In Search for Effectiveness and Fairness in Proving Algorithmic Discrimination in EU law

Grozdanovski, Ljupcho

Request a copy

Article (Scientific journals)

In Search for Effectiveness and Fairness in Proving Algorithmic Discrimination in EU law

Grozdanovski, Ljupcho

2021 • In Common Market Law Review, 58 (1), p. 99-136

Peer reviewed

Permalink
https://hdl.handle.net/2268/306510

Files (1)Send to Details Statistics Bibliography Similar publications

Files

Full Text

LG_EvidenceAlgorithmicBias_CMLRev.pdf

Author postprint (329.9 kB)

Request a copy

All documents in ORBi are protected by a user license.

Send to

RIS BibTex APA Chicago Permalink X Linkedin

Details

Abstract :

[en] Examples of discriminatory algorithmic recruitment of workers have triggered a debate on application of the non-discrimination principle in the EU. Algorithms challenge two principles in the system of evidence in EU non-discrimination law. The first is effectiveness, given that due to algorithmic opacity, the parties in algorithmic discrimination cases do not have easy and unrestricted access to facts enabling them to support their claims. The second is fairness, insofar as the launching and unfolding of the evidentiary debate requires lifting the veil of algorithmic opacity: a colossal task, placing unrealistic burdens of proof on claimants and respondents. Algorithmic discrimination thus seems impossible to prove and, consequently, falls outside the scope of application of EU non-discrimination law. Two possible solutions are proposed. First, as regards effectiveness, a right to access evidence in favour of victims of algorithmic discrimination should be recognized, through a joint reading of EU non-discrimination law and the General Data Protection Regulation. Second, to allocate the burden of proof more proportionately, an extension of the grounds for defence of respondents could allow them to establish that biases were autonomously developed by an algorithm.

Disciplines :

European & international law

Author, co-author :

Grozdanovski, Ljupcho ; Université de Liège - ULiège > Cité

Language :

English

Title :

In Search for Effectiveness and Fairness in Proving Algorithmic Discrimination in EU law

Publication date :

2021

Journal title :

Common Market Law Review

ISSN :

0165-0750

eISSN :

1875-8320

Publisher :

Kluwer Law International, Alphen aan den Rijn, Netherlands

Volume :

Issue :

Pages :

99-136

Peer reviewed :

Peer reviewed

Available on ORBi :

since 14 September 2023

Statistics

Number of views

19 (6 by ULiège)

Number of downloads

4 (3 by ULiège)

