Counterfactual Explanations Without Opening the Black Box: Automated Decisions and the GDPR

Sandra Wachter and Brent Mittelstadt were at the Oxford Internet Institute and the Alan Turing Institute when this was written, with Mittelstadt also at UCL; Chris Russell, the machine-learning researcher of the three and the source of the technical counterfactual method, was at the University of Surrey. Wachter and Mittelstadt are leading scholars of data-protection law and the ethics of automated decision-making, and the work was funded through the Alan Turing Institute under a public research grant, a provenance the authors state in the paper.

It was published in the Harvard Journal of Law & Technology, volume 31, in spring 2018, after a 2017 SSRN preprint; as a US law review it is student-edited rather than blind-refereed, the standard filter for legal scholarship. The paper's setting is the debate over whether Europe's General Data Protection Regulation gives individuals a "right to explanation" of automated decisions — a question the same authors had addressed in a companion 2017 article, Why a Right to Explanation of Automated Decision-Making Does Not Exist in the General Data Protection Regulation. Here they propose a constructive remedy: the counterfactual explanation, which states the smallest change to a person's data that would have produced a different outcome.

The piece sits as a secondary, legal-technical argument bridging machine-learning methods and data-protection law. Its framing — accounts of a decision "without opening the black box" — is the legal analogue of the post-hoc explanation that LIME provides, and it stands against Rudin's position that high-stakes decisions should be made by models transparent enough that no such workaround is needed. It supplies the concrete legal mechanism that Pasquale's broader case for transparency leaves open, narrowing "a right to understand" to "a right to know what to change."

The authors' stake is academic and reform-oriented. The argument takes a definite legal-technical position — that the GDPR lacks a binding right to explanation and that counterfactuals are the more workable remedy — which became influential in subsequent European regulatory debate, so the return is reputational and policy-shaping. The funding is public rather than commercial, and the paper reads as scholarship arguing for a remedy rather than as advocacy for any firm.