Explainable AI is the dominant response to automated decision-making, on the assumption that a system able to show how it produced an output has met the conditions for accountability. That assumption fails in a specific way. Selbst and Barocas established in 2018 that explaining a model does not show its decisions to be justified; the diagnosis has not since been converted into an enforceable mechanism. This article attempts that conversion. It argues, first, that normative and computational opacity are independent, so that interpretability progress makes embedded value commitments discoverable while leaving unanswered who adopted them. Second, that these commitments are located at conversion points, meaning any place where a score, rank or verdict becomes a differential consequence, each of which sets an exchange rate between two harms borne by different parties. Third, that disclosure must separate a discovery obligation running against the technical function from an adoption act running against institutional authority, since merging them produces a signature on commitments the signatory cannot see. Fourth, that the resulting Declared Operating Position becomes enforceable through a presumption triggered by an adverse decision and prima facie differential effect. Comparative analysis across European, United Kingdom, United States and Indian law shows this structure already exists in each. Indian administrative law supplies the presumption directly through the adverse inference drawn from unreasoned orders. The instrument also supplies the material that the less discriminatory alternative limb of disparate impact doctrine requires and that claimants presently cannot obtain.
Karthik Ramakrishna Suresh· Zenodo (CERN European Organi...· 0 citations
Explainable AI is the dominant response to automated decision-making, on the assumption that a system able to show how it produced an output has met the conditions for accountability. That assumption fails in a specific way. Selbst and Barocas established in 2018 that explaining a model does not show its decisions to be justified; the diagnosis has not since been converted into an enforceable mechanism. This article attempts that conversion. It argues, first, that normative and computational opacity are independent, so that interpretability progress makes embedded value commitments discoverable while leaving unanswered who adopted them. Second, that these commitments are located at conversion points, meaning any place where a score, rank or verdict becomes a differential consequence, each of which sets an exchange rate between two harms borne by different parties. Third, that disclosure must separate a discovery obligation running against the technical function from an adoption act running against institutional authority, since merging them produces a signature on commitments the signatory cannot see. Fourth, that the resulting Declared Operating Position becomes enforceable through a presumption triggered by an adverse decision and prima facie differential effect. Comparative analysis across European, United Kingdom, United States and Indian law shows this structure already exists in each. Indian administrative law supplies the presumption directly through the adverse inference drawn from unreasoned orders. The instrument also supplies the material that the less discriminatory alternative limb of disparate impact doctrine requires and that claimants presently cannot obtain.
Karthik Ramakrishna Suresh· Zenodo (CERN European Organi...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
