What the model says
Rumora is not an illustration of a conclusion — it is the thing that produced the conclusions. Every number below came out of the simulator on this page, averaged over 30 independent runs on freshly generated networks. You can reproduce any of them with the controls above.
Timing beats strength, and it is not close. A correction one-quarter as persuasive, reaching half as many people, released at tick 8 instead of tick 55, held the claim to 25% of the network instead of 97%. Released after the peak, correction strength stopped mattering entirely: quadrupling it left final reach unchanged to the nearest percentage point.
Why a late correction cannot win
The instinct is that a fact-check fights the rumour for the people who already believe it. It mostly does not. In the model — as in the literature it draws on — a correction's real work is inoculation: reaching people the claim has not gotten to yet, so that when it arrives they are 55% less likely to take it up.
That makes the value of a correction proportional to how many people are still unaware. Release it after the peak and that reservoir is gone. There is nobody left to protect, and the people who remain have spent the intervening ticks hardening.
| Correction released | Strength 0.2 | Strength 0.5 | Strength 0.8 |
|---|---|---|---|
| tick 8 — before the wave | 25% | 8% | 6% |
| tick 25 — during the climb | 79% | 68% | 60% |
| tick 55 — after the peak | 97% | 97% | 97% |
Final share of the network that ever believed the claim; lower is better. Read down a column: the same intervention loses most of its value by waiting. Read across the bottom row: after the peak, strength buys you nothing at all.
Amplification does not make true things spread. It lowers the bar for everything. A claim weak enough to die in an ordinary social network — reaching a median 6% of people — reached a median 97% when the feed showed it to strangers. Nothing about the claim changed.
The mechanic is specific, and worth separating from "the algorithm is bad". Amplification does two distinct things here. It biases who sees a post toward already-popular accounts, which is mostly a redistribution. And it delivers posts outside the sender's network entirely — which is not a redistribution, because it defeats the clustering that would otherwise contain a weak claim inside the neighbourhood it started in.
The second mechanic is the one that matters. Clustered networks are naturally resistant: a claim burns through a tight group, everyone has already heard it, and the spreaders lose interest before reaching a bridge. Out-of-network delivery hands the claim a fresh audience precisely when it was about to suffocate.
Near the threshold, the average outcome is a fiction. On identical settings, the 10th-percentile run reached 1% of the network and the 90th-percentile run reached 90%. The mean — 59% — describes almost none of the runs that actually happened.
This is why the sweep exists. Contagion near its critical point is bimodal: the claim either finds a bridge in the first few ticks or it does not, and that early coin-flip decides everything downstream. Load the Tipping point scenario and run 40 trials — the histogram comes out with two humps and a valley between them.
It has a blunt practical consequence. Two claims with the same underlying virality can produce a national story and a forgotten post, and the difference between them may be nothing more interesting than which three people saw it first. Post-hoc explanations of why a particular thing went viral are, quite often, explanations of a coin flip.
Rumours die of boredom, not refutation. With corrections switched off entirely, raising only the rate at which people lose interest — from 0.05 to 0.4 — cut final reach from 98% to 6%. Nobody was persuaded of anything. They just stopped bringing it up.
The engine borrows its stopping rule from Daley and Kendall's 1964 rumour model, and it is the piece most often missing from intuitions about misinformation. A spreader does not stop because they were corrected. They stop because they discovered the news was stale — everyone they told already knew — and repeating it stopped feeling like telling someone something.
Which means the most powerful lever in the whole simulation is not persuasion. It is whether the claim keeps finding rooms where it is still news. That is exactly the resource amplification manufactures, and exactly the one tight-knit clustering denies.
The finding that came out weaker than expected
The backfire effect — where a failed correction leaves someone more convinced — is in the model, and it does what it says. But driving it from zero to maximum did not change how far the claim travelled at all: 97.1% against 97.2%. What it moved was who was left unreachable (70% entrenched against 73%), how many the correction won back (28% against 27%), and how long the argument dragged on (121 ticks against 136).
