I. The code of the programme
A beautiful metaphor for the lines that follow is the cyberpunk series and film, “Ghost in the Shell.” Current EU data protection is the “shell” representing the individual-centric legal framework. In contrast, the “ghost” represents the inferred group data soul that the AI ActFootnote 1 and the GDPRFootnote 2 are struggling to rewire. This is because the dominant framework for data protection is largely individual-centric, where the General Data Protection Regulation (GDPR) focuses on safeguarding personal data at the individual level, primarily through a consent-based mechanism. While this paradigm has been in place for decades, the rapid evolution of data-driven technologies, especially Artificial Intelligence (AI), has begun to expose critical gaps in its protective reach. A key oversight in the current legislative architecture concerns the impact on groups of people who never consented to, nor knowingly contributed, the data that underlies automated decision-making, which may negatively affect them. Under the GDPR, personal data that is effectively anonymised is excluded from regulatory oversight, and the Regulation provides a limited legal instrument for collective redress. Yet modern AI systems can de-anonymise personal data or derive sensitive insights about entire demographics with relative ease. A paradigmatic example is the use of some service providers of “non-user’s information.” You may not use a service, but your personal data is still exploited as a friendly inference. This is how friendship requests work in practice on social media.Footnote 3
The paper’s thesis is that the GDPR’s safeguards are triggered by data “relating to” an identifiable person, whereas contemporary practices can de-anonymise data or generate actionable inferences at scale that affect not only individuals but also target groups of people as such, without requiring controllers to single them out as data subjects. In that sense, the privacy harm is real, but the decisive element lies within the scope of data protection and its remedial architecture (identifiability, Articles 77–82, limits of collective standing).
This is why this paper is framed primarily in EU data protection terms because the core regulatory problem it targets is not only “privacy” in an abstract sense, but the governance of data processing and inference when datasets are treated as anonymous while still enabling personal (including sensitive) inferences about persons.
In this context, the European Union has introduced the AI Act. This new legislative tool seeks to reduce AI’s risks to fundamental rights while promoting AI’s beneficial uses. Among its many features, it highlights the potential relevance of “groups of persons” as subjects of rights and potential harms. Given the interconnection between the GDPR and the AI Act, the AI Regulation provides insight into the potential of collective rights in data protection. This study builds upon an earlier analysis of the draft AI Act,Footnote 4 extending that work to incorporate the final text of the legislation and situating it more firmly within existing scholarship on group privacy and data protection. This paper uses “group data protection” to describe the extension of EU data protection law’s rights, obligations and remedies to groups of persons, including those constructed by AI and not individually identifiable. I reserve “group privacy” for the normative and theoretical debate that motivates such a shift.
Specifically, this paper makes two contributions. First, it identifies the need to evolve beyond a strictly individual-centric data protection framework, underscoring the unique vulnerabilities of groups, particularly when automated systems create or target them without their members’ explicit awareness or permission. Second, it explores both theoretical and practical applications of such a group within existing scholarship and legislation. It reviews existing interpretations of “group privacy” in the literature. It suggests potential pathways to reinterpret existing provisions and embed group-level data protections, specifically by the AI Act and the GDPR, as a remedy before any future large-scale reform.
In this vein, there are some technical limitations to point out as well. The study is limited by the evolving nature of the legislation in question. Given that the provisions discussed in the text are part of the first-ever Regulation on AI, it is not surprising that there is no reference jurisprudence on which this analysis could rely. In addition, the notions of “groups of persons” or “group privacy” represent an underexplored terrain in both theory and practice, thereby demanding future research on topics beyond the scope of this analysis (e.g., the tension between group versus individual privacy, and potential negative externalities of expanding data protection to group claims).
Methodologically, I combine a literature review (doctrine and case law) with close textual analysis of the GDPR, the AI Act and the RAD, read against their historical trajectories. The paper proceeds as follows: Section 2 (The European Data Protection Individualistic Shell) provides a background on the evolution of data protection, culminating in the centrality of the GDPR, and also examines some topical critiques (through Article 80 GDPR and RAD) of the main thesis defended here. Section 3 (The Network) reframes privacy as a public good and, drawing on the group privacy literature, motivates a shift to group protection. Section 4 (Uploading to the Cloud) revisits the draft of the AI Act to show where the adopted text diverges. Section 5 (A Sealed Ghost) examines how the current AI Act treats “groups of persons.” Section 6 (Hacking the Shell) uses online behavioural advertising to illustrate the loopholes that emerge from this analysis. Finally, Section 7 (Rewiring the Shell) lays out concrete recommendations, while Section 8 (System Sync) concludes.
