English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Nothing Deceives Like Success: How Success Bias Creates an Illusion of Understanding in Science

Forum topic · 二一 · 2026-05-01

Summary

A 2026 arXiv study by Avery W. Louis (Stanford) and Marina Dubova (Santa Fe Institute) uses 2,301 agent-based simulations to show how social learning strategies systematically distort science's ability to judge its own theories. In the model, agent-scientists collect data, build theories, and selectively copy peers' theories. The study distinguishes perceived success (fit to sampled data) from actual success (fit to ground truth). Key findings: agents systematically overestimate their theories' quality, worsened by problem complexity (ρ=0.817); success bias—learning from seemingly most successful peers—collapses theoretical diversity and amplifies the illusion; success bias only lifts the worst theories while harming the best; agents who rationally optimize perceived success actually degrade real performance, reproducing Matthew-effect-like citation inequality; and evaluation accuracy only matters when success bias is strong, creating a self-reinforcing vicious cycle. The authors link these dynamics to historical cases such as N-rays and the amyloid plaque hypothesis, arguing that scientific progress may depend less on picking the right theory than on maintaining diverse explanations.

Nothing Deceives Like Success: Social Learning and the Illusion of Understanding in Science

> Source: "Nothing Deceives Like Success: Social Learning and the Illusion of Understanding in Science" > Authors: Avery W. Louis (Department of Symbolic Systems, Stanford University), Marina Dubova (Santa Fe Institute) > Preprint: arXiv:2604.27188 [physics.soc-ph], April 2026 > Method: 2,301 agent-based simulations of collective theory building

The Core Problem: Science Is Blind Men Touching an Elephant

Science is cumulative and social—but scientists can never directly know how good their theories are. They evaluate theories only through proxy indicators: predictive accuracy, fit to existing data, parsimony, and internal consistency. History offers cautionary examples, from the early-20th-century "N-rays" reported by French physicists to the Alzheimer's amyloid plaque hypothesis, which dominated research for decades while clinical trials repeatedly failed.

Louis and Dubova call this the illusion of understanding: a systematic gap between a community's real-time judgment of theory quality and those theories' long-term true performance.

The Simulation Design

Each agent in the model behaves like a miniature scientist:

  • Data collection: actively sampling observations from the environment
  • Theory building: compressing data into representations (theories)
  • Social learning: exchanging data and selectively adopting peers' theories
  • The model distinguishes:

  • Perceived success: how well a theory fits the agent's own sampled data
  • Actual success: how well it fits the full ground truth
  • The gap between the two quantifies the illusion. Two social learning strategies were compared: success bias (interacting with peers whose theories look most successful) and community bias (interacting within one's own community).

    Five Key Findings

    1. The illusion is everywhere

    Even without social learning, agents systematically overestimate their theories—a scientific version of the Dunning-Kruger effect. The illusion worsens sharply with problem complexity (correlation ρ=0.817, p<0.001).

    2. Success bias amplifies the illusion

    Under strong success bias, agents rapidly converge on a few "star theories," collapsing diversity. Perceived-vs-actual success gaps are significantly larger in high-bias conditions (Mann-Whitney U=143,688, p<0.001). Success bias does not create the illusion—it amplifies and entrenches it.

    3. Success bias can only lift the floor, not raise the ceiling

    Regression analysis shows success bias significantly helps the worst-performing theories (bottom 25%), has weak effects on middle-tier theories, and is negative for the best theories (top 25%). It filters out bad theories but cannot discover new, better ones—which requires diversity, not convergence.

    4. Rationally optimizing perceived success degrades real performance

    When agents freely tune their strategies to maximize perceived success, they settle on high success bias plus strong community clustering—the combination most damaging to actual success. The resulting knowledge inequality strikingly resembles citation distributions in real scientific communities, suggesting the Matthew Effect may partly be a natural product of amplified cognitive illusions in social learning.

    5. Evaluation accuracy only matters under high success bias

    The community's ability to correctly identify good theories (measured by CrPC—Centrality and Relative-Performance Correlation) is irrelevant without success bias, but critical when it is strong: a single misjudged star theory gets magnified through mass imitation. This creates a vicious cycle—stronger success bias increases dependence on evaluation accuracy while simultaneously undermining it.

    Historical Echoes

  • N-rays (1903): French physicist Prosper-René Blondlot and dozens of colleagues "confirmed" a new form of radiation—until Robert Wood removed a key prism and Blondlot still saw it. The fraud was sincere self-deception.
  • Amyloid plaque hypothesis: Since the 1990s, Alzheimer's research bet nearly everything on β-amyloid. Trial after trial failed—not because the direction was entirely wrong, but because the community converged prematurely on a local optimum.
  • In both cases, the problem was not stupidity or insufficient data, but a community that *believed it understood when it did not*.

    Implications

  • For individual researchers: confidence in one's own theories may be the biggest cognitive trap. Genuine humility means clear-eyed awareness of one's limits.
  • For science policy: metrics like citations, h-index, and impact factors are proxies for "perceived success" that are systematically distorted under success bias. A system rewarding only apparent success quietly destroys the possibility of the next breakthrough.
  • For scientific communities: protect space for heresy. Nearly every paradigm revolution was first seen as deviant. Premature convergence turns science into an echo chamber.
  • The authors conclude that progress may depend less on choosing the right theory and more on maintaining the diversity of explanations worth remembering—in a world richer, stranger, and less perfect than any single theory can capture.

    References

  • Dunning, D. & Kruger, J. (1999). "Unskilled and Unaware of It." *Journal of Personality and Social Psychology*.
  • Merton, R. K. (1968). "The Matthew Effect in Science." *Science*.
  • Sloman, S. & Fernbach, P. (2017). *The Knowledge Illusion: Why We Never Think Alone*.
  • Louis, A. W. & Dubova, M. (2026). "Nothing Deceives Like Success." arXiv:2604.27188.

Tags

#philosophy-of-science#agent-based-modeling#social-learning#success-bias#dunning-kruger-effect#scientific-progress#cognitive-bias#arxiv

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177619021