The fragility of reputation effects

Abstract: I revisit the canonical reputation framework in which a long-lived player interacts with a sequence of short-lived opponents and may be either strategic or a commitment type who always plays the same, possibly mixed, action. I depart by allowing short-lived players to be uncertain not only about the long-lived player's type, but also about the signal structure. I show that even vanishingly small misspecified skepticism of short-lived players about commitment as an explanation of the observed signals can completely eliminate reputation effects: a patient strategic long-lived player's equilibrium payoff is bounded above by the canonical complete-information benchmark.

• The paper shows that even vanishingly small misspecifications in subjective models can completely eliminate reputation effects.
• It extends the canonical reputation framework by incorporating uncertainties over signal structures using dynamic relative entropy and equilibrium bounds.
• The findings imply that dynamic reputation mechanisms are fragile, urging careful empirical evaluation in markets, regulatory settings, and economic policy design.

The Fragility of Reputation Effects in Repeated Games

Introduction and Motivation

Reputation effects—where long-lived agents benefit from being perceived as “committed” to certain behaviors by a sequence of short-lived, observant opponents—form a core prediction in dynamic game theory. The canonical reputation framework, formalized by [Fudenberg & Levine 1989, 1992] and rooted in the original insights of [Kreps, Milgrom, Roberts, Wilson 1982], demonstrates that even an arbitrarily small probability of commitment can dramatically elevate the patient long-lived player's equilibrium payoff, compared to the complete information benchmark. The credibility of “commitment” as a working hypothesis for short-lived players is central to this logic. However, empirical observations suggest that agents often misattribute observed outcomes, using subjective or misspecified models that introduce narrative alternatives to commitment—such as luck, shocks, or unobserved heterogeneity.

This paper rigorously investigates the robustness of reputation effects to even vanishingly small forms of skeptic or misspecified model uncertainty in the interpretation of observable signals by short-lived opponents. It establishes that the celebrated reputation effect is highly fragile: if short-lived players assign any probability—even arbitrarily tiny—to models in which the commitment type is not the best explanation for observed signals, reputation payoffs collapse to the complete information Nash equilibrium benchmark.

Canonical Reputation Models and Their Fragility

The canonical framework involves a long-lived player (type: strategic or commitment) interacting with a sequence of short-lived players. With correct specification, imperfectly informative but public monitoring, and positive probability on commitment, the strategic type can achieve payoffs arbitrarily close to commitment as patience grows. The key is that short-lived players interpret observed signals using the true signal structure and are Bayesian updaters regarding type.

Building on recent empirical and theoretical literature on model misspecification and learning [Frick et al. 2020; Bohren & Hauser 2025], this paper departs from canonical assumptions by introducing the possibility that short-lived players are also uncertain or incorrect about the signal structure—not just the long-lived player’s type. Specifically, subjective models now allow for narratives or observationally plausible signal processes under the normal type that better fit the data than any commitment story.

Main Theoretical Results

Generalization and Extension

The paper extends the canonical model by allowing short-lived players to have beliefs over both types and a (potentially misspecified) family of signal structures, and crucially, they may assign zero probability to the true monitoring structure under which commitment could generate the observed signals.

Proposition on Misspecification and Entropy-Minimization

A formal sufficient condition is presented: if the set of subjective models includes any signal structure under which some “commitment-like” type can account for the observed signal distribution (in terms of minimal Kullback-Leibler divergence from truth), reputation effects persist, possibly in a distorted or misspecified form. This proposition clarifies that the payoff benefit of reputation is not tied to literal correctness, but to the existence of a close-enough commitment narrative in subjective beliefs.

Main Theorem: Complete Elimination of Reputation Effects Under Skeptical Misspecification

The core theorem of the paper states that for any vanishingly small sequence of model misspecifications (in which, along the sequence, some skeptical model remains with positive probability and explain signals better under the normal type than under any commitment explanation), the superior equilibrium payoff for the patient long-lived player is bounded above by the canonical complete information Nash equilibrium value:

$$\limsup_{n \to \infty} \limsup_{\delta \uparrow 1} \overline W(\delta; M_n, \pi_n) \leq \overline W^{CI}$$

even as the misspecification vanishes ($n \to \infty$). That is, the entire reputation effect disappears for all equilibria in this patient limit. The limiting order—first patience, then vanishing misspecification—is essential, mirroring the canonical logic of reputation payoffs.

The result is strong: it only requires the existence of models within the subjective prior of short-lived players, no matter how small their weight, as long as these models systematically explain the signal process better under the normal type.

Structural Tightness and Non-Genericity of Collapse

This fragility is not vacuously generic. If the subjective model space contains plausible commitment explanations for feasible signal distributions (i.e., if the observed data are close in relative entropy to some commitment type under some subjective model), reputation effects survive, as shown constructively. The collapse relies on uniform skepticism: the absence of any model where the commitment type can explain the data as parsimoniously as the normal type.

Proof Techniques

The analysis leverages a blend of dynamic relative entropy arguments, exponential concentration bounds on posterior decay, and a fine-grained extension of the folk theorem logic for repeated games with approximate best-responses. The latter is essential due to the two-stage limiting process (misspecification and patience) and endogeneity of individual learning, differing from the classical [Berk 1966] result for model misspecification. Stepwise, the strategy involves:

1. Demonstrating that—conditional on the normal type—the posterior on commitment decays exponentially along skeptical models.
2. Showing that short-lived players’ average payoff loss from failure to best-respond vanishes in the double limit.
3. Applying a robust extension of equilibrium payoff bounds for repeated games with approximate best-responses to sharpen the upper limit above.

Implications and Theoretical Significance

This work injects a significant degree of caution into the application of dynamic reputation arguments, especially for practical contexts such as industrial organization, labor markets, monetary policy credibility, competitive deterrence, and markets for certification or ratings. It demonstrates that the apparent “amplification” of small probabilities of commitment into large strategic payoffs is not merely sensitive but critically dependent on supporting narratives being both available and sufficiently credible in the worldviews of observers.

Theoretically, this non-robustness aligns reputation effects with other celebrated equilibrium selection mechanisms that require not only correct priors on types but also on data-generating processes. It suggests that attempts to structurally estimate dynamic contracts or strategic reputation in empirical contexts must be extremely attentive to how agents parse signals and what alternative explanations their learning models admit.

Moreover, the analysis indicates strong negative comparative statics: even vanishingly small misspecification can restore Nash equilibrium, hence suppressing reputation-based incentives in settings ranging from repeated price-setting and performance evaluation to entry deterrence and regulatory compliance.

Future Directions

Going forward, integrating richer learning dynamics with endogenous misspecification or allowing for active information design by the long-lived player could further clarify the boundaries of robust reputation formation. There is scope for developing frameworks in which feedback about the signal structure itself may be inferred or learned over time, partially restoring reputation under meta-learning or in the presence of at least some likelihood reweighting mechanisms.

Additionally, applying the insights of this fragility to AI systems, autonomous agents, or economic platforms where agents interact with heterogeneous and potentially algorithm- or human-generated signal interpretations may meaningfully inform the design and regulation of dynamic reputation systems.

Conclusion

The paper rigorously establishes that reputation effects in repeated games are non-robust to even arbitrarily slight forms of misspecified skepticism on the part of short-lived players. Unless all plausible subjective models for interpreting signals permit commitment-based narratives to explain the data as well as normal-type strategies, the long-lived player’s reputation payoff elevation vanishes entirely in the patient limit. This necessitates a reinterpretation of dynamic reputation arguments: their empirical force and theoretical applicability hinge on stringent—often unacknowledged—assumptions about the sophistication and openness of observer learning models.