II. The European data protection individualistic shell
In order to open the “shell” and explain this paper’s emphasis on data protection, it is important to clearly distinguish it from privacy as two distinct, yet closely related rights. There has been debate over whether privacy and data protection constitute separate rightsFootnote 5 and/or should be recognised as fundamental rights.Footnote 6 Whatever the scholarly discussions are, the European constitutional reality is unambiguous. Article 7 and Article 8 of the EU CharterFootnote 7 establish two fundamental rights, and this premise is fully adopted by this paper. The two rights differ primarily in their moral bases, duty bearers, scopes, and restrictions. Throughout this paper, I use privacy to mean Article 7 EU Charter and data protection to mean the distinct right under Article 8 EU Charter, operationalised primarily through the GDPR. Unless stated otherwise, the analysis proceeds in data protection terms, whereas privacy references are used only as a normative background.
Privacy is intrinsically intertwined with liberalism’s core normative principle, individualism,Footnote 8 whose central claim is the individual’s sovereignty over governmental interference,Footnote 9 enabling him to exercise his will freely.Footnote 10 What enables an individual to pursue freedom is his autonomy in a liberal-democratic society.Footnote 11 Thus, the “right to privacy consists essentially in the right to live one’s own life with a minimum of interference,”Footnote 12 drawing a boundary between the private self and the public, the individual from society. It is not surprising that the main duty bearer is the state, which must abstain from unjustified interference and lay down the necessary positive measures in its legal order.
While privacy’s moral framing may seem broad, its practical implications, compared with data protection, are narrower as interpreted by the Court of Justice of the European Union (CJEU). In Case T-194/04 Bavarian Lager, the Court explicitly stated that the scope of the right does not automatically extend to all personal data and, therefore, not all personal data falls under the notion of private life. This is not the case with data protection, where all information relating to an identifiable natural person is subject to the safeguards of the right; therefore, it encompasses virtually every kind of data falling within the definition provided in the GDPR Art. 4(1). Hence, data protection is broader in the information it covers but narrower in the persons it protects. This is why courts often do not address privacy because any assessment of interference must pass through the data protection lawfulness filter, meaning that if processing is not based on one of the legal bases set out in Article 6 GDPR, the analysis stops there and the processing is unlawful.
Despite its importance, where the analysis proceeds under the Charter, an interference with the right to respect of private life (understood as privacy, Article 7 Charter) must, where justified, satisfy the limitations clause in Article 52 (1) (legitimate aim, appropriateness, necessity and proportionality stricto sensu). The Charter provisions are addressed to the EU institutions and Member States only when they are implementing Union law (Article 51 Charter). Direct obligations of private parties are typically fulfilled through implementing secondary law, such as the GDPR, which addresses disputes at the level of lawfulness (Art. 6 GDPR).
Under the European Court of Human Rights (ECHR), the analogous review is performed under Article 8 (2) (“in accordance with the law,” “legitimate aim,” and “necessary in a democratic society”), typically involving contextual proportionality balancing and, importantly, systematic and temporal elements.Footnote 13
Similar to privacy, data protection aims to introduce an equalising element by remedying power asymmetries arising from mass data processing and collection through a set of rules, grounded in principles and operationalised through tools such as consent, whose compliance is monitored by independent authorities. This architecture is anchored in the key tenet of the rule of law embodied in the principles in Article 5 GDPR, in particular, Article 5 (1)(a), lawfulness, fairness and transparency. Thus, EU data protection, and as some authors frame it, has an “essential procedural nature that makes it more objective as a right.”Footnote 14
In sum, both rights protect individuals against undue interference. Privacy should be regarded as a core right protected by a data protection filter. EU data protection exercises its filtering function through procedural rulesFootnote 15 grounded in principles such as fairness, transparency, lawfulness, independent oversight, democracy, and the rule of law.
This is why this paper locates “group protection” primarily within data protection law. The problem is not only interference with private life but also the governance of data processing and inferences from anonymous datasets. Privacy (also, group privacy in Section 3) provides a conceptual justification, but alone it does not provide an equivalent operational toolbox for regulating large-scale processing externalities or for designing enforceable procedural safeguards. Other legal branches, such as non-discrimination, consumer and tort law, remain relevant but do not replace data protection. They offer specific answers, bound to the field’s arsenal of remedies, tailored to the field’s development, methods, harms and safeguards.
The following paragraphs examine specific aspects of data protection to substantiate, both theoretically and normatively, the argument that the next evolution in data protection entails a move from an individualistic approach to one that safeguards groups as such, and not as a bundle of individual rights or a right of organisations.Footnote 16
This represents a potential for a shift in the current individual-centred data protection policy framework. The next paragraphs aim to expand this argument by providing an overview of the cornerstone developments that have shaped the present-day paradigm and by laying out the background for the subsequent discussion of the term “groups of persons” explored in another paper,Footnote 17 as potentially adversely affected parties by an AI system.
I identify three main factors driving the current individualistic approach to EU data protection: continuous technological evolution, led by AI; historical context; and the main law and its product, the GDPR.
