TRANSMISSION // 2611.22484v1

When incentives make consensus less informative

Source published

A theoretical social-learning study finds that rewards for endorsing one account can strengthen or weaken false consensus, depending on private evidence and how endorsements update beliefs.

AI-generated conceptual illustration for When incentives make consensus less informative
AI-generated conceptual illustration; not documentary evidence.

Why might a group settle on a false account of who caused a harm, even when people have private evidence and understand the incentives behind others’ statements? A model by Ronen Gradwohl, Arye Hillman and Rok Spruk shows that rewards for endorsing one side have two effects: they make that endorsement more tempting, while changing how much observers can learn from it.

The puzzle: what does agreement tell us?

A public consensus can be wrong about responsibility for a harm. The paper calls this “moral inversion”: those responsible are approved, while those harmed are blamed. Its examples include the Dreyfus affair and racial terror lynchings in the United States. The authors focus on one possible ingredient in such cases: people may gain an expressive, reputational or professional benefit from endorsing a particular account.

The question is not simply whether a reward makes people more likely to repeat the favored account. It does. The harder question is what that change does to the information carried by each endorsement—and, eventually, to the likelihood and trustworthiness of group agreement.

A model of people learning in sequence

The authors build a theoretical model, not an experiment with participants or an analysis of historical records. The true state is one of two possibilities, H or L. People act one after another. Each sees earlier endorsements and receives a private signal about the state, then endorses H or L.

An endorsement of the correct state brings a benefit, and endorsing L brings an additional common expressive payoff. People know this incentive and the model’s assumptions. The researchers hold the private-signal system fixed while varying the payoff for L.

As in standard models of social learning, early endorsements can influence later choices. Eventually, people may enter a “cascade”: they all choose the same side, even though later individuals’ private signals no longer affect their actions. The paper examines how often L becomes the consensus when H is true, and how often an L consensus is actually correct.

Two effects pull in different directions

A larger reward for endorsing L changes both behavior and the meaning of behavior. First, it makes L more attractive, so an incorrect L consensus can become more likely. Second, it changes what endorsements reveal: an L endorsement provides weaker evidence in favor of L, while an H endorsement provides stronger evidence for H. The weaker evidence from L can sometimes keep learning going rather than trigger a cascade immediately.

Under a condition the authors call “single-peakedness,” reference private beliefs—those calculated from a prior of one-half—are most concentrated around one-half, with density that does not increase as beliefs move farther away. The condition includes uniform, triangular and symmetric truncated-normal distributions. It does not mean beliefs are necessarily concentrated around the model’s actual prior. Under this condition, the analysis finds the intuitive outcome: increasing the expressive payoff raises the chance of an incorrect consensus on L and reduces the reliability of an L consensus.

But that conclusion is not universal. When the condition fails, beliefs can move past the point where a cascade would begin, a feature the authors call “overshooting.” They give a constructed example in which a higher payoff makes an initial L endorsement less conclusive. More evidence is then needed before an L cascade forms. In that example, the incorrect L consensus becomes less likely and an L consensus becomes more reliable. The paper also shows that, in this example, the public record under the higher payoff is more informative in the formal Blackwell sense: it can be used to reproduce the lower-payoff record through a state-independent transformation, but not vice versa.

The reversal is not a general prediction for every case outside single-peakedness. A further theoretical result says that, for a fixed private-information structure, a sufficiently large increase in the payoff restores the benchmark direction: the favored incorrect consensus becomes more likely, and its reliability falls. A smaller increase can therefore have a different effect from a larger one in the same information environment.

What the results do—and do not—show

The paper clarifies a trade-off in a particular sequential-learning framework. An incentive can make an account more popular while simultaneously changing how diagnostic endorsements are. Which effect dominates depends on the distribution of private evidence and the size of the incentive change.

This is a mathematical result under stated assumptions, not evidence that expressive rewards caused any particular historical episode, or that real observers always update as the model assumes. The authors say the model leaves out features of historical moral inversion, including coercion, prejudice and suppression of evidence. It isolates a common, understood reward for endorsing one account; it does not represent every source of public pressure or misinformation.

The source is an arXiv preprint. The material provided does not establish whether it has been peer reviewed. The manuscript includes proofs and technical details, but its conclusions remain conditional on its assumptions and model. The findings are best read as showing how consensus can conceal information—not as a direct diagnosis of any real-world dispute.

Reviewed cross-references

These connections were reviewed by a model, not independently fact-checked by a human. Similarity does not imply agreement; consult the original papers and the evidence below.

  • Why rational choices can change faster than the world they track

    Shared method: Both use sequential social-learning models in which agents observe predecessors’ actions and receive private signals. The comparison is limited: the origin studies how an expressive payoff affects consensus with a fixed underlying state, while Huang studies action volatility as the state changes over time.

    Model review: · openai-codex/gpt-6-luna

    Evidence and provenance

    Selected excerpts, not complete papers. PDF extraction can omit figures or damage equations, tables and reading order; sources may be preprints.

    • 2611.22484v1 — Original paper

      Excerpt reference: static/papers/2611.22484v1.md:285-404
      Basis: Paper text
      Source SHA-256: f78c5f14478e50d9b4fa277076c9f5e09d5035d7c7940b95d87d3a46157768b8

    • 2208.14570v3 — Original paper

      Excerpt reference: static/papers/2208.14570v3.md:98-183
      Basis: Paper text
      Source SHA-256: 2704559875625da9e0666b1aa96d3e24d1b7baf1b88f2084a5e620343fd22ff8

  • When does an information score really reward better information?

    Shared method: Both use the Blackwell order to compare informativeness: the origin applies it to the public record generated by social learning, while Cheng and Kim characterize when information measures respect Blackwell and Lehmann rankings. The latter is a general characterization, not evidence about the origin’s model or its comparative-statics results.

    Model review: · openai-codex/gpt-6-luna

    Evidence and provenance

    Selected excerpts, not complete papers. PDF extraction can omit figures or damage equations, tables and reading order; sources may be preprints.

    • 2611.22484v1 — Original paper

      Excerpt reference: static/papers/2611.22484v1.md:15-53
      Basis: Paper text
      Source SHA-256: f78c5f14478e50d9b4fa277076c9f5e09d5035d7c7940b95d87d3a46157768b8

    • 2404.15158v5 — Original paper

      Excerpt reference: static/papers/2404.15158v5.md:74-147
      Basis: Paper text
      Source SHA-256: 6cec2096e7ba7c50658f6eb0bedffee7bd50d784329dcb61ac8cb39869026c34

Automatically matched by shared topic keywords. These links do not imply that the sources agree or establish a consensus.