Epistemic injustice in algorithmic decision-making in social protection systems: operationalising fricker's testimonial and hermeneutical injustice for automated case- management

The article addresses a precise question that arises at the intersection of Miranda Fricker's theory of epistemic injustice and the ongoing automation of social-protection case- management systems across Europe, the United Kingdom, and Australia: does the user lose her epistemic standing when she is categorised by an algorithm whose interpretive horizon she does not share, and whose categorical apparatus she cannot inspect? Drawing on Fricker's foundational distinction between testimonial injustice (a credibility deficit assigned to a speaker on the basis of identity-prejudice) and hermeneutical injustice (a structural deficit in shared interpretive resources that disadvantages a marginal group in articulating its experience), the article develops a typology of five modes of algorithmic epistemic closure through which automated case-management systems systematically reproduce both forms of injustice. The five modes are: (1) categorical translation, by which the user's narrative is compressed into pre-defined system fields that admit only of the kinds of testimony the algorithm can recognise; (2) calibration silencing, by which an algorithm trained on biased historical data systematically discounts certain classes of testimony as statistically less credible; (3) opacity, by which the black-boxed character of the model deprives the user of the interpretive resources required to make sense of the decision; (4) temporal compression, by which the speed of automated processing forecloses the space for nuance, contextualisation, and follow-up that human caseworkers would otherwise provide; and (5) appeal asymmetry, by which formal redress procedures presuppose a level of algorithmic literacy and infrastructural access that the affected user characteristically lacks. The typology is applied empirically to three documented cases — the Dutch SyRI welfare-fraud detection system, the United Kingdom's 2020 Ofqual A-level grading algorithm, and the Australian Robodebt automated debt-recovery scheme — drawing on the peer-reviewed analyses of these cases by van Bekkum and Borgesius (2021), Smith (2020), and Nikidehaghani, Andrew and Cortese (2023). The article argues that the five modes do not act in isolation; they form an integrated apparatus whose cumulative effect is the conversion of a person seeking social support into what I term an “algorithmically illegible subject” — a position from which neither testimony nor self-interpretation can be successfully advanced within the system as it stands.