For five lessons we’ve handed common knowledge a great deal of power: it turns a room full of private doubts into a public revolt, tips a healthy bank into a run, and lets strangers converge on the same street corner without a word. Time to turn the model over and inspect its underside. Because the honest truth is that true common knowledge — the infinite tower of “I know that you know that I know…” — is almost never literally achieved. Humans run on a cheaper, fragile approximation of it, and knowing exactly where that approximation cracks is what separates someone who uses the model from someone the model uses.
Before you read — take a guess
Guess: in real human coordination, how high does the tower of 'I know that you know that I know…' actually get climbed?
Common knowledge is an idealisation
Picture a ladder that is supposed to reach the moon. In the textbook, it has infinitely many rungs and everyone has climbed every one. In real life, people step on the first two or three rungs — I know it, you know it, I’m pretty sure you know I know it — glance up, and simply assume the rest of the ladder is there. That assumption is doing enormous quiet work.
What people actually have is almost-common knowledge: a fact is known by everyone, and known to be known up to some finite depth, after which each person merely presumes the recursion continues without ever checking. Formally, if true common knowledge is “known at every level to infinity”, almost-common knowledge is “known and mutually known up to level , then taken on faith”. For the vast majority of coordination, level 2 or 3 plus faith is indistinguishable from the real thing — the coordination fires anyway.
The catch is the faith. It holds until something makes a player doubt one rung, and because every rung leans on the one above it, a single crack can bring the whole assumption down. That is why coordination built on almost-common knowledge is usually fine but structurally brittle at the edges.
The word that does the quiet work: approximate
Almost-common knowledge is knowledge shared and mutually known up to a few levels, then assumed to continue forever. It is what humans actually run on. It behaves like the real thing right up until someone has a reason to doubt a single level — at which point the borrowed certainty above it can vanish all at once.
Name what humans actually run on.
Pick the right option for each blank, then check.
True common knowledge needs an recursion nobody ever verifies. In practice people rely on knowledge — known up to a few levels, then simply to continue — which usually suffices but is at the edges.
The coordinated-attack problem: why confirmation never closes
Here is the cleanest proof that the infinite tower cannot be built by talking. Two allied generals sit on hilltops on opposite sides of a valley, with the enemy in between. If they attack together at dawn, they win. If only one attacks, that lone army is destroyed. They must agree on the timing — but the only way to communicate is to send a messenger through the valley, where the enemy might intercept them.
General A sends: “Attack at dawn.” But A cannot move until she knows the message arrived — a lone charge is suicide. So B, on receiving it, sends an acknowledgement: “Got it — dawn.” Now B has the plan… but B cannot move either, because B doesn’t know his acknowledgement got through. If it was intercepted, A never marches, and B charging alone dies. So A must acknowledge the acknowledgement: “Got your confirmation.” And now A is in exactly B’s earlier position — A can’t be sure that message landed. Every acknowledgement creates a new final message whose delivery is uncertain, and whoever sent the last one can’t safely act on it.
The trap in one line
Whenever the last message sent might have been lost, its sender cannot rely on it — so no finite exchange of confirmations ever reaches common knowledge. The tower needs an acknowledgement for every acknowledgement, forever.
Let’s tabulate the regress so the impossibility is undeniable. Each row adds a level of knowledge, and each row still leaves someone holding an unconfirmed final message:
| Message | After it (if delivered) | Who still can’t be sure |
|---|---|---|
| A → B: “Attack at dawn” | B knows the plan | B — did A learn I received it? |
| B → A: “Confirmed” | A knows B knows | A — did B learn I got the confirm? |
| A → B: “Got your confirm” | B knows A knows B knows | B — did A learn I got that? |
| … forever | one more level of knowing | always the last sender |
No finite number of rounds ever removes the doubt; it only pushes it up one level. Since real channels can fail, common knowledge here is genuinely unreachable, and the generals — each rational, each willing to win together — may never dare to attack.
Rubinstein’s electronic-mail game
You might hope automation rescues the generals: let computers fire confirmations back and forth thousands of times a second. Ariel Rubinstein’s electronic-mail game shows it doesn’t. Two players benefit from jointly taking a risky action, but only in one “state of the world” that just one of them observes directly. Their computers automatically confirm each message, and confirm the confirmations — building a tower hundreds of levels high. Yet because there’s always a tiny chance the last automatic message was the one that got lost, the reasoning unravels from the top down exactly as with the generals. The startling result: no matter how many automatic confirmations pass, the players still fail to coordinate on the risky action. A hundred levels of “I know that you know” is not common knowledge, and the gap between “very high finite tower” and “actually infinite” is the whole difference between coordinating and not.
