Share of people sharing, by round
Click or drag to move through the rounds.
What people are passing on this round
illustrative wordingShare of people sharing, by round
Click or drag to move through the rounds.
What people are passing on this round
illustrative wordingEach dot is a person, and each line a contact who can see what they post. In round 0 a few people see an accurate report. In every round, a person who is not yet sharing looks at their contacts. If the share of contacts they see sharing is larger than their private conviction Δ, the cost to them of repeating something they doubt, they start sharing too. They pass on the most alarming version they saw, and sometimes step it up a level.
Fear changes both steps. A frightened person puts more weight on each contact who is sharing, and is more likely to embellish. A correction goes out once, to a random share of people. Sharers it reaches weigh it against what the claim means to them and how many of their contacts still share it. People it reaches before they share become harder to recruit.
The model is a teaching device. These findings are what it should be read against.
Models
Literature
The Mass Game (Walberg and co-author) pits an individual against the masses. θ is the share of the group endorsing the claim and does two jobs at once: it is the benefit of joining the wave and the ostracism cost of holding out. Δ is private conviction. If the masses share, the individual gets θ − Δ from sharing and −θ + Δ from holding out, so the best response is to share if and only if Δ < θ. In a population with convictions drawn from F, the share evolves as θt+1 = F(θt), and full conformity is absorbing.
On a network each person sees only their own contacts. Person i shares when Δi < θ̂i, with θ̂i = min(1, λ·mi/ki + a), where mi of their ki contacts are sharing. The paper treats θ̂ − θ, the gap between perceived and true endorsement, as the lever a disinformer pulls; a is that lever. The fear gain λ comes from Walberg’s Narrative Warfare paper, where fear raises the weight a person puts on evidence and, on a network, “catching a neighbor’s emotion raises one’s own gain on all subsequent private signals.” Scaling the observed share by λ is this tool’s simplification of that proposition, not the paper’s own network model. Sharing is absorbing: people do not quietly stop, they only retract.
Convictions are LogNormal with median Δ and shape σ (the paper’s numerical section uses shape 0.5), drawn once per seed so that moving a slider shifts every person together. A person who starts sharing passes on the most alarming version among their sharing contacts and steps it up one level with probability min(0.95, ε·λ).
The correction follows Walberg’s Sticky Affect model. Belief is revised only if α(E)·evidence > (1 − α(E))·UI + C, where α(E) is the weight on accuracy, which falls as emotional activation E rises, UI is the identity value of the claim, and C = g(N, S) is the cost of abandoning it, rising in narrative coherence N and social reinforcement S. Here E = (λ − 1) + v/4 for a sharer holding version v, α(E) = 1/(1 + E), S is the share of the person’s contacts still sharing, and N grows with the rounds they have held the claim, saturating at 10. An in-group voice carries three times the evidence weight of an out-group fact-check, following the roughly 3:1 ratio in the Moral Panic paper’s falsification sweep. People reached before they share add α(E)·evidence to their conviction.
Everything numeric in this tool is a teaching value, not an estimate. The outputs are properties of the model under these values, not findings about any real audience, platform or event.
Nine players with intensity pairs (xk, yk) from (2, 10) to (10, 2) choose Left or Right. Each earns yk per same-side partner on the Left and xk on the Right. From the paper’s starting split (players 1–5 Left), player 6 moves first because 7 × 3 = 21 < 5 × 5 = 25. Under “All who gain switch,” the card reproduces the paper’s four-round payoff table exactly. The paper defines a one-at-a-time rule; “Largest gain first” applies it and reaches the same unanimous end state in five rounds, with players 8 and 9 moving separately.
Randomness comes from a seeded mulberry32 generator. The run with and without the correction share the same draws, so they are identical until the correction round.