Selection of the fittest or selection of the luckiest: the emergence of Goodhart’s law in evolution

Zsófia Talyigás, Francesco Paparella, Bastien Mallein, Emmanuel Schertzer · 2025

Read Selection of the fittest or selection of the luckiest: the emergence of Goodhart’s law in evolution by Zsófia Talyigás, Francesco Paparella, Bastien Mallein, Emmanuel Schertzer (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-21849.txt.

Introduction

Zsofia Talyigas, Francesco Paparella, Bastien Mallein, and Emmanuel Schertzer ask whether evolution selects the fittest or merely the luckiest in a 2025 mathematical biology study. Selection judges phenotypes, the traits an organism shows, and chance events blur the link between a good genotype and reproductive success. The authors model that blur and its long-run consequences.

Their analysis shows a Goodhart-like effect emerging inside evolution itself: once a trait becomes the target of selection, luck in expressing it can swamp the underlying quality. The result reframes how biologists should read adaptation, since winners may owe their place partly to fortune.

Catalog record

CREATORZsófia Talyigás, Francesco Paparella, Bastien Mallein, Emmanuel Schertzer
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE125 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-21849

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Selection of the fittest or selection of the luckiest: the emergence of Goodhart’s law in evolution Zsófia Talyigás, Francesco Paparella, Bastien Mallein, Emmanuel Schertzer Abstract Natural selection is commonly assumed to become more effective as it becomes stronger. However, selection acts on phenotypes rather than directly on genotypes, and phenotypic success is inherently noisy. Here we study how this mismatch shapes long-term evolutionary dynamics. Using a minimal stochastic model in which individuals inherit genetic fitness while selection acts on noisy phenotypic expressions, we show that increasing selection strength accelerates adaptation only up to a critical threshold. Beyond this point, stronger selection paradoxically slows evolution and erodes genetic diversity by favoring the luckiest individuals rather than the genetically fittest. We identify two distinct evolutionary regimes—selection of the fittest and selection of the luckiest—separated by a sharp transition. This transition corresponds to a previously unrecognized change in the structure of traveling fitness waves, from semipulled to fully pulled fronts, with profound consequences for adaptation speed and genealogical structure. Our results reveal a biological instance of Goodhart’s law: when phenotypic measures become overly optimized targets, they cease to reliably promote genetic improvement. These findings highlight intrinsic limits to the effectiveness of strong selection and suggest that optimal evolutionary outcomes require intermediate selection strength in noisy environments. 1 Introduction Natural selection is often assumed to become more effective as it becomes stronger. Yet selection acts on phenotypes rather than directly on genotypes, and phenotypic success is inherently shaped by chance. As a result, the individuals favored by selection within a given generation are not always those carrying the most heritable advantages. This distinction, already emphasized by Fisher in his dictum that “natural selection is not evolution” [ 24 ] , highlights a fundamental separation between short-term selective success and long-term evolutionary change. Selection filters phenotypes within generations, whereas evolution depends on the reliable transmission of genetic information across generations. This raises a fundamental question for evolutionary theory: how does the strength of selection shape long-term evolutionary outcomes, not only in terms of the rate of adaptation, but also in terms of the maintenance of genetic diversity? When optimization relies on noisy or imperfect proxies, intensifying selection can produce unintended and even counterproductive outcomes. In economics and the social sciences, this failure mode is known as Goodhart’s law [ 27 , 18 ] : when a measure becomes a target, it ceases to be a good measure. Closely related phenomena have since been identified across a wide range of disciplines, including statistics, control theory, computer science, and machine learning where optimizing imperfect objective functions can lead to overfitting and instability [ 35 , 39 , 49 , 1 , 53 , 44 , 45 ] . In all these contexts, stronger optimization does not necessarily improve performance, but can instead amplify noise and distort the relationship between measured targets and underlying goals. Whether an analogous phenomenon can arise in biological evolution—where selection measures phenotypic prowess as proxy for underlying genetic fitness—has remained largely unexplored. Yet