In the two-generals problem, why doesn't sending one more acknowledgement ever solve it?
False common knowledge: manufactured, not true
Now the pitfall that turns the model from a tool into a trap. A public signal manufactures common knowledge about what everyone will act on — and it does this whether or not the underlying claim is true. The bank in lesson 5 didn’t run because it was insolvent; it ran because a rumour became common knowledge that others would run, and pulling your money first is rational once you believe that, solvent bank or not.
Call it false common knowledge: a state where “everyone knows that everyone knows ” has been established by a public signal, while itself is false. The recursion is perfectly real; only the content is wrong. And crucially, false common knowledge is just as behaviourally potent as the true kind — people coordinate on it identically, because what drives their action is the belief about others, not the buried fact.
Worked example. A viral post claims a popular payment app is “about to freeze all withdrawals at midnight.” It’s fabricated. But it’s public — millions see it, and each viewer sees that millions of others saw it. Now everyone knows that everyone knows about the alleged freeze. The rational move, given that belief, is to withdraw now, before the crowd does. The stampede of withdrawals then really does strain the app — and the false claim becomes true because it was commonly believed. This is reflexivity wearing a mask: the signal manufactured the reality it described.
The trap: acting on it is not evidence it's true
Never read “everyone is acting on ” as ” is true.” A public signal creates common knowledge about coordination, not about facts. Rumour cascades, moral panics, manufactured consensus, and viral misinformation all run on genuine common knowledge attached to a false core. The recursion is real; the content can be garbage.
A stock forum erupts: 'Everyone knows this company is finished — look how many people are selling!' The price craters. What's the reasoning error a careful thinker spots?
Not every coordination needs FULL common knowledge
Having warned that true common knowledge is unreachable, we must avoid the opposite overcorrection — concluding that coordination is therefore impossible. It plainly isn’t; people coordinate constantly. The resolution is that most coordination never needed the full infinite tower in the first place.
Two cheaper substitutes usually do the job. First, a focal point — a solution so obvious it needs no communication (meet “at noon”, split “50/50”, rally at “the main square”) — lets people converge on almost-common knowledge alone, because each correctly assumes the other’s mind runs the same obvious way. Second, a little critical mass inside a dense-enough core can be enough to tip the rest: approximate common knowledge within one tightly connected cluster — a friend group, a trading desk, a neighbourhood — can start a cascade that pulls in the loosely connected periphery afterward. You don’t need the whole society to simultaneously know that everyone knows; you need a connected seed where it’s almost true, plus a threshold structure that lets the seed spread.
Enough, not perfect
Coordination rarely requires literal common knowledge. A focal point lets minds converge on the obvious with no messages at all; a dense critical-mass core lets approximate common knowledge among a connected few ignite the many. The model’s demand for an infinite tower is the idealisation — real tipping runs on ‘enough,’ not ‘perfect.‘
Sort each situation by whether its coordination rests on TRUE common knowledge (a genuine, verifiable public fact) or FALSE common knowledge (a public signal attached to an untrue core).
Place each item in the right group.
- A fabricated viral post triggers a withdrawal stampede on a solvent bank
- A televised referee blows the final whistle; both teams and every fan know the match is over
- A rigged "everyone agrees" chorus manufactures the sense of a consensus that does not exist
- A published, verifiable timetable makes "the train leaves at 9:00" common knowledge
- An official, broadcast election result both sides can independently verify
- A moral panic spreads a debunked rumour that "everyone knows" a danger is real
Where the model lies — and how it gets misused
Every powerful model has a failure surface. Common knowledge’s three biggest misuses share a root: treating a subtle, degree-laden idea as a blunt, binary one.
| The misuse | Why it’s wrong | The corrective |
|---|---|---|
| Treating common knowledge as binary (“we have it or we don’t”) | It’s a matter of degree — depth of recursion and density of the network both vary continuously | Ask how common: how many levels, how connected the core, how confirmable |
| Assuming a public signal ALWAYS ignites coordination | A signal only tips if the underlying grievances or thresholds already exist (lesson 5) — shout “revolt!” to the content and nothing happens | Check the private preferences and thresholds before crediting the signal |
| Forgetting false common knowledge is a weapon | Manufactured consensus and rumour cascades are engineered deliberately, not just accidental bugs | Treat “everyone knows” as a claim to audit, especially when someone benefits from you believing it |
The through-line: common knowledge is not a switch, it’s a landscape with depth, density, and truth-value as separate dials. Whenever you catch yourself saying “it’s common knowledge, therefore…”, stop and ask which dial you’re actually leaning on.
