Explore the Atlas

Alignment can come apart in surprising ways.

Seeing one form of alignment does not tell you that another is present. Sometimes two forms are merely independent. Sometimes improving one can actually make another worse.

These cases are useful because they show why it is dangerous to treat alignment as a single property.

Goals → behavior

Shared goals do not guarantee coordinated action. →

People can genuinely want the same outcome while failing to translate that agreement into compatible plans or behavior.

Example: strong intentions often fail without concrete implementation procedures.

Agreement → concordance

People can coordinate without agreeing. →

Compatible trajectories do not require common values, goals, beliefs, or preferences.

Example: buyers and sellers can want opposite outcomes and still coordinate through a market.

Values → Relevance

Good values do not guarantee seeing what matters. →

A system can care about the right outcome and still fail because it does not notice the relevant difference, recognize the situation, or perceive an available response.

Example: a driver can genuinely care about pedestrian safety yet fail to notice that someone is about to enter the road.

Alignability → Values

A system can be steerable without sharing your values. →

Changing its incentives, constraints, or environment can redirect behavior while its underlying objectives remain the same.

Example: firms change their behavior in response to prices while remaining profit maximizers

Performance → Robustness

Looking aligned now does not mean staying aligned under change. →

Two systems can behave identically in familiar conditions and diverge when they are perturbed.

Example: Two drones can hover equally well in still air. A sudden gust reveals that one can detect the deviation and recover while the other drifts away or crashes.

The point is not the list.

Each example is a way of asking a different question about an apparently aligned system. The Atlas collects these distinctions so that one kind of success is not silently used as evidence for another.