AI wrote that a campaign caused the increase. Does your report show more than “after”?

Our view: keep the observed change separate from its proposed cause. A fluent explanation cannot supply a comparison the data does not contain.

NIST distinguishes association from causation. HM Treasury’s policy-evaluation guidance likewise separates observing an outcome from attributing it to an intervention. These are background evaluation principles, not a new finding about every AI report.

A true change can carry an unsupported cause

In an invented shop report, orders rise from 10 to 20 after a new promotion. A holiday starts in the same period. The orders doubled, but those two counts do not isolate the promotion’s effect. The holiday is a possible alternative explanation, not a cause we proved.

Change the claim before polishing it

Our proposed wording is: “Orders rose after the promotion; this comparison does not establish how much of the increase it caused.” Do not let a more confident rewrite turn timing into attribution. It is also unwarranted to conclude that the promotion had no effect.

Ask which comparison supports the explanation

For an AI-written report, our editorial questions are what was measured, what else changed, and what evidence describes the outcome without the intervention. A comparable control or another justified evaluation design needs actual data and assumptions; asking AI for a causal story does not create them.

Leave an unresolved cause unresolved

When the available report only supports a descriptive statement, publish that narrower statement. Save proposed explanations as hypotheses and identify what evidence would help assess them. Choosing and evaluating a causal design takes more than this short column.

Published October 14, 2026. Editorial analysis using fictional shop counts and existing evaluation guidance. The HM Treasury HTML page is marked updated May 15, 2026; this is not October news. No campaign, causal experiment, business dataset or AI output was tested.

Sources: NIST · Correlation and causality · HM Treasury · Impact evaluation.