More statistics

Scopus citations^®

Scopus citations^®
without self-citations

Bibliography

1. Artificial Intelligence “refers to a class of computer programs designed to solve problems requiring inferential reasoning, decision-making based on incomplete or uncertain information, classification, optimization, and perception”. See Bathaee, “The Artificial Intelligence black box and the failure of intent and causation”, 31 Harvard Journal of Law & Technology (2018), 890-938, at 898.
2. Dastin, “Amazon scraps secret AI recruiting tool that showed bias against women”, Reuters (2018), available at (all websites last visited 30 Nov. 2020).
3. Autor, “Why are there still so many jobs? The history and future of workplace automation”, 29 Journal of Economic Perspectives, (2015), 3-30.
4. Acemoglu, “Technical change, inequality and the labour market”, 40 Journal of Economic Literature, (2002), 7-72; Acemoglu and Restrepo, “Robots and jobs: Evidence from US labor markets”, NBER Working Paper 23285 (2017), available at .
5. Nedelkoska, Diodate and Neffke, “Is our human capital general enough to withstand the current wave of technological change?”, Working Paper No. 93 (2018), available at .
6. See Acemoglu (2002), op. cit. supra note 4. For the EU, see Grozdanovski, “Will technological skill bias exacerbate residual market inequalities? Lessons from EU non-discrimination law”, 1 Labour Law Journal (2020), 58-70.
7. Hacker, “Teaching fairness to artificial intelligence: Existing andnovel strategies against algorithmic discrimination under EU law”, 55 CML Rev. (2018), 1143-1186, at 1145.
8. See infra, section 3.
9. For the purpose of this study, algorithmic opacity will be understood as the lack visibility of computational processes such that humans are not able to inspect the inner workings of these processes to ascertain for themselves how the results and conclusions were computed. See Paudyal and Wong, “Algorithmic opacity: Making algorithmic processes transparent through abstraction hierarchy”, Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 61-1 (2018), 192-196.
10. Bathaee, op. cit. supra note 1, 890-938, at 903.
11. Ibid., at 892.
12. See Solum, “Legal personhood for Artificial Intelligences”, 70 North Carolina Law Review (1992), 1231-1287; Cerka et al., “Is it possible to grant legal personality to Artificial Intelligence software systems?”, 33 Computer Law & Security Review (2017), 685-699.
13. Hacker, op. cit. supra note 7.
14. The system of evidence in EU non-discrimination refers to the rules governing the establishment and rebuttal of the presumption of discrimination in the Non-discrimination Directives. These directives, and the rules on evidence they contain, will be examined infra, section 3.2.1.
15. See Katyal, “Private accountability in the age of Artificial Intelligence”, 66 UCLA Law Review (2019), 56-141, at 106.
16. Evidence will be defined as “anything presented to the senses and offered to prove the existence or nonexistence of a fact”. See Nutter, “Machine learning evidence: Admissibility and weight”, 21 Journal of Constitutional Law (2019), 919-958, at 920.
17. See Walton, Legal Argumentation and Evidence (Pennsylvania State University Press, 2002), atp. 17.
18. Autonomy implies self-regulation, the ability to work in, and adapt to, a given environment with limited knowledge and resources. See Pennachin and Goertzel, “Contemporary approaches to Artificial General Intelligence” in Pennachin and Goertzel (Eds.), Artificial General Intelligence (Springer, 2007), pp. 1-30, atp. 10.
19. See infra, section 4.
20. General Data Protection Regulation (EU) 2016/679 of the European Parliament and of the Council on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, O.J. 2016, L 119/1 (GDPR).
21. See infra, section 4.2.
22. See infra, section 3.
23. Art. 2(2)(b) Directive 2000/78 establishing a general framework for equal treatment in employment and occupation, O.J. 2000, L 303/16. The focus on indirect discrimination is also explained by the fact that in such cases, different treatment may be justified under the EU non-discrimination principle. Justif ication of algorithmic discrimination will be addressed infra, section 4.
24. Bathaee, op. cit. supra note 1, at 901.
25. Human supervision refers to the level of human oversight over the programming or use of AI. The European Commission’s High-Level Expert Group on AI (hereafter HLEG) has distinguished three levels of human involvement (human in the loop, human on the loop and human in command) for the purpose of effectively detecting, auditing and rectifying incorrect decisions taken by “perfectly” functioning systems, for those situations where the AI system’s decisions significantly affect individuals. See HLEG, “Policy and investment recommendations for trustworthy Artificial Intelligence” (2019), available at
26. Shmueli, “To explain or to predict?” 25 Statistical Science (2010), 289-310, cited in Vijayakumar and Cheung, “Replicability of machine learning models in the social sciences. A case study in variable selection”, 226 Zeitschrift für Psychologie (2018), 259-273, at 259.
