EU non-discrimination law lets a claimant win by establishing only a prima facie case of discrimination, whereafter the burden of proof shifts to the respondent. Machine-learning decision systems (notably in recruitment and employment) destabilize this enforcement mechanism: the statistical complexity and opacity of the systems prevent claimants from showing the causal link the equality directives require, while respondents facing the shifted burden can invoke causes hidden inside the model. This record poses the open problem: what exactly a claimant must prove to establish a prima facie case of direct and indirect algorithmic discrimination (including proxy discrimination), how the burden of proof should be allocated once shifted, and through which legal gateways the evidence needed to meet these thresholds can be obtained. Despite a decade of scholarship, the 2024 AI Act, the GDPR explanation rights, and a growing body of case law, no CJEU ruling, EU legislative act, or settled doctrine answers these questions as of 2026; every inspected treatment proposes rather than resolves.

Background

The EU equality directives — Directive 2000/43/EC (racial and ethnic origin), Directive 2000/78/EC (religion, disability, age, sexual orientation in employment), Directive 2004/113/EC (goods and services) and Directive 2006/54/EC (employment and occupation) — prohibit direct discrimination (less favourable treatment on a protected ground) and indirect discrimination (a neutral practice that particularly disadvantages a protected group without objective justification). Their central enforcement device is a procedural rule: the claimant need only establish a prima facie case — facts from which a court may presume discrimination — whereafter the burden of proof shifts to the respondent, who must prove there was no breach of the equal-treatment principle (Article 8 of Directive 2000/43/EC and the corresponding provisions of the other directives, as interpreted by the Court of Justice of the European Union, CJEU). The CJEU has never precisely fixed the content of the prima facie requirement; in opaque-decisionmaker contexts such as pay discrimination it developed evidential relaxations, e.g. the Danfoss presumption, and the EU legislature has since gone further in the 2023 Pay Transparency Directive (EU) 2023/970, whose Article 18(2) places on the employer the burden of proving that there has been no direct or indirect pay discrimination where the employer has not implemented the pay-transparency obligations of Articles 5, 6, 7, 9 and 10 of that Directive. Automated decision-making (ADM) systems trained by machine learning strain this framework in a new way. Their inner workings are often incomprehensible even to their deployers ('algorithmic opacity' or the 'black box' problem), so claimants can neither identify what caused their treatment nor obtain comparative or statistical evidence; in direct discrimination this blocks the prima facie case, while in indirect discrimination causation resurfaces after the burden shift as a quasi-defence of objectively justified, ground-unrelated factors. A recurring hard case is proxy discrimination: a model using a facially neutral feature (e.g. postcode) that is inextricably linked to a protected ground (e.g. ethnicity). The CJEU's 'inextricable link' doctrine, developed in CHEZ Razpredelenie Bulgaria (C-83/14) for human conduct, might classify such proxies as direct discrimination, but its transfer to statistical, partially unintelligible models is contested, because direct discrimination admits no objective justification and therefore triggers strict liability. Candidate evidentiary gateways exist outside the equality directives: the General Data Protection Regulation (GDPR) grants access rights (Article 15(1)(h)) and constrains fully automated decisions (Article 22), as clarified in the CJEU's SCHUFA and Dun & Bradstreet Austria rulings; the Artificial Intelligence Act (Regulation (EU) 2024/1689) adds transparency obligations for high-risk systems and, in Article 86, a right to an explanation of individual decisions; the proposed AI Liability Directive — one remedial avenue suggested in the sources — was withdrawn by the Commission in 2025. What remains unresolved is how these pieces combine into a workable law of proof: what suffices for a prima facie case of algorithmic discrimination, who must prove what after the shift, and how claimants reach the evidence at all.

Problem Statement

The problem is to make the EU non-discrimination prohibition justiciable for machine-learning decision systems by fixing the law of proof that the equality directives presuppose but do not supply. It comprises three linked questions that share this single objective and are to be answered together, on the basis of EU primary and secondary law and the CJEU's jurisprudence: (1) Prima facie case — what facts and evidence must a claimant establish for a court to presume direct or indirect algorithmic discrimination; in particular, does a facially neutral input that is inextricably linked to a protected ground (proxy discrimination) bring the claim within direct discrimination, and is a positive indication of a causal relationship between the protected ground and the treatment required of the claimant? (2) Allocation after the shift — once the burden of proof shifts to the respondent under Articles 8–10 and 19 of the equality directives, what must the respondent demonstrate to rebut the presumption or to objectively justify a neutral practice, and can it discharge the burden by showing that the bias was autonomously developed by the algorithm rather than intended or knowable? Whether such a showing would operate as rebuttal of the presumption, as objective justification of a neutral practice, or as neither is not presupposed here and is itself part of question (2), given that direct discrimination admits no objective justification. (3) Access to evidence — through which legal gateway or combination of gateways can claimants actually obtain the model, data, and explanation evidence needed to meet the thresholds in (1) and (2), in particular the equality directives' procedural rules read jointly with the GDPR's access and explanation rights (Articles 15(1)(h) and 22) and the AI Act's transparency and Article 86 explanation obligations, without excluding other routes grounded in EU primary and secondary law, such as the right to an effective remedy under Article 47 of the Charter of Fundamental Rights or the information-gathering and litigation powers of equality bodies? The operative rules are sought for machine-learning decision systems generally, within the protected grounds and material scopes of the four equality directives, with recruitment and employment as the salient and most-developed application context rather than a scope limit; because Article 86 AIA reaches only high-risk systems and disclosure under the GDPR and the AI Act remains conditional, the protected grounds, material fields, and system classes (e.g. AI Act Annex III high-risk systems versus other systems) for which a proposed rule claims validity form part of that rule. The answer sought is the operative rule for each of the three questions and a coherent account of how the three rules interact — in particular how structural transparency obligations and evidence gateways enable the prima facie threshold and relate to the shift of the burden of proof — framed in a form the CJEU or the EU legislature could adopt.

