Meridia Insight Science Breakthroughs Knowledge

How Society Agrees on Who’s Good and Who’s Bad — Even When We Disagree

Even when people disagree, they still tend to agree — and that hidden correlation determines whether cooperation survives.

94% of people are seen as "good" — but only if everyone agrees. When judgments diverge, it drops to 22%.

Under a set of eight powerful social rules, 94% of people are seen as "good" by others — but only if everyone shares the same opinions. When individuals form private judgments, that consensus collapses to as low as 22%, and cooperation falters. Yet even in disagreement, people still tend to agree: if one person sees someone as good, others are up to 85% more likely to do the same. This hidden correlation, long unexplained, has now been mathematically pinned down — and it reveals how fragile moral consensus really is.

These numbers come from a new analysis of the "Leading Eight" norms, a family of social rules that sustain cooperation through reputation. For two decades, these norms have served as a gold standard in understanding how indirect reciprocity — helping others to build a good reputation — can hold societies together. But they were built on a fantasy: the idea that everyone sees each other the same way. In reality, reputations are private, fractured, and often conflicting. One person’s hero is another’s villain.

The breakthrough lies not in simulating millions of interactions, but in a simple equation that predicts how likely two observers are to agree on someone’s character — a quantity called . For the first time, researchers have derived a closed-form solution for this correlation across all eight norms, using only two variables: the average fraction of "good" assessments () and the conditional probability that one observer judges a person good given that another does (). This pairwise approximation slashes prediction errors by up to 70% compared to older models and recovers a known three-way classification of the norms — all without running a single simulation.

The implications stretch far beyond game theory. They speak to how online communities moderate content, how hiring committees evaluate candidates, and how justice systems weigh character testimony. In a world of algorithmic judgment and polarized discourse, understanding the mechanics of moral consensus isn’t just academic — it’s urgent.

The Science

The study, by Jong-Won Hong, Sunhee Chae, and Hye Jin Park at Inha University, revisits indirect reciprocity — a mechanism where cooperation persists not because people repay favors directly, but because they want to maintain a good reputation. In this model, every interaction involves a donor deciding whether to help a recipient, while others observe and update their assessment of the donor. These assessments — labeled simply "good" (G) or "bad" (B) — are stored in an image matrix, where each entry represents person ’s view of person

.

The rules governing these updates are the assessment norms. The Leading Eight, identified in 2004 by Ohtsuki and Iwasa, are the only known binary norms that sustain full cooperation under public assessment — that is, when all observers agree. They differ in how they judge actions based on the donor’s and recipient’s reputations. For example, all eight agree that helping a good person is good and refusing to help a bad person is acceptable. But they diverge on edge cases: what should we think of someone who helps a bad person? Or refuses to help one when they themselves are bad?

Under private assessment, each observer updates their view independently, and small errors — say, mishearing an action or misapplying a rule — accumulate. With probability , an observer flips their judgment, introducing noise into the system. This leads to divergent reputations: two people can look at the same individual and see opposite moral standings.

To quantify this, the authors use two key variables. The first is , the population-averaged goodness — the fraction of times any individual is judged good by others. Since donors cooperate only with those they see as good, also measures the level of cooperation in the population.

The second is , the conditional probability that one observer judges a player good given that another does. This captures assessment correlation: even when opinions differ, they’re not random. If Alice thinks Bob is good, Charlie is more likely to agree — not because they talk, but because they’ve seen Bob act the same way.

Prior work by Murase and Hilbe (2024) showed that and together determine whether cooperation is stable. But while could be predicted from a self-consistency equation, had no analytical solution — it had to be measured from simulations. This left a gap: we knew these two variables mattered, but we didn’t know how they emerged from the norms themselves.

Hong, Chae, and Park close that gap with a pairwise approximation. They assume that the joint probability of three assessments — say, donor, recipient, and observer — can be factored into pairwise conditionals. Specifically, they approximate

which means that two observers’ views of the same player are conditionally independent given that player’s true behavior. This allows them to derive a closed equation for that, when paired with the equation for , fully determines the system’s steady state.

The model assumes a well-mixed population of individuals, with random donor-recipient pairs chosen at each step. Actions are error-free, but assessment errors occur with probability , ranging from 0.001 to 0.1. Simulations run for Monte Carlo steps to ensure convergence.

What They Found

The results reveal a stark divide in how the Leading Eight perform under private assessment. At low error rates (), the two-variable model predicts with remarkable accuracy — but with telling exceptions.

For norms L1, L3, L4, and L7, the model’s predictions align almost perfectly with simulations. For instance, under L1 — the most forgiving norm — the predicted is 0.94, matching the simulated value. Under L3, it’s 0.89, again spot-on. These norms maintain high cooperation because they allow redemption: even if someone is mistakenly judged bad, they can regain goodness by helping a good recipient.

But for L2 and L5, the model underestimates cooperation by as much as 15 percentage points. At , simulations show for L2, while the model predicts only 0.50. The gap narrows as increases, disappearing entirely at .