27. Nutter, op. cit. supra note 16, at 929.
28. Kamath, Liu and Whitaker, Deep Learningfor NLP and Speech Recognition (Springer, 2019), atp. 42.
29. Faliagka et al., “Taxonomy development and its impact on a self-learning e-recruitment system”, 8th International Conference on Artificial Intelligence Applications and Innovations (AIAI)(2012), 164-174, atp. 164.
30. Singh, Rose, Visweswariah, Vijil and Kambhatla, “PROSPECT: A system for screening candidates for recruitment”, Conference: Proceedings of the 19th ACM Conference on Information and Knowledge Management, 2010, Toronto, Canada, available at , at p. 3.
31. Ibid., at pp. 4-5.
32. Saaty, “The analytical hierarchy process - What it is and how it is used”, 9 Math ’I Modelling (1987), 161-176.
33. Ranjith and Cheung, op. cit. supra note 26, at 260; Kamath et al., op. cit. supra note 28, at p. 41.
34. Glauner, Valtchev and State, “Impact of biases in big data”, ESANN 2018 Proceedings, European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, Bruges, Belgium, 25-27 April 2018, available at , 645-654, at p. 648.
35. See Wajcman, “New connections: Social studies of science and technology and studies of work”, 20 Work, Employment & Society (2006), 773-786, at 781. See also Lambrecht and Tucker, “Algorithmic bias? An empirical study of apparent gender-based discrimination in the display of STEM career ads”, 7 Management Science (2019), 2966-2981, Khademi and Honovar, “Algorithmic bias in recidivism prediction: A causal perspective”, 34th AAAI conference on Artificial Intelligence (AAAI-20), available at .
36. The training data essentially provides statistics of biased human behaviour. E.g. annual bonus numbers, as an objective metric of employee performance may reflect a deeply rooted gender bias. See Baer, Understand, Manage, and Prevent Algorithmic Bias. A Guide for Business Users and Data Scientists (Apress, 2019), at p. 70.
37. Ibid., at p. 76.
38. Vijayakumar and Cheung, op. cit. supra note 26, at 260.
39. Baer, op. cit. supra note 36, at p. 90.
40. Ibid., at p. 91.
41. Kowert, “The foreseeability of human-artificial intelligence interactions”, 96 Texas Law Review (2017), 181-204, at 204.
42. Ibid., at 192.
43. See Ho, A Philosophy of Evidence Law: Justice in the Search for Truth (OUP, 2008), at p. 86.
44. White Paper, “On Artificial Intelligence - A European approach to excellence and trust”, COM(2020)65 final.
45. This alternative will be examined infra, section 4.
46. See in particular HLEG, Policy and investment recommendations cited supra note 25, at p. 41: “We urge policy-makers to refrain from establishing legal personality for AI systems or robots. We believe this to be fundamentally inconsistent with the principle of human agency, accountability and responsibility, and to pose a significant moral hazard.”
47. See Bathaee, op. cit. supra note 1.
48. In the Louis Marx case of 1958, a US customs court classified a robot as “a mechanical device or apparatus, a mere automation, that operates through scientific or mechanical media”. See 40 Cust. Ct. 610, 610 (1958) cited in Calo, “Robots in American Law”, University of Washington School of Law Research Paper No. 4 (2016), available at , atp. 14.
49. See Solum, op. cit. supra note 12.
50. Paisner, Cox and Maynord, “Goal-driven autonomy for cognitive systems”, Proceedings of the Cognitive Science Society (2014), available at , 2085-2090, at 2085.
51. Korinek and Stiglitz, “Artificial Intelligence and its implications for income distribution and unemployment”, NBER Paper (2017), available at , atp. 3.
52. Bathaee, op. cit. supra note 1, at 919.
53. Ibid., at 934.
54. Ibid.
55. 1989 A.M.C. 1955 (1989), cited in Calo, op. cit. supra note 48, at 17.
56. This case deals with robots used to explore a ship sunk in the Atlantic Ocean. The robot made findings which raised the issue of whether the human explorers could gain custody, control and possession of those findings, even though they did not, themselves, make the discovery. The US court answered aff irmatively, ruling that robots performed tasks and were “able to generate live images of the wreck and had the further ‘capability to manipulate the environment ’ at the direction of people”.
57. See e.g. Case C-157/15, Achbita, EU:C:2017:203.
58. See e.g. Case C-143/16, Abercrombie & Fitch Italia Srl, EU:C:2017:566.
59. See Case C-109/88, Danfoss, EU:C:1989:383.
60. See infra, section 4.
61. Reasonableness in the context of legal evidence will be defined as a form of rational reasoning, conducive to arguments supported by appropriate premises. See Picinali, “The meanings of ‘reasonableness’: Dispelling the ‘floating’ reasonable doubt”, 76 Modern Law Review (2013), 845-875, at 856.
62. See supra, section 2.
63. Directive 2000/78.
64. Council Directive 2000/43 implementing the principle of equal treatment between persons irrespective of racial or ethnic origin, O.J. 2000, L 180/22.
