Taking the Whys Seriously: Limitations of Counterfactual Explanations in Justification and Recourse

TEXT START: Counterfactual explanations (CEs) are widely used in explainable artificial intelligence (AI) to show how a model's outputs would change if the input features were manipulated.

The Dissection

This paper strips CEs of their supposed neutrality. It argues that the “why” is preloaded by measurement choices, labels, business requirements, validation, and success metrics—not merely generated by the final explanation method. Its real contribution is relocating responsibility from the visible model output to the organization that defined the model’s reality.

The Core Fallacy

It mistakes exposure of arbitrariness for recovery of agency. Showing that a counterfactual is an organizational artifact does not mean the affected person can force a different decision, or that recourse remains economically meaningful.

Under the Discontinuity Thesis, the deeper problem is not simply that CEs answer the wrong “why.” AI is converting people into objects of automated allocation while eroding the productive role that once gave them bargaining power. A more relational explanation can make the gatekeeping machinery more legitimate without making it less coercive.

Hidden Assumptions

• Institutions will remain answerable to the people they classify.
• A decision-maker can still be compelled to act differently.
• Transparency about upstream choices will create accountability rather than better compliance theater.
• Recommended changes will remain feasible for the claimant.
• Human justification and recourse will remain central after cognitive work is automated.
• Organizational choices can be reformed without confronting who owns and controls the automated system.

Social Function

Primary classification: partial truth serving as transition management.

The paper is a useful forensic instrument for locating buried discretion. But its reformist frame risks becoming ideological anesthetic: institutions may learn to disclose, justify, and optimize exclusion while leaving the underlying power structure untouched. In a mass-displacement regime, recourse can become a customer-service window on a machine whose economic answer is already fixed.

The Verdict

A sharp autopsy of counterfactual explanations, but an incomplete autopsy of the system deploying them. It exposes how “why” is manufactured upstream while avoiding the harsher conclusion: once AI severs productive participation, justification and recourse become governance features of managed dependency—not restored rights within a functioning labor order. This is not a rebuttal to the Discontinuity Thesis. It is a quality-control memo for the rationing machinery.