Stationary goodness as a function of assessment error

Simulated and predicted average fraction of 'good' assessments across the Leading Eight norms

Stationary goodness as a function of assessment error
LabelValue
L1 Sim0.94 %
L1 Two0.94 %
L2 Sim0.65 %
L2 Two0.5 %
L3 Sim0.89 %
L3 Two0.89 %
L6 Sim0.22 %
L6 Two0.23 %

Figure: Stationary goodness as a function of assessment error . The two-variable model (green) tracks simulations (purple) closely for most norms, but underestimates cooperation for L2 and L5 at low error rates.

The same pattern holds for . The model predicts that if one observer sees someone as good, another observer will agree with probability between 0.6 and 0.9, depending on the norm. For L1, , meaning strong consensus. For L6 and L8, it’s closer to 0.55 — barely above chance.

Conditional probability of agreement

Probability one observer judges a player good given another does

Conditional probability of agreement
LabelValue
L10.85 %
L20.6 %
L30.8 %
L60.55 %
L80.56 %

Figure: Stationary conditional probability as a function of . The model (green) captures the trend but underestimates agreement for L2 and L5 at low error rates.

The authors also compute the covariance between observers’ assessments:

This quantity measures how much observers agree beyond random chance. For L1, , indicating strong consensus. For L6 and L8, , suggesting near-independent judgments.

Covariance of assessments

Measure of agreement beyond random chance

Covariance of assessments
LabelValue
L10.08 unitless
L20.05 unitless
L30.07 unitless
L60.01 unitless
L80.01 unitless

Figure: Covariance between observers’ assessments. The model reproduces high consensus norms well but underestimates agreement in L2 and L5.

Crucially, the model recovers the three-class structure of the Leading Eight identified by Fujimoto and Ohtsuki (2024):

  • Class 1 (L1, L3, L4, L7): High cooperation, high consensus
  • Class 2 (L2, L5): Moderate cooperation, fragile consensus
  • Class 3 (L6, L8): Low cooperation, weak consensus

This classification hinges on how a norm judges cooperation with a bad recipient. Norms that punish such cooperation (like L2 and L5) are more sensitive to assessment errors, leading to fragmented reputations and lower .

Why This Changes Things

For two decades, the Leading Eight have been treated as a monolith — eight equally valid paths to cooperation. This work shatters that illusion. Under private assessment, they behave radically differently. Some sustain near-universal goodwill; others fracture into echo chambers of moral disagreement.

The real breakthrough is not just predicting , but doing so analytically. Previous models either ignored correlations (mean-field) or required solving complex integral equations. This new framework reduces the entire dynamics to two coupled equations — a dramatic simplification that makes the model usable for real-world design.

Consider online platforms. Reddit, Wikipedia, and Stack Overflow all rely on reputation systems to moderate behavior. But they assume public consensus: upvotes are visible, karma is shared. In reality, moderation is often private — admins see things users don’t, algorithms make hidden judgments. This model shows how such systems can silently diverge, with different users seeing the same contributor as hero or troll.

Or consider hiring. A job candidate’s reputation is built from private references, internal reviews, and algorithmic screening. If different evaluators use subtly different norms — one forgives past failures, another doesn’t — the resulting assessments may correlate weakly, leading to unfair outcomes. This model quantifies that risk.

Even in law, character testimony plays a role. If judges and juries apply different moral rules, their assessments of a defendant’s character may disagree — not due to bias, but due to the structure of the norms themselves. This work suggests that some rules inherently produce more consensus than others.

The fact that L2 and L5 — norms that punish helping bad actors — perform worst under private assessment is deeply suggestive. It implies that moral rigidity undermines social cohesion when information is imperfect. In a noisy world, forgiveness isn’t just kind — it’s functional.

What’s Next

The model’s failure on L2 and L5 at low error rates is not a bug — it’s a clue. The authors suggest that higher-order correlations, beyond pairwise, may be at play. In other words, the agreement between three observers isn’t fully determined by their pairwise relationships. This could reflect clustering, echo chambers, or feedback loops in assessment dynamics.

Future work could test this by measuring third-order correlations in simulations or real-world data. Social media platforms, with their rich logs of user interactions and moderation decisions, could provide a natural testbed.

Another frontier is dynamic norms. The Leading Eight are fixed rules, but real societies evolve their moral codes. Could a population learn to adopt L1 over L2 when private assessment dominates? Machine learning models of norm evolution could explore this.

Finally, the model assumes binary reputations. Real judgments are graded: "trusted," "suspended," "banned." Extending this framework to multi-level reputations — or even continuous scores — would bring it closer to platforms like eBay or Uber.

The biggest challenge may be empirical validation. We lack large-scale data on how people form and update reputations in private. But as digital interactions leave ever-denser traces, that may change. When it does, this model offers a ready-made toolkit for diagnosing the health of a community’s moral ecosystem.

For now, the message is clear: agreement is not guaranteed. Even under the best rules, cooperation depends on how we see each other — and how much we expect to agree. In a world of private judgments, the most resilient societies may not be the most moral, but the most forgiving.

Comments (0)

No comments yet. Be the first to share your thoughts.