65. Directive 2006/54 of the European Parliament and of the Council on the implementation of the principle of equal opportunities and equal treatment of men and women in matters of employment and occupation (recast), O.J. 2006, L 204/23.
66. Directive 2000/78, Art. 10(1); Directive 2000/43, Art. 8(1); Directive 2006/54, Art. 19(1). Each of these provisions take over the rule of evidence originally set out in Art. 4(1) of Directive 97/80/EC on the burden of proof in cases of discrimination based on sex, O.J. 1998, L14/6, repealed by Directive 2006/54. See also Case C-415/10, Meister, EU:C:2012:217, para 22; Case C-303/06, Coleman, EU:C:2008:415, para 54; Case C-81/12, Asociaba Accept, EU:C:2013:275, para 55; and Case C-83/14, CHEZ, EU:C:2015:480, para 85.
67. See supra, section 1 (Introduction).
68. See Case C-415/10, Meister, para 22.
69. See namely Directive 2000/78, Recital 31.
70. See Case C-54/07, Feryn, EU:C:2008:397.
71. Opinion in Case 109/88, Danfoss, EU:C:1989:228.
72. Ibid., para 30.
73. See infra, section 3.2.2.
74. The confusion stemming from the absence of detail in the definition of the standard of proof to be applied to the indicia has been reinforced by the different linguistic versions of the Non-discrimination Directives, as was shown in Case C-394/11, Belov. The referring court asked the ECJ if the facts established by a victim of discrimination must allow a conclusion that there has been discrimination, or whether the mere presumption that there has been discrimination is sufficient, under Directive 2000/43. See Opinion in Case C-394/11, Belov, EU:C:2012:585, para 87.
75. See Directive 2000/78, Art. 10(1); Directive 2000/43, Art. 8(1); Directive 2006/54, Art. 19(1).
76. Ibid.
77. E.g. the ECJ has declared statistics admissible as evidence of indirect discrimination. See e.g. Joined Cases C-4 & 5/02, Schönheit, EU:C:2003:583, para 72. See also Directive 2000/78, Recital 15.
78. We allude to the res ipsa loquitur principle, traditionally referring to facts beyond doubt or dispute.
79. CaseC-54/07, Feryn.
80. Ibid., para 32.
81. See Directive 2000/78, Recital 15; Opinion in Case C-400/93, Royal Copenhagen, EU:C:1995:48, para 48.
82. Situation testing “requires the selection of pairs of individuals who have similar profiles for all the characteristics relevant for accessing the job, goods or service in question, but who differ with regard to one protected ground, such as sex, origin or disability”. See Ringelheim, “The burden of proof in anti-discrimination proceedings. A Focus on Belgium, France and Ireland”, 2 European Equality Review (2019), 49-64, at 58-59.
83. Opinion in Case 109/88, Danfoss.
84. Ibid., para 52 (emphasis added).
85. Ibid., para 54.
86. Case C-531/15, Otero Ramos, EU:C:2017:789.
87. Ibid., para 28.
88. Council Directive 92/85 on the introduction of measures to encourage improvements in the safety and health at work of pregnant workers and workers who have recently given birth or are breastfeeding, O.J. 1992, L 348/1.
89. Case C-531/15, Otero Ramos, para 51.
90. Ibid., para 60.
91. Art. 19(4)(a) of Directive 2006/54 states that the rules reversing the burden of proof in Art. 19(1) also apply to situations covered by Directive 92/85, in so far as discrimination on grounds of sex is concerned. See Case C-531/15, Otero Ramos, para 53.
92. Case 109/88, Danfoss.
93. Case C-415/10, Meister.
94. Ibid., para 13.
95. Ibid., para 15. In a similar vein, in a case from 1988, dealing with separate recruitment processes for male and female candidates applying for positions as public servants, the ECJ found that the criteria used for the allocation of positions to male or female workers was not transparent. However, the Court examined this lack of transparency not through the prism of the employer’s duties, but through the prism of the Commission’s possibility to supervise situations derogating from the non-discrimination principle under the Equal pay Directive (76/207, at the time of the ruling); see Case 318/86, Commission v. France, EU:C:1988:352, paras. 25-29.
96. Case 170/84, Bilka, EU:C:1986:204.
97. Case C-127/92, Enderby, EU:C:1993:859.
98. Ibid., para 14 (emphasis added).
99. Case C-415/10, Meister.
100. Ibid., para 13.
101. Ibid., para 46.
102. Ibid., para 47 (emphasis added).
103. This evidentiary difficulty does not comply with the requirement of adequate judicial protection against victimization. See Directive 2000/78, Recital 30. This Directive also states that “the rules on the burden of proof must be adapted when there is a prima facie case of discrimination and, for the principle of equal treatment to be applied effectively, the burden of proof must shift back to the respondent when evidence of such discrimination is brought” (Recital 31, emphasis added).
104. Case C-415/10, Meister, para 47.
105. See infra, section 4.1.