Once a public signal has made a false claim common knowledge, a single person knowing it’s false doesn’t fix anything — because their best action still depends on what everyone else will do, and everyone else is still acting on the signal. Debunking the fact privately doesn’t dissolve the coordination; only a counter- signal that is itself public — one that makes “it’s false, and everyone now knows that” common knowledge — can. This is why the antidote to manufactured common knowledge is rarely quiet truth and almost always a louder, more credible public signal. The recursion that trapped the crowd can only be undone at the same level it was created.
The latticework — how this whole course connects
Common knowledge isn’t a lone gadget; it’s a joint that links several models you already hold. Wiring it in is what makes it durable.
- Critical mass & tipping points. Common knowledge is the informational engine behind a threshold cascade. A public signal doesn’t add protesters directly — it tells each person that enough others will show up to cross their personal threshold, so they all move at once. Thresholds are the fuel; common knowledge is the spark.
- Nash equilibrium & equilibrium selection. Coordination games have multiple equilibria (everyone stays home / everyone revolts; keep your money in / run). Common knowledge is the mechanism that selects which equilibrium the group lands on — it doesn’t change the payoff matrix, it moves everyone’s beliefs to the same corner of it.
- Reflexivity & self-fulfilling dynamics. False common knowledge is reflexivity with a public address system: a belief about what others will do reshapes reality until the belief comes true. The healthy bank that gets run on, the app whose fabricated freeze causes a real one — the signal manufactures the fact it claimed.
- Incentives & preference falsification. Before a public signal, people hide their true preferences because each privately fears being the only one to act. The signal doesn’t create the grievance — it reveals that the grievance was common all along, flipping falsified preferences into shared, actionable knowledge.
- Forward pointer: information cascades. Coming up in your latticework is a close cousin — cascades where people rationally ignore their own information to copy the visible actions of others. Common knowledge is about aligning beliefs about beliefs; cascades are about inferring facts from behaviour. They interlock: watch for where one shades into the other.
Match each concept from this lesson to its precise meaning.
Pick a term, then click its definition.
When to reach for common knowledge (and when not to)
- Is coordination the bottleneck? Reach for the model when everyone wants the same outcome but each is waiting on the others. If the problem is conflicting interests, not aligned-but-uncertain ones, use game theory more broadly.
- Is there a genuine public signal? The model bites when something is seen and seen to be seen. No public signal, no manufactured common knowledge.
- Do the underlying thresholds exist? A signal only tips a crowd whose private grievances or incentives already reach their thresholds. Check the fuel before crediting the spark.
- Am I confusing coordination with truth? If your only evidence for a claim is that everyone’s acting on it, you may be looking at false common knowledge. Audit the core separately from the crowd.
- Is the channel reliable enough? Where confirmation can silently fail — the two-generals regime — expect coordination to stay brittle no matter how many messages fly.
Big picture
Where common knowledge breaks
- Limits & latticework
- It is an idealisation
- Infinite tower never literally built
- We run on almost-common knowledge
- Unreachable by talking
- Two-generals: confirm the confirmation, forever
- Rubinstein e-mail game: high finite ≠ infinite
- False common knowledge
- Public signal ignores truth
- "Everyone acts on it" ≠ "it is true"
- Enough, not perfect
- Focal points and a dense core often suffice
- Latticework
- Critical mass, Nash, reflexivity
- Preference falsification → cascades
- It is an idealisation
Key takeaways
- True common knowledge is an idealisation. The infinite recursion is never literally built; humans run on almost-common knowledge — a few verified levels plus faith — which usually suffices but is fragile at the edges.
- Confirmation never closes the loop. The coordinated-attack problem and Rubinstein’s e-mail game prove that over a failable channel, no finite exchange reaches common knowledge — coordination stays brittle where messages can be lost.
- A public signal manufactures common knowledge regardless of truth. False common knowledge — rumours, panics, runs on healthy banks — coordinates people just as forcefully as the true kind. Never read “everyone acts on it” as “it’s true.”
- You rarely need the full tower. Focal points and a dense critical-mass core let approximate common knowledge do the work; the model’s demand for infinity is the idealisation, not the requirement.
- Common knowledge is a dial, not a switch — depth, density, and truth-value are separate — and it’s the joint connecting critical mass, Nash equilibrium selection, reflexivity, preference falsification, and the information cascades coming next.
You now hold the whole course: the ladder from private to mutual to common knowledge, how public signals and focal points move a crowd to the same equilibrium, how bank runs and revolutions ignite — and, from this lesson, exactly where the model thins out, how it’s faked, and how it wires into the rest of your latticework. Carry the humility with the power into the final exam.