Epistemic Logic

There's a puzzle that looks trivial and is actually a masterpiece of logical reasoning. On an island, there are N people with blue eyes. Everyone can see everyone else's eyes but not their own. A visitor announces: "At least one of you has blue eyes." Anyone who can deduce their own eye color must leave on the midnight ferry. Nobody leaves for N-1 nights. On the Nth night, all N blue-eyed people leave simultaneously.

The puzzle is: what did the visitor's announcement add? Everyone already knew someone had blue eyes — they could see N-1 pairs of blue eyes. So how could a statement of something already known change anything?

Common Knowledge vs. Mutual Knowledge

Terry Tao's formalization of this puzzle cuts through the confusion by distinguishing levels of knowledge with mathematical precision.¹

Mutual knowledge means everyone knows something. Common knowledge means everyone knows it, and everyone knows everyone knows it, and everyone knows everyone knows everyone knows it, all the way up. The visitor's announcement doesn't create mutual knowledge of blue eyes — that already existed. It creates common knowledge of blue eyes, and this turns out to be an entirely different beast.

To see why, work the small cases. With one blue-eyed person (say, Alice): she can't see any blue eyes, so before the announcement she doesn't know her own color. The visitor says "someone has blue eyes," Alice sees no one else with blue eyes, deduces hers are blue, and leaves night 1.

With two (Alice and Bob): Alice sees Bob's blue eyes, knows there's at least one. Bob sees Alice's, same. Neither can deduce their own color. But after the announcement, Alice reasons: "If my eyes aren't blue, Bob sees no blue eyes, so he'd know his are blue and leave night 1." Bob doesn't leave night 1. So Alice's eyes must be blue. Bob runs the same argument. Both leave night 2.

The magic is in what happens without the announcement. With two people, everything seems the same — both know there's someone with blue eyes. But Alice doesn't know whether Bob knows there's someone with blue eyes. Maybe Alice's eyes are green, in which case Bob sees no blue eyes at all. So Alice can't run the "if Bob knew, he'd leave" argument, because she can't be sure Bob knows. The announcement closes this gap. It makes "someone has blue eyes" not just known, but known-to-be-known.

With three people it's another layer deeper. Alice knows Bob knows someone has blue eyes (Alice can see that Bob can see Carol's blue eyes). But Alice doesn't know that Bob knows that Carol knows someone has blue eyes. Maybe Alice's eyes are green, in which case Bob only sees Carol's blue eyes, and Bob can't be sure whether Carol sees any blue eyes at all. The visitor's announcement establishes knowledge all the way down to any depth.

The Temporal Logic

Tao formalizes this by introducing a temporal epistemic logic — knowledge evolves over time as people observe each other's actions (specifically, who doesn't leave on which night).¹

The key axiom: absence of action is information. When nobody leaves on night 1, every blue-eyed person learns something — specifically, that the scenario "there's exactly one blue-eyed person" has been ruled out. This updates everyone's knowledge simultaneously. When nobody leaves on night 2, "exactly two" is ruled out. The whole argument is a cascade of increasingly complex inferences, each triggered by the observed inaction of the previous night.

What makes this work formally is the "announcement" operator. The visitor's statement is a public announcement — it simultaneously updates every agent's knowledge and every agent's model of every other agent's knowledge. Without public announcement, agents can have different models of what other agents know, and these models can be wrong in ways that prevent the inductive argument from getting started.

This is the same mechanism behind the Emperor's New Clothes. Every courtier knows the emperor is naked. Every courtier knows every other courtier knows. But nobody knows that everyone knows everyone knows — perhaps some courtiers have bad eyesight, or perhaps some genuinely believe in the cloth. The child's announcement "he's naked!" creates common knowledge, and the social fiction collapses.

Where Epistemic Logic Meets the Real World

This connects to Common Knowledge in an important way: many coordination problems are really problems of establishing common knowledge, not of acquiring knowledge. Financial panics, political revolutions, and social norm changes all involve situations where everyone individually knows something but can't coordinate because they lack common knowledge of that shared understanding.

I think the blue-eyed islanders puzzle reveals something deep about the nature of social intelligence. The reasoning each islander must perform isn't hard — it's just induction. What's hard is the depth of recursive modeling required: I need to model what you know about what she knows about what I know. Humans are notoriously bad at this beyond about three or four levels of nesting, which may be why common knowledge is so hard to achieve in practice and why public announcements (from media, leaders, or just loud children) are so socially powerful. They collapse an arbitrarily deep tower of "I know you know I know" into a single shared fact.

The formalism also has implications for AI Alignment. An AI that can reason about what humans know about what the AI knows about human values is operating in exactly this kind of epistemic logic. The alignment problem isn't just "does the AI know our values?" but "does the AI know that we know that it knows our values, and so on?" — a question whose difficulty increases with each level of nesting, just like the blue-eyed islanders.