1. The armour: built in the post-war for today’s digital challenges
One could treat the distinction between privacy and data protection as part of a larger historical development. European data protection emerged from the post‑war constitutional settlement that sought to constrain state surveillanceFootnote 18 and from the administrative possibilities opened by large‑scale record‑keeping,Footnote 19 while preserving persons’ “inherent dignity.”Footnote 20 Privacy and human‑rights jurisprudence recognised that information practices can implicate private life. Still, early data protection regimes were designed to regulate the collection, storage and use of personal information by public authorities and other powerful information holders.Footnote 21
In the late 1970s, amid rapid computerisation, European states adopted the first modern data‑protection laws. A decisive development followed in 1983 with the German Federal Constitutional Court’s Census judgment,Footnote 22 which framed “informational self‑determination” as an autonomy interest and embedded the individual as the central point of informational control.Footnote 23 The EU translated these principles into secondary law.
Directive 95/46/ECFootnote 24 harmonised national rules and introduced consent and purpose limitation as central safeguards. The General Data Protection Regulation (GDPR) 2016/679 later modernised and strengthened that framework, making the data subject’s rights directly enforceable across the Union through principles, lawfulness requirements, and accountability duties. The GDPR encapsulates Europe’s data protection philosophy, emphasising citizens’ rights and dignity. It prioritises individuals as the main managers of their data and rights.
Nevertheless, this individual‑centric design premise is increasingly misaligned with contemporary data practices. The economic and governance significance of data now lies less in their initial collection than in their secondary uses and inferences.Footnote 25 Modern analytics can act on populations without collecting data from each member, producing group‑level predictions and effects that reach persons who never provided data or consented.Footnote 26 At the same time, techniques relied upon to neutralise the risk, such as anonymisation or differential privacy, do not reliably prevent re‑identification or the extraction of meaningful inferences.Footnote 27 This mismatch frames the next subsection: the GDPR is broad in the information it covers, yet structurally narrow regarding who can trigger rights and remedies.
2. Sovereign of the GDPR: the individual in data protection
The GDPR deliberately centres the data subject as “both the object and the subject”Footnote 28 of protection. In practice, the regime is triggered by data that “relate to” an identifiable natural person and is enforced primarily through that person’s rights and remedies. Doctrinally, this individual focus is built into the definition of “personal data” (Art. 4(1) GDPR) and the identifiability threshold, under which a person is not (directly or indirectly) identifiable, and in such cases, the GDPR’s safeguards are not engaged. The same logic structures the rights catalogue and the remedial architecture in the Articles 77–82, which presuppose a claimant who can link themselves to a processing operation and contest its lawfulness and effects.Footnote 29 Also confirmed by the CJEU’s case law.Footnote 30 A topical criticism of that conclusion involves Article 80 GDPR and the Representative Actions Directive (RAD).Footnote 31
Article 80 GDPR allows non‑profit bodies to act on behalf of data subjects, and the RAD provides collective procedures for consumer harms. But these mechanisms do not displace the identifiability premise,Footnote 32 they mainly set who litigates and which remedies are practically available.
First, Article 80 operates as a representation of identified or identifiable persons. Second, Article 80(2) enables mandate‑free actions, but its practical function is typically to obtain injunctive relief rather than to distribute compensation to persons who cannot be identified.Footnote 33 Third, the availability and design of mandate‑free collective actions remain highly dependent on national procedural and consumer‑law choices, which tends to reintroduce an identification requirement once monetary redress is sought. Therefore, it could be concluded that para. 2 of Article 80 is rather an exception, “an option,”Footnote 34 which should be implemented by national governments according to their legislation,Footnote 35 rather than a feature of the legislative architecture underlining the data protection legislation, in particular, the GDPR.
The RAD makes this split explicit. Article 8 RAD involves measures to stop or prohibit infringing activity, with no consumer mandate under Article 8 (3). Redress measures under Article 9, like compensation, require consumers to identify (Article 9 (2)). Both injunctions and mass claims against GDPR infringements can be pursued under Article 2(1) and Annex I RAD.
However, given that GDPR would take precedence in data protection matters, redress without a mandate, ergo, without identification, would be unlawful because it would contradict the legislators’ intention with Article 80 (2) where pursuing compensation under Article 82 GDPR without the express mandate of the damaged is prohibited with the intention to prevent frivolous claims driven by pecuniary incentives.Footnote 36 The practical consequence is that stopping an infringement can therefore benefit everyone exposed to a practice, including those who are “anonymous” to the defendant in practice. Compensation, by contrast, is confined to those who can be singled out as beneficiaries. This is why generally Member States permit mandate-free (GDPR) mass claims but confine them to injunctions. When compensation is sought, the national provisions require an opt-in mechanism, i.e., a mandate or a court-approved beneficiary list.Footnote 37
Across Article 80 GDPR, the RAD, and broader remedies, a gap may leave affected members of inferred groups uncompensated even if an illegal practice is proven and stopped. The group privacy literature explains why harms occur at the group level, even when individual rights and remedies are the default.
In this context, it is worth situating the following paragraphs within the broader scholarly debate on groups in privacy and data protection. The next Section explains the “network” around and why the individual focus creates public‑good externalities in an AI age and clarifies what kinds of “groups” any collective remedy must cover.
