A Leak That Fixes Itself

Opening Scene

A plumber notices that a certain fixture’s low water pressure always coincides with a nearby pump running at high speed, and initially assumes the pump is causing the pressure drop. Further investigation reveals the opposite: the pump is a compensating booster, automatically running harder specifically because it detects low pressure downstream. The plumber had the direction of causation exactly backward — not because the correlation was wrong, but because the arrow was pointing the other way.

In Plain English

Reverse causation occurs when a variable assumed to be the cause is actually the effect, with the true causal direction running opposite to what seemed intuitive. This is a genuinely common and easy-to-miss mistake, because a correlation alone can’t reveal which direction causation runs — “A and B are correlated” is equally consistent with “A causes B,” “B causes A,” or the confounding covered in Article 4.

The Old Way

Before this had formal causal inference language, mistaking the direction of causation produced confidently wrong conclusions across many fields:

Early observers noted that sick people often visited doctors, and might have naively concluded doctors cause sickness — an extreme but illustrative example of exactly this trap.
A business might notice that struggling stores get more manager attention, and mistakenly conclude manager attention causes stores to struggle, when the true direction runs the other way: struggling stores prompt more attention.
A researcher might notice that happier employees have higher-performing teams, and assume happiness causes performance, when in reality strong performance might also be boosting employee happiness.

In each case, the correlation was real, but the assumed direction of causation was backward.

What’s Changing (and Why AI Is the Reason)

Formal causal graph methods, covered in Article 5, force an explicit statement of assumed causal direction, making it easier to spot and challenge a potentially reversed arrow before it goes unquestioned.
Time-ordered data — knowing which variable’s change genuinely preceded the other’s — has become an increasingly available and increasingly used tool for ruling out at least some reverse causation possibilities, though temporal precedence alone doesn’t fully settle the question.
As AI systems increasingly learn correlational patterns from data without any inherent understanding of causal direction, the risk of a model implicitly encoding a reversed causal relationship — and confidently acting on it — has become a genuine, practical concern.

The Metaphor, Fully Extended

Behind the Wall	Reverse Causation Concept
A pump assumed to cause a pressure drop	A variable assumed to be the cause
The pump actually responding to and compensating for the pressure drop	The variable actually being the effect, not the cause
A correlation that’s real, but pointing the wrong direction in the plumber’s mind	A correlation that’s real, but with the causal direction backward
Recognizing the booster pump’s true, reactive role in the system	Recognizing a variable’s true role as effect rather than cause

For Beginners: What to Actually Do

Practice explicitly considering the reverse-direction explanation for any causal claim you encounter, alongside the confounding possibility from Article 4.
Learn to ask “could the assumed effect actually be causing the assumed cause instead?” as a standard check.
Use time-ordering where available as one useful, though not fully conclusive, way to help rule out reverse causation.

For Practitioners and Leaders: The Deeper Layer

Require an explicit justification for assumed causal direction in any analysis presented for a major decision, not just an assumed intuitive arrow.
Watch specifically for reverse causation in any “attention causes problems” or “intervention causes struggle” type finding, where the reverse explanation is often more plausible than it first appears.
Recognize that machine learning models trained on purely correlational data have no inherent way to determine causal direction, making this a genuine, structural risk worth actively guarding against.

Quick Recap

Reverse causation occurs when the assumed cause is actually the effect, with the true causal direction running the opposite way.
A correlation alone can never reveal which direction causation runs.
Formal causal graphs and time-ordered data both help identify and challenge a potentially reversed causal arrow.
AI systems trained on correlational data have no inherent way to determine causal direction, making this a structural, practical risk.

Where This Fits in the Series

Article 13 covered getting the direction of flow backward. Article 14 covers tracing the full path an effect takes through a system, not just its final destination.