So the effect is real here, and it is narrow: backfire does not spread a claim, it hardens and prolongs one. That happens to match where the empirical literature landed after the early results failed to replicate cleanly. I have left it un-juiced rather than tuning it into a better story — a slider that produced a dramatic swing would have been easy to build and would have been a lie about what the mechanism does.
How the model works
Every agent is in one of four states. Unaware has not heard the claim. Believer heard it, believes it, and is still repeating it. Doubter heard it and rejects it, and may push the correction. Fatigued still believes it but has stopped bringing it up — a silent carrier, and the reason a wave can come back.
Agents are not switches. Each carries a conviction that rises every time the claim is reinforced by someone they know, and conviction does three things: it makes them talk about it more, it makes them harder to correct, and it makes them resistant to losing interest. That single variable is what turns a crowd that merely heard something into a crowd that cannot be talked out of it.
What each control changes
- Virality — base chance a single telling lands on someone who has not heard it.
- Social proof — how much a listener is swayed by the share of their own contacts already repeating it. This is complex contagion: hearing it from four friends is not four times one friend.
- Repetition — the illusory-truth effect. Each failed exposure makes the next one more likely to land.
- Scepticism — mean resistance across the population; each agent draws its own value around it, so there are always credulous and stubborn tails.
- Amplification — how much the feed intervenes: popularity bias among contacts, out-of-network delivery, and a wider audience per post.
- Losing interest — the Daley–Kendall stifling rate. How fast a spreader gives up once the news stops being news.
- Correction strength / release tick / reach — how persuasive the fact-check is, when it lands, and how much of the network it is seeded into.
- Backfire — how much a failed correction hardens its target instead of moving them.
Networks
Topology is generated fresh for every run: Watts–Strogatz small world (clustered, with a few shortcuts — the closest thing to offline acquaintance), Barabási–Albert scale-free (a few hubs hold most of the ties), a stochastic block model for echo chambers, and Erdős–Rényi random as a null case. A node's audience per post scales with the square root of its degree, which is what makes a hub an actual superspreader rather than just a well-informed listener.
Honest limitations
This is a model, which is to say it is wrong in useful ways. Ticks are not hours. Nobody changes their mind for reasons of their own. There is one claim, never a competing pair. Networks are static — in reality, believing a thing changes who you are connected to, and that feedback loop is probably the most important thing missing here. Treat the numbers as statements about the model's logic, not measurements of the world.
Play it
There is a strategy game built on this. You are the claim, and you are trying to get enough of the world to believe something firmly enough that correcting it stops being possible.
Play Rumora Free, no account, runs in the browser · play.rumora.appIt runs the same mechanics as the model above, compressed from individual agents to population fractions so that fourteen regions can run at once. The two are relatives rather than the same program: its constants are tuned for a game, and every number quoted on this page belongs to the agent model here, not to the game. What the game inherits is the logic — that entrenched belief is what matters rather than reach, that a claim dies of boredom rather than refutation, and that suppression in a low-trust, polarised place hardens people instead of moving them.
Playing the spreader is the point rather than a provocation. It is the same reasoning behind Cambridge's Bad News: naming a technique is what makes it recognisable when it is used on you.
The source
A claim about how something behaves is worth nothing if nobody can check it, so the methods page states the model precisely enough to argue with, lists the parameters every published figure was produced at, and shows how to regenerate them.
It also links the source directly. This site is static and unminified, so it serves the very files the simulator above is running — no repository to clone and nothing to take on trust. The model is MIT licensed; the game is source-available under a non-commercial licence.
If you find an error in the model, it is an error in the findings too, and I would like to know.
Built on Daley & Kendall (1964) for the stifling dynamic, Granovetter thresholds and Centola's complex contagion for social proof, and the prebunking literature — notably van der Linden and colleagues — for inoculation. Any errors in translating those into code are mine.