III. The network: privacy as a public good and data protection’s structural dimension
Since the inception of the European data protection framework, the focus has been on “the information regarding individuals, without distinguishing between their public and private nature.”Footnote 38 In recent decades, some authors have restated the importance of privacy as a public good as well. I therefore treat privacy as a public good whose under‑provision justifies group‑level legal guarantees.
On privacy, Regan sustains that it functions simultaneously as a common, public and collective value,Footnote 39 one that individuals share, democracies needFootnote 40 and markets undersupplyFootnote 41 of value shared by everyone. On a similar note, Fairfield and Engel argue that privacy has public-good aspects and thus externalities that spread across society and individuals, even though they may have put in place the highest levels of privacy protection.Footnote 42 Thus, the authors argue that privacy as a public good is poised to be what economists call a tragedy of the commons and invoke the environmental analogy.Footnote 43 This is because sharing private data yields an immediate benefit that outweighs the long-term risks or externalities (the so-called discounting effect in economics). In a world of AI, big data can infer correlations from seemingly unrelated data points that could adversely affect unaware and unidentifiable individuals whose data has never been collected.Footnote 44 This is precisely the gap this paper evidences: the traditional (individualistic) approach is insufficient, as individuals cannot properly protect themselves given the amount of information provided by third parties.Footnote 45 This is better exemplified by the example of someone who understands the significance of sharing a group picture for the people in the picture, but it is much more difficult to grasp the spill effect for third parties.
On data protection, Ben-Shahar draws a similar parallel between environmental damage, oil spills, data pollution and data spills, acknowledging the negative societal effect of data breaches but from the perspective of data protection. He ultimately calls for “an environmental law for data protection.”Footnote 46 The “structural dimension” of data protection is extensively examined by Felix Bieker and helps explain this structural insufficiency.
In line with the author’s thesis, the notion of data protection has two dimensions. On the one hand, it is the protection of “individuals and (minority) groups … from the processing of data by means of individual rights and structural guarantees.” On the other, the “structural” protection, which covers “systemic aspects of data protection such as institutional guarantees and organisational requirements.”Footnote 47 This means according to the author that data protection has two faces, one protects individuals (and groups) through specific rights, while the other regulates the rules and requirements, the State or corporate entities should abide irrespective of whether an individual is concerned and by this preventing any adverse effects on society. This, in turn, again, benefits individuals and groups in society. As defended throughout those paragraphs, Bieker acknowledges that specific groups of people (e.g., minorities), “even though there is no information on one specific individual,” are not protected sufficiently, if at all, despite the technological capacity to affect them adversely. This stanceFootnote 48 overlaps with and confirms Malgieri’s findings regarding the incomplete protection of vulnerable data subjects (and groups).Footnote 49 This is why Bieker concludes that full protection, especially with regard to profiling intended to predict behaviour, “the personal scope of the right (data protection), de lege ferenda, must be extended in order to enable groups of individuals to enforce this right collectively.”
Together, these insights reveal the core flaw of the EU data protection framework: systemic data protection harms are like pollution, widespread, collectively borne, and impossible for any individual data subject to avoid. To address these externalities, the law needs a clear explanation of what types of “groups” justify collective protection.
To begin answering this question, the most comprehensive anthology of perspectives on the types of groups is the book edited by Taylor, Floridi, and van der Sloot, where they ask, “whether, and how, we may be able to move from ‘their’ to ‘its’ privacy with regard to the group” in the context of data processing practices such as profiling.Footnote 50
Some authors refer to self-proclaimed and framed groups. Framed groups are formed voluntarily or through individual will, while self-proclaimed groups are not. Most authors describe them as designed or discovered, meaning they are identified during data processing or analytical processes based on specific datasets. The authors state most groups are “enforced,” thus designed or discovered.
Closely related is the distinction between self-aware and unaware groups. Self-aware groups include professional groups, unions or interest groups formed purposefully. Not-self-aware groups comprise individuals who share genetic features and are unable to foresee the implications of analysing their genetic code, either as a group or individually. This is because genetic information identifies the group and makes each member identifiable.
A genetic code is a stable feature, which is why some focus on distinguishing stable and fluid groups. Stability comes from awareness and will to belong. However, in data processing, group criteria may change quickly. For example, a person going to the gym wants to be a member, but using different transport modes each day momentarily clusters him with other passengers, some regulars, some tourists. While he is aware and willing to be grouped with gym peers, he probably would not see himself as part of a short-term fluid group travelling together.
Another distinction is made based on a group’s hierarchical or egalitarian structures. Legally recognised persons such as supervisors or legal representatives have a hierarchical nature, enabling them to defend collective interests before authorities, especially minority or lobbying interests. Such groups are often law-recognised due to their stability and self-proclaimed status. In contrast, horizontal groups lack legal recognition; each member must claim liability individually, even when harmed as part of a group, exemplified by LGBTQ + rights.