biological selection is unavoidably noisy, shaped by environmental contingencies, developmental variability, and stochastic interactions between organisms and their environment. These sources of noise suggest that overly strong selection might, paradoxically, undermine the effectiveness of the evolutionary mechanism. To investigate this possibility, we study an idealized evolutionary model in which individuals inherit a quantitative genetic fitness, while selection operates on noisy phenotypic expressions of that fitness. Only a fraction of individuals reproduce each generation, with reproductive success determined by phenotypes that imperfectly reflect the genotypes transmitted to offspring. This framework captures a minimal but generic feature of biological evolution: selection acts on observable traits, whereas evolution proceeds through the transmission of genetic information. By analyzing the long-term dynamics of this system, we examine how reproduction skewness and phenotypic noise jointly shape evolutionary trajectories. First, we show that strengthening selection by allowing fewer individuals to reproduce—thereby magnifying small phenotypic differences—accelerates adaptation only up to a critical threshold. Beyond this point, stronger selection systematically favors the luckiest individuals rather than the genetically fittest, leading to a slowdown of evolutionary adaptation. This change in behavior reflects a qualitative shift in evolutionary dynamics between two distinct regimes, which we term the selection of the fittest and the selection of the luckiest. In the selection of the fittest regime, moderate restrictions on reproduction allow selection to reliably preserve the highest-fitness genotypes, resulting in efficient and sustained adaptation. By contrast, in the selection of the luckiest regime , overly strong restrictions amplify stochastic fluctuations in phenotypic success, causing reproductive success to be dominated by chance and leading to the systematic loss of high-fitness genotypes. We characterize these regimes and the transition between them through a combination of simulations and mathematical analysis. At a mathematical level, this transition is accompanied by a change in the nature of the underlying evolutionary waves [ 33 ] . We show that the dynamics shift from semipulled waves in the selection of the fittest regime to fully pulled waves in the selection of the luckiest regime, highlighting a previously unrecognized distinction within the class of pulled waves [ 42 , 61 ] . Beyond its mathematical interest, this distinction has direct biological implications. Our newly identified fully pulled regime represents an extreme form of tip-dominated evolution, in which the contribution of the bulk is negligible. As a consequence, genetic drift at the wave front is expected to be even stronger than in previously studied semipulled waves, leading to an especially pronounced loss of genetic diversity [ 31 , 32 ] . The emergence of Goodhart’s law in evolution reflects a non-linear dependence of the rate of adaptation on selection strength. We uncover a second instance of such non-monotonic evolutionary behavior by examining how genetic diversity responds to phenotypic noise. Reducing phenotypic noise can be viewed as an amplification of the evolutionary signal. We find that genetic diversity depends non-monotonically on noise amplitude: in the selection of the luckiest regime, amplification of the evolutionary signal by reducing phenotypic noise decreases genetic diversity, whereas, once the selection of the fittest regime is reached, further noise reduction has the opposite effect. Taken together, these results highlight non-monotonic responses of evolutionary outcomes to selection parameters. Continuous changes in how selection operates—whether by restricting reproduction or by modifying the fidelity of phenotypic information—can lead to qualitatively different evolutionary regimes. Our results suggest that biological evolution is governed by intrinsic non-linearities that limit the effectiveness of strong selection and give rise to complex evolutionary behavior. 2 The model We consider an asexual population of N\gg 1 individuals, of which K are selected to reproduce and generate N offspring. We will assume that each individual possesses a genotype , identified as a real number, representing the genotypic fitness of the individual. The genotype consists of the inherited information transmitted from parent to child. The expression of that genotype in a given environment is the phenotype , which is also represented by a single real number in our idealized model. In Biology, the genotype to phenotype map may be influenced by extrinsic and intrinsic noises such as developmental noise [ 26 ] , phenotypic heterogeneity [ 5 ] , cellular noise [ 37 ] , biological noise [ 22 ]

Read the complete text