness functions and bias coefficients. To handle zero counts and rare moves, a correction term is added to the Dirichlet parameters, and rare moves are aggregated into an "other" category.

Key Results
The model was applied to three distinct positions, revealing different transmission dynamics:

Negative Frequency-Dependent Bias (Anti-Conformity): In the Queen's Pawn opening (ply 2) and the Caro-Kann opening (ply 5), the authors observed that certain moves (e.g., $d5$ and $Nf6$ in the Queen's Pawn; $exd5$ in the Caro-Kann) exhibited negative frequency dependence. These moves were selected with higher probability when they were rare and lower probability when common. This suggests players actively avoid popular lines, possibly to take opponents out of their preparation.
Success Bias: In the Caro-Kann opening, the decline of the move $exd5$ was significantly correlated with its lower win rate in the general population. The model indicates that players are avoiding this move because it leads to unfavorable outcomes, consistent with computer engine analysis showing $e5$ as superior.
Prestige Bias: In the Najdorf Sicilian (ply 11), the rapid rise of the move $h3$ was not driven by frequency dependence but was strongly correlated with the win rates of the top 50 players ( $\beta_{top50-win}$ ). This suggests that elite players, likely influenced by computer engine analysis, adopted the move, which was then imitated by the broader community.
Game Sample Size ( $N_s$ ): The authors introduce a concept of "game sample size" ( $N_s = \bar{f}_t N_t$ ), representing the effective number of games players observe when making decisions. They found that $N_s$ is significantly lower than the total number of games for early positions but increases for later positions, suggesting players sample a larger fraction of available data as positions become more complex and less standardized.

Significance and Claims
The paper claims to provide a quantitative framework for understanding cultural transmission in a domain with discrete, high-stakes choices. The primary contributions are:

Demonstration of Biases: The study empirically demonstrates that different transmission biases operate simultaneously in chess. While anti-conformity (negative frequency dependence) is prevalent in early openings, prestige and success biases play significant roles in specific contexts, particularly where computer analysis has shifted theoretical consensus.
Methodological Advancement: By employing a Dirichlet-multinomial model, the authors successfully model overdispersion in count data and separate the effects of selection (fitness) from random drift. This approach allows for the estimation of growth rates and the quantification of uncertainty in cultural transmission dynamics.
Connection to Chess Theory: The quantitative results align with qualitative knowledge held by professional chess players and historians. For instance, the model's identification of prestige bias in the Najdorf Sicilian corresponds to the known influence of computer engines and elite grandmasters in introducing new "side pawn pushes."
Broader Applicability: The authors suggest that their statistical approach can be extended to other cultural domains involving circumscribed discrete data, offering a tool to test narratives about cultural trends with statistical evidence.

The paper concludes that chess serves as a robust model for studying how social learning biases shape the evolution of cultural traits, specifically highlighting the interplay between individual innovation, elite influence, and the collective avoidance of over-prepared strategies.