Finally, Floridi notes that traditionally, the distinction is made between real and fictional groups. However, he contends that groups are neither discovered nor invented but are intentionally designed. The legitimacy of these groups depends on the justification for choosing a specific level of abstraction (LoA). He claims that a group’s “naturalness” merely reflects how intuitive that level is. According to him, groups exist in relation to a justified purpose (epistemological rather than ontological) and insists that the “naturalness” of a grouping depends on the context. Therefore, the key point is not whether groups are separate entities, but which particular lenses are used to perceive them. Additionally, Floridi argues that designed groups should have rights as groups, not just as the sum of their individual members. I use this to support the view that legal protection should not depend on a group’s “naturalness,” allowing AI-designed groups to challenge processing and seek collective redress.
The AI Act is the most prominent legislative act, containing a direct and explicit reference to “groups of persons.” The following Sections analyse relevant provisions to examine how the concept of “groups,” evolved since the law’s proposal, how it fails upload to the legislative cloud redress, and how it has been coded into this context so to develop specific suggestions for enhancement at the end, freeing the ghost.
IV. Uploading to the cloud: AI groups of persons. From ambition to compromise
A brief look back at the European Parliament’s draft AI ActFootnote 51 serves functional purposes. First, it reveals which tools (definition of “group,” internal complaint procedures, group standing to lodge complaints) were available but finally discarded, thereby illuminating the precise gap the final Act leaves open. Second, the draft text supplies ready‑made language that could be revived in the mandatory 2028 review (Article 112 AI Act), showing that stronger group protection is politically and legally feasible.
1. Groups of persons in the proposed AI Act
The EP’s amendments deliberately introduced the notion of “groups of persons” in core definitional, transparency and redress provisions. This is reflected in multiple mentions of groups throughout the text. For example, the definitions of “significant risk,”Footnote 52 “affected person”Footnote 53 and “emotion recognition”Footnote 54 explicitly refer to the impact of AI on groups. The references to “groups” could be clustered into three main axes: potential discrimination against groups,Footnote 55 collective transparency and public trust through accuracy and impact assessmentsFootnote 56 and robust redress mechanisms via internal complaint procedures and a right to lodge group complaints.Footnote 57
Despite that, the draft left open several issues inherent to that introduction, such as what counts as “group” or “groups of persons,” whether providers can feasibly describe the system’s logic for each affected cohort,Footnote 58 and which groups would meet admissibility thresholds for lodging complaints.
2. The divergence in the final AI Act, and why it matters
The Artificial Intelligence Act was finally adopted on 13 June 2024. The adopted Act preserves the draft’s anti‑discrimination approach but removes every binding redress device.Footnote 59
The adopted AI Act retains several core provisions from the draft aimed at preventing discriminatory AI outcomes.Footnote 60 It upholds prohibitions on systems that distort human behaviour or exploit vulnerabilities, particularly regarding age, disability or socioeconomic status.Footnote 61 Recitals 57 and 58 and Annex IV preserve the draft’s transparencyFootnote 62 commitment by requiring systems’ instructionsFootnote 63 to disclose performance accuracyFootnote 64 broken down by the persons or groups to which the AI would be applied,Footnote 65 yet the draft’s stronger clause mandating group’s participation in impact assessments (Article 27 (c)(d)) was moved to recitals.
Interestingly enough, where the EP draft established that human oversight should be ensured “where decisions based solely on automated processing … produce legal or otherwise significant effects on the persons or groups of persons on which the system is to be used,”Footnote 66 this is missing in the adopted version.
Thus, the adopted version of the AI Act follows closely the EP’s proposal in the provisions acknowledging the role of AI in multiple aspects of life as well as the responsibility of third parties along the deployment chain. Nevertheless, this attestation, along with additional industry obligations, does not amount to an effective protection as much as a proper redress mechanism in case of a tort.
Crucially, however, the categorisation of provisions referring to “groups” has been modifiedFootnote 67 so that the recognition of the potential negative impact of AI practice on collective entities remains, while the redress mechanisms are entirely withdrawn.
Besides the Recital 96 AI Act, which recommends that deployers set up “complaint handling and redress procedures,” there is no compulsory provision which operationalises complaint mechanisms as suggested in the draft. In reality, the right to lodge a complaint (Article 85 AI Act) is limited to a public authority and to the individual in question. Therefore, the wording in Amendment 628 that “every natural persons or groups of natural persons shall have the right to lodge a complaint with a national supervisory authority” has been eliminated. This deletion re-creates the exact structural gap identified previously, namely, the law acknowledges collective risk but denies collective standing and compensation.
In sum, the term “groups” has been finally introduced, and some inconsistencies have been purged.Footnote 68 The concerns related to the definition of “groups of persons,” the system’s description requirements, and, most importantly, the mechanism for redress in the AI Act were largely left unresolved, and, with this, the structural gap highlighted above. Injunction-level protection survives, but collective redress has been written out. How those provisions would work is a question of time and jurisprudence, however, because the Parliament’s language shows that stronger protection was once regarded as feasible, those discarded clauses provide a ready blueprint for the first revision cycle mandated by Article 112 AI Act.
