When the AI Got Causation Backwards: A 70%-Confidence False Conclusion

At 2:00 AM our anomaly detection flagged a campaign whose impressions had dropped 67% — 100 impressions versus a historical mean of 301 — and generated an anomaly event with a score of 67. The analyst on call clicked “reason causal chain,” and the LLM did its job. It analyzed 30 preceding candidate events and state snapshots, produced eight evidence links, and delivered a conclusion with 70% confidence: the 43 negative keywords added the day before had over-restricted the campaign's reach.

That conclusion was exactly backwards.

The truth: the negative keywords were the fix, not the bug

The 43 negative keywords — brand-adjacent terms like “webox”, “login”, “food”, and “catering” — were a deliberate, correct optimization. They had been cutting off competitor terms mis-matching our brand keywords, and their removal of that noise was why impressions fell. The AI had taken a correct optimization action and labeled it a mistake.

Five limitations found in the review

That night's human review — which happened the evening of July 15 — surfaced five structural limitations:

• Inverted causation. The model saw “more negatives + fewer impressions” and concluded over-restriction. It had no business semantics to tell it the negatives were our own brand terms.
• No business context. The LLM did not know these strings were our own brand names, nor that in B2B lead generation these are purely mis-matched queries no prospect would convert on.
• Inflated confidence. A 70% causal hypothesis was really a correlation. The evidence supported “impressions and negative count moved together,” not “negatives caused the drop.”
• Missing the real cause. A fully-loaded RSA push may have been the true driver, but it was too recent to sit inside the preceding-candidates window, so the model could not see it.
• Unfounded external guesses. Competitor spend and algorithm updates were floated without any data behind them.

What we changed

Three changes came out of this case. First, campaign-level business context is now injected into the prompt: business type, target audience, and the brand keyword list, so the model knows what a term means before it reasons about it. Second, humans can label each evidence link as agree or disagree, and that label is stored with the chain. Third, confidence is now adjusted dynamically by evidence strength — a correlation does not get to wear a causal confidence score.

The value of the closed loop

This case is why we believe in the machine-finds-signal, human-decides loop: the AI detected the anomaly, alerted, reasoned about it, and a human corrected it — in time, before the wrong conclusion could reach a client report. Four lessons we now apply to any AI causal reasoning:

1. Never trust a causal label without checking whether the evidence actually supports causation.
2. Give the model the business semantics — brand names, audience, offer — before asking it to reason.
3. Treat high-confidence causal claims built on correlational evidence as a system bug, not a quirk.
4. Always keep a human review step between an AI conclusion and any external action.

The machine found the signal; the human kept it honest. That division of labor, with evidence chains and a human review loop, is exactly what our attribution and growth engine at iport is built around.