A counterfactual explanation says which input changes would change the outcome

What it is

Beside a loan denial: “if monthly income were four thousand higher, this one would pass.” The sentence does not say how many layers the model has, and does not rank features. It gives an input world that did not happen, and the output in that world. A counterfactual explanation says which input changes would change the outcome.

Wachter et al. offered it as a form of account for algorithmic decisions: the nearest change that would flip the result. It answers “how would it be different,” not “what temperament the system has overall,” and not “which layers the process walked.”

Why it happens

A decision boundary parts two classes of output in input space. The shortest, or most tellable, steps from the current point into the other class are one counterfactual. What the user gets is a path, not a device manual. The path’s use is to make dependence of output on input visible: some directions flip at a touch, some never leave the original side however you walk.

This is not a local surrogate. A local surrogate fits “why this time” in a neighbourhood; it need not point to a change that would flip. A counterfactual can skip “because” entirely and only say “if.” Do not weld the two questions into one sentence.

Studying it

Pair the same denial with a counterfactual (an input change that would flip), a feature ranking, or no account. Watch whether people can point to a change that really flips, and whether they mistake it for a causal statement. Independent variables: type of account, whether the counterfactual was generated from the true model. Dependent variables: whether the pointed-to change actually flips on the model, rate of reading “if” as “because.”

The counterfactual must be verified on the original model. An “if” generated by another surrogate is not measuring counterfactual explanation; it is another after-the-fact story.

Where it stops holding

When output is insensitive to input (creation, random sampling), there is no stable “change these and it changes,” and this entry’s object vanishes. If the user cannot change inputs at all (read-only results), the counterfactual has no use. Actionability, choosing among many counterfactuals, being used to game, and whether other conditions held constant, are later entries. This one only defines what this kind of account is pointing at.

Applying it

• Title it “if these items became like this, this result would become…,” not “reasons” or “how the system works.”
• Show only changes that have been run on the current model and confirmed to flip. If they do not run, do not write “if.”
• Present it apart from rankings and heatmaps, so users do not read “if” and “because” as one sentence.
• Check: have people revise an input as the copy says and resubmit. If the result does not change, this is not a counterfactual; if they restate it as “denied because income is low,” the type has already been misread.

Related

• Same group: L5.08.2 The explanation is actionable only when the named input is one the user can change · L5.08.3 The same outcome has many counterfactuals; choosing which to show is itself a value judgement · L5.08.4 Counterfactual explanations leak decision rules and may be used to game them · L5.08.5 A counterfactual assumes other conditions held constant, which rarely holds in practice
• Nearby: L5.01 Types of Explainability · L5.02 Local and Global Explanations · L5.05 The Moderation Principle of Transparency
• Search terms: counterfactual explanation · Wachter · if-then explanation