Intelligence Is Applied Causality

Every few months another AI benchmark falls, and the same argument restarts: is this thing intelligent, or is it a trick? The argument never resolves because the people having it rarely agree on what intelligence is. So here is a working definition, one I now use to evaluate people, plans, and machines alike. Intelligence is deciding and acting based on a profound understanding of cause and effect relationships.

Notice what the definition leaves out. It says nothing about how much you know or how fluent you sound. Knowledge and fluency are assets, but both can be high while intelligence is low. A system can memorize everything that ever happened in a market and still destroy value the week conditions change, because it learned what happened without learning why. The same holds for an executive who can recite every metric in the deck but cannot say which lever moves which number.

Researchers have circled this for decades. Shane Legg and Marcus Hutter once collected more than seventy published definitions of intelligence and found one recurring core: the ability to achieve goals across a wide range of environments. That framing is right but incomplete, because it names the outcome without naming the ingredient that produces it. You cannot reliably achieve goals in an environment you can only describe. You have to know which of its levers move which of its outcomes.

Intelligence is deciding and acting based on a profound understanding of cause and effect relationships. Everything else is supporting equipment.

Prediction is not understanding

The rooster crows and the sun rises. The crowing predicts the sunrise with near perfect accuracy, and it causes nothing. Judea Pearl built a career on this distinction. In The Book of Why he argues that no amount of passive data, however large, can by itself answer causal questions, because data records what happened, never what would have happened under a different choice.

This matters right now because large language models are the strongest association engines ever built. They are trained to predict the next token from past patterns, and they do it so well that the output feels like understanding. Sometimes real causal structure comes along for the ride, because the training text contains humans explaining causes. Often it does not. François Chollet made a related point about measurement: a test score shows skill at a task, and skill can be bought with data and practice. Intelligence is the thing that acquires new skill when the task changes. Both arguments land in the same place. The scoreboard tells you what a system did. It does not tell you whether the system knows why.

The ladder from seeing to imagining

Pearl gives the definition its structure with his ladder of causation: three rungs that different kinds of reasoning occupy. The bottom rung is association, seeing patterns in data. The middle rung is intervention, predicting what happens when you change something on purpose. The top rung is the counterfactual, reasoning about what would have happened if you had chosen differently.

IMAGININGCounterfactuals: what would have happened if we had acted differently?

DOINGIntervention: what happens to the system when we change this input?

SEEINGAssociation: what tends to occur together in the data we already have?

Most corporate analytics lives on the bottom rung. A dashboard is rung one. An A/B test is rung two: you changed one thing deliberately and watched the result. A serious post mortem is rung three: the team asks what would have happened had they shipped a week later or priced ten percent lower. The word "profound" in the definition is doing real work here. Profound understanding means you can operate on all three rungs: read the pattern, predict the intervention, and imagine the alternative.

Why deciding and acting belong in the definition

Understanding alone is scholarship. It becomes intelligence when it drives decisions, and the decisions become actions in the world. Acting earns its place in the definition for two reasons.

First, action is the only honest test of a causal model. Explanations are cheap; anyone can tell a plausible causal story after the fact. An intervention puts the story at risk. If you believe a long checkout flow causes cart abandonment, shortening it is both a decision and an experiment, and the result will tell you whether the belief was knowledge or narrative.

Second, action is where causal knowledge comes from in the first place. Science did not climb Pearl's ladder by watching more carefully. It climbed by intervening: controlled experiments and deliberate changes with measured outcomes. The loop is the same whether the actor is a scientist, a product team, or an autonomous agent.

MODELForm a causal picture of how the system works

→

DECIDEChoose the intervention the model says will work

→

ACTRun the intervention in the real world

→

OBSERVECompare the outcome against the prediction, then revise the model

A team that models but never acts learns nothing new. A team that acts without a model is gambling. Intelligence is this loop running: the model earns the decision, the action tests the model, and the observation revises it.

What this means for AI systems

Bernhard Schölkopf and his coauthors state the machine learning version plainly: most current systems learn statistical dependencies rather than causal structure, which is why they degrade when the world drifts away from the training distribution. A model of correlations is a photograph of one world. A causal model is a machine for generating predictions about many possible worlds.

Agentic systems change the terms of this problem. An agent with tools can intervene: run the code, execute the test, query the live system, and observe what actually happens. That moves it from rung one to rung two, and it is why a coding agent wrapped in a good test harness behaves so much more intelligently than the same model producing code it never runs. The feedback loop supplies the causal grounding that training data alone could not.

The definition also gives you a sharper way to evaluate AI inside your own organization. Ask which rung the system stands on rather than how impressive its output sounds. Does it summarize patterns? Does it predict interventions and then verify them? Can it explain what would have happened on the path not taken? The same three questions work on vendors, on internal tools, and, uncomfortably, on your own leadership team.

Definitions are tools, and a good one changes what you look for. Carry this one around for a week and you notice how much activity that passes for intelligence is rung one work: summarizing and describing what already happened. You also start to notice the quieter people and systems doing the actual thing, running the loop from model to decision to action to revision.

What leaders should do

A definition matters when it changes behavior. This one converts into four habits.

1. Require every significant proposal to state its causal claim in one sentence: we believe X causes Y, so doing Z should produce W. If nobody in the room can write that sentence, you have a pattern, not a plan.
2. Shift budget from passive analysis toward cheap interventions. One controlled pilot that changes something teaches more about your business than another quarter of dashboards.
3. Evaluate AI systems on rung two, not rung one. Before scaling a pilot, test its recommendations against a scenario the training data never saw and check whether the reasoning survives.
4. Make counterfactual review a standing ritual. After every major win or loss, write down what the team believes would have happened under the alternative decision, and treat the update to the causal model as the real output of the review.