106. Art. 4(2) GDPR: “‘processing’ means any operation or set of operations which is performed on personal data or on sets of personal data, whether or not by automated means, such as collection, recording, organisation, structuring, storage, adaptation or alteration, retrieval, consultation, use, disclosure by transmission, dissemination or otherwise making available, alignment or combination, restriction, erasure or destruction”.
107. See supra, section 1 (Introduction).
108. Kowert, op. cit. supra note 41, at 191.
109. White Paper cited supra note 44, at p. 18.
110. Case C-415/10,Meister, para 22.
111. GDPR, Recital 71 (emphasis added).
112. The grounds for discrimination in this provision are religion or belief, disability, age, and sexual orientation as regards employment and occupation.
113. The grounds for discrimination in this provision are sex, race, colour, ethnic or social origin, genetic features, language, religion or belief, political or any other opinion, membership of a national minority, property, birth, disability, age, and sexual orientation.
114. Algorithmic recruitment may be an automated decision within the meaning of Art. 22(1) ifjob applicants are shortlisted solely by the algorithm, i.e. without human intervention. See Veale and Edwards, “Clarity, surprises, and further questions in the Article 29 Working Party draft guidance on automated decision-making and profiling”, 34 Computer Law & Security Review (2018), 398-404, at 400. For the sake of simplicity, the assumption in this study will be that a shortlist of job applicants, generated by an algorithm without any additional intervention by a human recruiter, may qualify as a decision solely based on automated processing.
115. The term “individual” is emphasized in order to distinguish the general rights and safeguards for data subjects (Ch. 3, Sec. 2, Arts. 13-15 GDPR) and the Art. 22 safeguards, applied in the context of individual decision-making. This distinction is relevant because there has been some debate, which will not be examined here, on whether the general requirements apply in cases of individual automated decisions.
116. This seems to be the prevailing interpretation of Art. 22(1) GDPR. See also Veale and Edwards, op. cit. supra note 114, at 399.
117. Art. 22(2) sets out three derogations to the right to object, established in Art. 22(1): (a) when an automated decision is necessary to enter into, or perform a contract; (b) when it is authorized by Union or Member State law to which the controller is subject and which also lays down suitable measures to safeguard the data subject’s rights and freedoms and legitimate interests; and (c) when the decision is based on the data subject’s explicit consent.
118. Art. 22(2)(a) GDPR.
119. See Mendoza and Bygrave, “The right not to be subject to automated decisions based on profiling” in Synodinou et al., EU Internet Law. Regulation and Enforcement (Springer, 2017), pp. 77-98.
120. Art. 22(3) GDPR.
121. Ibid., (emphasis added).
122. Art. 9(2)(a) GDPR.
123. Art. 9(2)(b) GDPR (emphasis added).
124. Art. 1, Directive 2000/78.
125. Art. 1(a), Directive 2006/54.
126. Art. 9(1), GDPR.
127. Watcher, Floridi and Mittelstadt, “Why a right to explanation of automated decision-making does not exist in the General Data Protection Regulation”, 2 International Data Privacy Law (2017), 1-25, at 3.
128. It should be stressed that Art. 22(3) does not explicitly refer to a “right to an explanation”. Such a right has been derived from a combined reading of that Article and of Recital 71 GDPR. According to Watcher et al., the ex post explanation, under Art. 22(3) would include “specific decisions, i.e. the rationale, reasons, and individual circumstances of a specific automated decision”. See Watcher et al., op. cit. supra note 127, at 3.
129. Veale and Edwards, op. cit. supra note 114, at 402-403.
130. See “Amazon scrapped ‘sexist AI’ tool”, available at .
131. Watcher et al., op. cit. supra note 127, at 3.
132. Ibid.
133. Assuming that age corresponds to the stage of professional advancement, a 38-year-old and a 41-year-old would, presumably, be considered to be at equivalent stages of advancement. This would make it difficult to justify an age difference (of three years) as the reason to shortlist the former but not the latter. Age discrimination would have been more visible if a claimant was 41 years old, but the selected candidates were younger than 30. In that case, age would appear to be a clearer criterion to give preference to entry-level over early-mid-level workers.
134. See supra, section 3.1.
135. Case C-143/16, Abercrombie & Fitch Italia.
136. See Wajcman, op. cit. supra note 35.
137. See e.g. Case C-506/06, Mayr, EU:C:2008:119, in which a female worker was dismissed because she was at an advanced stage of in vitro fertilization, raising a high probability of maternity leave.
138. See Bathaee, op. cit. supra note 1, at 920.
139. See e.g. Case C-143/16, Abercrombie & Fitch Italia.