V. A sealed ghost: groups in the AI Act
Groups of potentially affected people are expressly protected under Article 5 AI Act. AI systems manipulating human behaviour “by appreciably impairing their ability to make an informed decision” or by exploiting “any of the vulnerabilities of a natural person or a specific group of persons due to their age, disability or a specific social or economic situation” are banned (Article 5 (1)(a) and (b)). In addition, systems, which perform evaluation or classification of “persons or groups of persons” based on their “behaviour or known, inferred or predicted personal or personality characteristics,” leading to a detrimental and/or unjustified outcome in other social contexts than those for which the data was collected, are prohibited as well (Article 5 (c)). While the Regulation prohibits emotional analysis in educational or workplace contexts,Footnote 69 it allows certain AI-driven evaluations (e.g., recruitment) under a high-risk label.Footnote 70 Recitals 56 and 57 elucidate the rationale behind by both referring to concerns related to historical discriminatory patterns to “women, certain age groups, persons with disabilities, or persons of certain racial or ethnic origins or sexual orientation.” This is why high-risk AI is subject to stricter rules in order to ensure accountability, transparency and ultimately trust in users.
Furthermore, to achieve better protection in line with the previous, Article 9 (9) foresees a risk management system, which takes into account the “adverse impact on persons under the age of 18 and, as appropriate, other vulnerable groups.” Recital 67 establishes that data sets should include statistical properties having into account a group dimension “in particular for persons belonging to certain vulnerable groups, including racial or ethnic groups.” This requirement is reflected in Article 10 (3). As already mentioned, transparency is at the core of the Regulation. Hence, Article 13 (3)(b)(v) mandates that the instructions for use should include the performance concerning “specific persons or groups of persons on which the system is intended to be used.” Article 27 establishes the requirements for the fundamental rights impact assessment for high-risk AI systems whose scope should include an evaluation of the consequences for groups. In that sense Recital 96 mandates that the objective is for the deployer “to identify the specific risks to the rights of individuals or groups of individuals.” This practically means that those collective entities should be identifiable ex-ante and hence stable and predictable. In similar vein, go other transparency provisions.Footnote 71 The groups referred to in the Annexes, in particular, when it comes to law enforcement (Annex III (6)(d)), the logic of the AI included in the technical requirements (Annex IV (2)(b)), or the functioning of the system (Annex IV (3)). The provisions adopted by the AI Act have their critics, however, who consider them inadequate, ambiguous or even additionally harmful.Footnote 72
It is apparent from the wording of the Articles and their corresponding Recitals mentioned here as well as those referring to discrimination throughout the Regulation, and in general, the adverse effects of AI practices that the lawmaker has specific and identifiable groups in mind. This conclusion is supported by the explicit wording of the provisions, which describe specific, identifiable and stable groups as well as legally recognised ones, following the taxonomy overviewed in Section 3. Also, multiple Recitals make express reference to age, gender and race as well as to religious, ethnic and sexual minorities.Footnote 73 All of these point out to groups, which are identifiable, stable and even legally recognised. For the sake of this recognition, it does not matter if they are aware or not of their grouping.
Interestingly enough, the Regulation insists also on vulnerability as a ground for data subject’s (also, groups’) protection. Although, as established elsewhere, European data protection legislation does not have a coherently developed notion of “vulnerability,” specific conditions or characteristics of individuals or groups make them particularly susceptible to exploitation, harm, or manipulation by AI systems because they may struggle to make informed decisions or protect their interests.Footnote 74 Thus, vulnerability is associated with dependency, weakness or being affected disproportionately from external factors, which is often perceived as deserving specifically targeted protection. Examples of vulnerable data subjects are minorities, socially or economically disadvantage people. Safeguards related to those communities are scattered in sectoral regulations.Footnote 75 Take for example asylum seekers and refugees, already labelled as vulnerable groups.Footnote 76 In this specific context of vulnerability, throughout the years, the European Data Protection Supervisor puts forward group-based arguments in order to halt data processing of “particularly vulnerable groups (e.g., asylum seekers and refugees),” which may be considered lawful but potentially exposing to risk “a group, as group … in a way that cannot be covered by ensuring each member’s control over their individual data.”Footnote 77 Another paradigmatic example of vulnerable groups are minors recognised as such vulnerable group in several acts.Footnote 78
While the exploration and analysis of the issues around vulnerable subjects go beyond this research, the term fits the concept of groups as meant in the Regulation. This is groups of persons who, by virtue of their age, condition or legal status, find themselves particularly susceptible to AI’s adverse effects. When reviewing Bieker’s taxonomy of data protection dimensions it has been divided in individual and structural, where the former, despite focused on single persons includes groups exemplified by minorities. Therefore, although the AI Act acknowledges that personal data may be used to harm groups, it does not go beyond the already established individual data processing dimension. Because neither Article 85 AI Act nor the GDPR’s Article 80(2) regime supplies a mandate-free, compensatory remedy, the RAD remains the only (partial) venue for group redress, yet even this Directive requires identifiability (Art. 9(2) RAD), leaving anonymous and affected groups rightless.
