The explanation is actionable only when the named input is one the user can change

What it is

“If you were ten years younger, this one would pass.” On the model the sentence can be true: age is the nearest flip. The applicant cannot change age. A counterfactual is actionable only when the input it names is one the user can change. An unchangeable “if” is a true empty sentence.

Actionability is not truth. True and immovable still cannot become a next step.

Why it happens

Algorithms that generate counterfactuals usually minimise distance, and distance does not care whether a variable can move in a life. Age, ethnicity proxies, a default that has already happened, are good directions at short distance and dead directions in life. Faced with a dead direction, the user either gives up acting or goes to change what should not be changed (faking age). The account is finished in mathematics and has given no act in the procedure.

Changeable also splits into “legally can change” and “in time for this round.” A degree can change and still miss this window, so it is still unactionable for this instance. Actionability has to be pinned to the current flow, not to the scale of a life.

Studying it

The same model generates two sets of counterfactuals: nearest flip with no mutability limit, and flip only on mutable fields. Watch what people change next, whether the change goes through, and whether they touch immutable items. Independent variables: how the mutable set is defined, whether “this item cannot be changed” is marked. Dependent variables: share of submitted edits that are mutable, attempts on immutable items, abandonment.

The mutable set must be given by the domain; the algorithm must not guess it. A wrong guess and the whole experiment is measuring the wrong actionability.

Where it stops holding

Some users can change items others cannot (a company applicant can change a registered address, a person cannot). Actionability varies with role; define the mutable set by role. If the user only wants to understand “what the system is sensitive to,” an immutable counterfactual still has epistemic value, but that is not action, and the interface must not label it “you can do this.” Which of several counterfactuals to pick is the next choice problem.

Applying it

• Add a mutability constraint at generation: look for a flip only on fields the user can change this time. If none, say “there is no change you can make now”; do not fall back to age.
• Mark each listed item “can be supplemented this round / cannot be changed.” If an immutable item is still shown, it can only be a limit, not a suggestion.
• Refuse to put historical identity variables into a “if you…” suggestion sentence.
• Check: of people who revise as the copy says, are the fields all ones this flow accepts. If age, sex, or a default that has already happened appear, actionability is lying.

Related

• Same group: L5.08.1 A counterfactual explanation says which input changes would change the outcome · 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: L5.05 The Moderation Principle of Transparency · L5.02 Local and Global Explanations · L5.01 Types of Explainability
• Search terms: actionable counterfactual · mutable features · recourse