140. Art. 4(1) applies when a difference of treatment is based on a characteristic that constitutes a genuine and determining occupational requirement, provided that the objective is legitimate and the requirement is proportionate. The exception to discrimination under Art. 4(2) applies when, in occupational activities within churches and other public or private organizations the ethos of which is based on religion or belief, a person’s religion or belief constitute a genuine, legitimate and justified occupational requirement, having regard to the organization’s ethos.
141. Seee.g. Case C-157/15, Achbita, EU:C:2017:203.
142. This was, mutatis mutandis, the issue in Case C-68/17, IR, EU:C:2018:696.
143. See e.g. Case C-143/16, Abercrombie & Fitch Italia, paras. 41 et seq.
144. Case C-229/08, Wolf, EU:C:2010:3.
145. See the ECJ’s case law relative to the assessment of health or environment threats, justifying the enactment of safeguard measures. See Case C-236/01, Monsanto, EU:C:2003:431, para 106.
146. Evidence is here defined as being an accurate (or at least, convincing) representation of reality, on the basis of which fact finders may reach a true verdict or a true finding on a disputed proposition. See Ho, op. cit. supra note 43, at p. 86.
147. See supra, section 1 (Introduction).
148. See supra, section 3.
149. See supra, section 2.
150. US Court of Appeals for the 7th Circuit, US v. Coscia, 866 F.3d 782 (2017).
151. Bathaee, op. cit. supra note 1, at 909.
152. It is true that statistical evidence has generally been declared admissible in support of evidence offered by victims (see Joined Cases C-4 & 5/02, Schönheit). However, in the event that a respondent is given the possibility to rebut the presumption of human agency, it may be assumed that statistical evidence can also be declared admissible in their offer of evidence.
153. See Grimshaw, Fagan, Hebson and Tavora (Eds.), Making Work More Equal. A New Labour Market Segmentation Approach (Manchester University Press, 2017), at p. 251. See also Case C-1/95, Gerster, EU:C:1997:452.
154. For a definition of overfitting see supra, section 2.
155. Case C-97/08 P, Akzo, EU:C:2009:536.
156. See supra, section 3.1.
157. Bathaee, op. cit. supra note 1, at 936.
158. An analogy can be drawn on this point with the ECJ’s test on the liability of online platforms, under the eCommerce Directive (Directive 2000/31 of the European Parliament and of the Council on certain legal aspects of information society services, in particular electronic commerce, in the Internal Market, O.J. 2000, L 178/1). The main criterion used by the Court to determine whether online platforms - as intermediaries between providers and beneficiaries of online services - are liable seems to be the degree of intervention (the active or passive role) of the platforms in defining the modalities according to which these services are provided. See Case C-324/09, L’Oréal, EU:C:2011:474 and Case C-610/15, StichtingBrein, EU:C:2017:456. Van Ecke observed, on the assessment of the active/passive role of platforms, that “the ‘passive-reactive’ scheme founded on the notice-and-take-down procedure, is… gradually being replaced by ex-ante monitoring of users”. See Van Ecke, “Online service providers and liability: A plea for a balanced approach”, 48 CML Rev. (2011), 1455-1502, at 1459. The possibility and extent of the ex ante monitoring of users by an online platform like Facebook was, inter alia, raised in Case C-18/18, Glawischnig-Piesczek, EU:C:2019:821. For a comment of this case, see Kuczerawy and Rauchegger, “Court of Justice injunctions to remove illegal online content under the eCommerce Directive: Glawischnig-Piesczek”, 57 CML Rev. (2020), 1495-1526. However, the main differences between AI users and online platforms regarding the proof of liability are: (1) AI use implies control and general command (an active role) of the user; platforms are prima facie classified as “mere” intermediaries (i.e. presumably assume a passive role); (2) because AI use includes control and command, the accountable agent is presumed to be the user. The same presumption does not apply to platforms; their liability is not presumed but established.
159. See 1989 A.M.C. 1955 (1989) cited in Calo, op. cit. supra note 48.
160. See Council Directive 85/374/EEC on the approximation of the laws, regulations and administrative provisions of the Member States concerning liability for defective products, O.J. 1985, L 210/29. As per Art. 4 of this Directive, “the injured person shall be required to prove the damage, the defect and the causal relationship between defect and damage” (emphasis added). Assuming that an AI user managed to establish the presence of an embedded or machine-learnt unfair bias in a discrimination case, they could rely on the evidence adduced in that case, to meet the evidentiary requirements set out in Art. 4.
161. Bathaee, op. cit. supra note 1, at 893.
162. According to some predictions, superintelligent systems (i.e. systems that outperform humans across the board) may come about as soon as 2075. See Müller and Nich, “Future progress in Artificial Intelligence: A survey of expert opinion” in Müller (Ed.), Fundamental Issues of Artificial Intelligence (Springer, 2016), pp. 553-571, at p. 553.