In conclusion, the AI Act falls short of freeing the sealed ghost. Other groups, emerging from AI processing, namely, fluid, designed, unaware and/or unstable groups remain unaddressed. In order to translate this concern into practical terms, the next Section examines those challenges through the lenses of the particular, yet revealing, example of online behavioural targeting for advertising.
VI. Hacking the shell: the unprotected groups in profiling and online behavioural advertising
A defence and justification for a sui generis right of groups is made elsewhere,Footnote 79 and empirically.Footnote 80 This Section continues the discussion by extending and developing the arguments made earlier by evidencing where the system is hacked and where a shift towards group data protection, with the challenges online behavioural advertising (OBA) poses to “non-traditional groups” is needed.
OBA operationalises exactly the gap identified above. Platforms and ad‑tech providers create and act upon AI‑designed groups that are neither self‑aware nor readily identifiable. Injunctions can stop the practice for everyone, but compensation remains unreachable for those cohorts because neither the GDPR nor the AI Act gives them collective standing.
OBA is the most widely used marketing strategy, enabling marketers and advertisers to target individuals or specific consumer groups differently.Footnote 81 It raises concerns about data protection, transparency, and societal impact.Footnote 82 While businesses and consumers benefit from targeted content,Footnote 83 risks include opacity in data collection, discrimination and adverse profiling.Footnote 84 Unlike general online personalisation, which includes both tailored pricing and advertising, OBA specifically profiles users based on their browsing behaviour and assumed affinities.Footnote 85 Users may voluntarily provide personal data (e.g., social media activity or consent), but much data is inferred through browsing history, location tracking, and predictive analytics.Footnote 86 Often, this information is shared with third parties through mergers or data sales, making its scope “virtually infinite.”Footnote 87 Personal data collection, however, is the first step in profiling consumers. The next one is the anonymisation of the data. It has been proven that anonymised data could be de-anonymised and re-assigned with minimal input by AI.Footnote 88 Yet, there is no need to carry out complex procedures to signal out someone, as profiling can occur without any individual identification.Footnote 89 When it comes to group’s profiling, AI systems can evaluate other data such as third parties’ data or data that does not fall within the scope of the GDPR in order to identify and target successfully groups of seemingly unrelated data subjects.
OBA’s profiling practically bypasses and undermines the available data protection guarantees. This means that individuals lack effective data protection because the anonymisation techniques used in online behavioural targeting are not covered by the GDPR.Footnote 90 Even the elevated requirements in Article 9 concerning sensitive data fall short as they fail to “acknowledge a potential relationship between assumed interests and sensitive personal traits.”Footnote 91 Moreover, the rights enshrined in Articles 15–22 GDPR do not apply in those cases where, after retrieving personal information, it has been anonymised.Footnote 92 In addition, all those rights are invoked solely by data subjects, and only when they establish a credible suspicion that they have been victims of some tort or discrimination. But even then, they would need to figure out a way to prove that they have been treated adversely compared to some others.Footnote 93 This is particularly dangerous when it comes to data collection and processing, which could be fed with diverse and unlimited sources of information, in order not only to infer opinions and tastes and profile users but also to induce a particular behaviour. Although the AI Act forbids AI uses which manipulate human behaviour, it does not backlist OBA technology.
Sandra Wachter explores how affinity profilingFootnote 94 turns out to challenge citizens’ privacy and data protection rights, discriminates against marginalised groups, and most importantly, affects adversely “non-traditional groups.”Footnote 95 OBA’s affinity profiling is defined as a technique which infers indirectly users’ or groups’ assumed interests. By presenting two cases,Footnote 96 the author evidences that a potentially affected person need not to be part for a disadvantaged group or have any relationship with a protected group in order to suffer adverse treatment or discrimination (e.g., the content shown in advertisements or the services they are provided as it is in the CHEZ judgement). Even with the broadest interpretation of data protection law, its framework fails to protect “new types of groups created through inferential analytics that do not map onto historically protected attributes.”Footnote 97
Here, the notion of “groups” comes to aid, however, not the same as the one established in the AI Act. This research calls for the protection of non-traditional or designed, fluid and unaware groups as autonomous entities, independent from the individual rights each person in the group holds. In this way, not only individuals who are aware and proactive in their behaviour on the Internet, not only specific, stable and legally recognised groups, would be protected, but also all those people who have been targeted as perceived to pertain to an AI designed group whose characteristics, in the end, do not fit into the GDPR or the AI Act’s framework. Those groups exist and as argued throughout this paper have no protection and means to seek redress.