References

  1. Ljupcho Grozdanovski (2021). 'In search of effectiveness and fairness in proving algorithmic discrimination in EU law.' Common Market Law Review 58(1): 99–136. DOI 10.54648/cola2021005. https://doi.org/10.54648/cola2021005
  2. Christine Carter (2024). 'Why the algorithmic recruiter discriminates: The causal challenges of data-driven discrimination.' Maastricht Journal of European and Comparative Law 31(3): 333–359. DOI 10.1177/1023263X241248474. https://doi.org/10.1177/1023263X241248474
  3. Henni Parviainen (2024). 'Challenges of Direct Discrimination in Algorithmic Recruitment: Insuperable or Not?' International Journal of Comparative Labour Law and Industrial Relations 40(4): 437–466. DOI 10.54648/ijcl2024017. https://doi.org/10.54648/ijcl2024017
  4. Anna W. Ghavanini (2026). 'The burden of proof in EU anti-discrimination law: A civil procedure analysis.' International Journal of Discrimination and the Law (online 31 March 2026). DOI 10.1177/13582291261439759. https://doi.org/10.1177/13582291261439759
  5. Sandra Wachter, Brent Mittelstadt, Chris Russell (2021). 'Why fairness cannot be automated: Bridging the gap between EU non-discrimination law and AI.' Computer Law & Security Review 41: 105567. DOI 10.1016/j.clsr.2021.105567. https://doi.org/10.1016/j.clsr.2021.105567
  6. Aislinn Kelly-Lyth (2023). 'Algorithmic discrimination at work.' European Labour Law Journal 14(2): 152–171. DOI 10.1177/20319525231167300. https://doi.org/10.1177/20319525231167300
  7. Lucas Michael Haitsma (2023). 'Regulating algorithmic discrimination through adjudication: The Court of Justice of the European Union on discrimination in algorithmic profiling based on PNR data.' Frontiers in Political Science 5: 1232601. DOI 10.3389/fpos.2023.1232601. https://doi.org/10.3389/fpos.2023.1232601
  8. Anna Capellà Ricart (2024). 'The role of European equality bodies to address algorithmic discrimination.' International Journal of Discrimination and the Law 24(3): 194–216. DOI 10.1177/13582291241257201. https://doi.org/10.1177/13582291241257201
  9. Ljupcho Grozdanovski (2025). 'Non-discrimination law, the GDPR, the AI act and the – now withdrawn – AI liability directive proposal offering gateways to pre-trial knowledge of algorithmic discrimination.' AI and Ethics 5(5): 5039–5062 (online 29 May 2025). DOI 10.1007/s43681-025-00754-0. https://doi.org/10.1007/s43681-025-00754-0
  10. Vibeke Blaker Strand (2025). 'The Burden of Proof Rule Under EU Non-discrimination Law and Algorithmic Discrimination: Addressing Challenges and Possibilities.' Ljubljana Law Review (Zbornik znanstvenih razprav) 85(1): 165–182 (online 30 December 2025). DOI 10.51940/2025.1.165-182. https://doi.org/10.51940/2025.1.165-182
  11. Konstantinos Lamprinoudis (2026). 'The courts have spoken: examining judicial responses across the EU to algorithmic bias in automated decision-making through the lens of non-discrimination law.' European Law Open (Cambridge, online 27 March 2026): 1–27. DOI 10.1017/elo.2026.10073. https://doi.org/10.1017/elo.2026.10073
  12. Janneke Gerards, Frederik Zuiderveen Borgesius (2025). 'Protected Grounds and the System of Non-Discrimination Law in the Context of Algorithmic Decision-Making and Artificial Intelligence.' arXiv:2509.08837. https://arxiv.org/abs/2509.08837
  13. Birte Böök, Linda Senden (2025). 'The puzzle of algorithmic discrimination in European equality law: Putting together some practical, legal and conceptual pieces.' In: Research Handbook on European Anti-Discrimination Law, Edward Elgar, 482–509. DOI 10.4337/9781789906318.00035. https://doi.org/10.4337/9781789906318.00035