The Alibi That Doesn't Hold

Opening Scene

The rejected loan applicant from Article 1 doesn’t just want to know which factors mattered. They want to know something more useful: what would have had to be different for the answer to change? That’s not a question feature importance or even SHAP directly answers. It’s a genuinely different kind of question — the same one a detective asks when testing an alibi: if this one detail had been different, would the whole story have held up?

That’s exactly what a counterfactual explanation provides.

In Plain English

A counterfactual explanation describes the smallest change to an input that would have flipped the model’s decision — “if your income had been $4,000 higher, this application would have been approved.” Rather than attributing credit to existing features, it identifies the nearest alternative version of the case that produces a different outcome, which is often exactly the actionable information a person actually wants.

The Old Way

Before counterfactual explanations were formalized, people reasoned this way instinctively, without any systematic method:

A student asking a teacher what they needed to do differently to get a better grade, receiving an informal, often vague answer.
An athlete asking a coach what specifically would have changed the outcome of a close game, based on the coach’s judgment rather than a rigorous method.
A rejected job candidate asking, informally, what they could improve for next time, with no guarantee the feedback given actually reflected the real decision boundary.

The instinct to ask “what would have had to change” long predates any formal technique for answering it precisely.

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

Counterfactual explanation algorithms can now search a model’s decision boundary systematically, finding the smallest, most realistic change that would flip an outcome, rather than relying on a human’s informal guess.
Regulatory frameworks, covered later in this series, increasingly favor counterfactual explanations specifically because they’re actionable — telling someone what to change is often more useful and more legally defensible than describing feature contributions alone.
Modern counterfactual methods increasingly constrain suggested changes to be realistic and achievable, avoiding the failure mode of technically correct but practically meaningless suggestions like “if you were 15 years younger.”

The Metaphor, Fully Extended

The Investigation	Counterfactual Concept
Testing whether a small detail change would break an alibi	Finding the smallest input change that flips a decision
The nearest alternative version of events that changes the verdict	The nearest alternative input that changes the outcome
A realistic, testable alternative scenario	A realistic, achievable counterfactual, not an arbitrary one
Giving someone an actionable path to a different outcome	A counterfactual explanation’s practical usefulness

For Beginners: What to Actually Do

Practice framing explanations as “what would need to change” alongside “what factors mattered” — they answer genuinely different, complementary questions.
When generating counterfactuals, check that suggested changes are realistic and achievable, not just mathematically valid.
Recognize counterfactual explanations as often the most useful format for the actual person affected by a decision, even when SHAP or LIME are more useful internally.

For Practitioners and Leaders: The Deeper Layer

Offer counterfactual explanations directly to end users affected by consequential decisions — they’re often more actionable and more satisfying than raw feature attributions.
Constrain counterfactual search to realistic, actionable changes as a matter of policy, not just a technical nicety.
Recognize regulatory frameworks increasingly expect this specific kind of explanation, a connection this series returns to in Article 19.

Quick Recap

Counterfactual explanations describe the smallest realistic input change that would flip a model’s decision.
They answer a genuinely different, often more actionable question than feature attribution methods like SHAP or LIME.
Modern methods constrain suggestions to be realistic and achievable, not just mathematically valid.
Regulators increasingly favor counterfactual explanations for their practical usefulness to affected individuals.

Where This Fits in the Series

Article 7 covered the forward-looking “what would have changed this” question. Article 8 steps back to look at the difference between explaining a model everywhere versus explaining just one case.