Potential drawbacks might be considered as well. For example, implementing group rights can be a complex in terms of fitting into the existing data protection architecture and it should be balanced against existing rights. Moreover, it could be argued that this novelty may lead to the opposite situation where individual users shielded by the collective protection engage in a frivolous conduct in which they accept whatever conditions posed by the data controller with the idea that their rights would be covered by the enhanced group protection. Also, by relying on a collective approach, the unique preferences and concerns of individual users may be overlooked, leading to a one-size-fits-all approach that may not be suitable for all individuals. The criticism may be continued even further by including proposals to curb the predictive power of AI systems.Footnote 98
While this criticism may prove helpful for anticipating potential loopholes in establishing group rights in data protection, automatically generated groupings already occur. My intention has never been to limit technological advancements but to bring European protection of the people constituting inferred groups up to date, as they currently have no protection and no means to seek redress.
The architecture and gears of the group approach in digital legislation are the subject of another paper, which analysed “groups of persons” prior to the adoption of the AI Act.Footnote 99 This is why, in those lines, I build on it further, now that the AI Act has been adopted. The next Section provides practical suggestions for operationalising the concept of groups within existing legislation.
VII. Rewiring the shell: practical guide on operationalising groups
Although non-exhaustive, the following lines provide concrete and attainable objectives with the current legal framework for the introduction of non-traditional groups in the AI Act and the GDPR. Curiously enough, both Regulations receive their evaluation reports in 2028. This is a chance not to be overlooked.
1. Define “groups of persons” in the AI Act
It has already been argued that the AI Act should not only foresee provisions which protect groups of data subjects from discrimination, providing them with transparency tools and redress mechanisms, but also it should delimit the contours of those clusters. As already explored in the previous Sections, the AI Act refers to specific groups of people who, by virtue of their age, gender, or social and economic status, might be markedly vulnerable. Nevertheless, delimitation and interpretation of the term used in the AI Act has been done in this analysis, and not in the Regulation, which leaves the door open to regulatory and judicial interpretation and consequently legal uncertainty and inconsistencies in application and enforcement. Clarifying the interpretation of “groups” in Article 3 AI Act, recognising that groups can be machine constituted based on assumed, inferred or designed characteristics (e.g., inferred demographics, behavioural data) where individuals are not aware of their inclusion in the group and where those groups are ephemeral and thus distinct from those already established in the legislation would be a big step.
This raises the question and critique of whether such groups, conceived as single units distinct from the members from whom they are composed, could ever be “identifiable” or possess their own identity from an operational standpoint. However, “identity” is not the most appropriate term. It replicates the individualist approach described throughout this paper, implying stability and individual awareness, which does not fit impersonal groups or datasets. Modern automation designs groups artificially for specific goals. This design or designation is, most importantly, imposed rather than naturally arising. Therefore, groups’ identity is that identity designed and designated by an external actor for a specific purpose.
2. Reintroduce the group right to complaint in the AI Act and provide guidelines
While the draft of the AI Regulation included a redress mechanism, the adopted text has forsaken it. It is not clear whether the recognition of any potential damages ensuing from biases or discrimination and the information disclosed based on the transparency provisions provided would work out, given that there is no way to claim damages as a result of a particular decision-making result, based on perceived affinities, behaviour or interests. Once again, the regulatory approach focuses on the individual action and claim before a supervisor in order to assert rights. In addition, the claimant should prove damage, which is practically impossible when inferred groups are considered, as pointed out in the cases mentioned in the preceding Section. In practice, this means that no significant change ensues for citizens except the conclusion that they may be discriminated on the grounds of their age or gender for example. Thus, a mechanism for lodging a collective complaint, defined procedural and substantive requisites and legal recognition of this right should ensure higher legal certainty and better legal protection.
Without further progress towards group data protection, realistically, a redress mechanism under the AI Act must be circumscribed to the types of groups the proper law recognises. The blueprint has already been laid out in the past by the European Parliament, which proposed Article 68a (new), establishing a right to lodge a complaint to every natural person or “groups of natural persons” with a national supervisory authority. As already mentioned, the groups in the AI Act are rather the traditionally recognised ones: stable, legally recognised and visible, which makes the reintroduction of the group right to complaint feasible, due to the existence of unions or other interest groups, which may serve as representatives of those groups in the manner of the Article 80 GDPR.
3. Group impact assessment in the AI Act
While the EU legislation is still a long way to go in order to speak about group rights impact assessment, despite Article 27 (c) and (d) AI Act, a new Article 9a would establish a specific and focused impact assessment for groups of persons. This would be a mandatory evaluation of how an AI system’s operation affects groups (and their data protection and privacy rights) and whether it exacerbates biases, discrimination or collective harms. In this way, concrete and transparent rules would consolidate certain provisions for groups, facilitating the Regulation’s application by deployers and its interpretation by experts.
The scholarship has already argued in favour of the reintroduction of some of the provisions proposed but ultimately dropped, as described in a previous paper.Footnote 100 For example, Mantelero calls for group representatives in fundamental rights impact assessment (FRIA) as a venue for groups’ participation,Footnote 101 which was already proposed by the Parliament in Amendments 408 and 413 (Article 29 (a) new)).Footnote 102 Other authors propose a severity assessment, which includes “right-holders; their representatives, especially from groups mostly affected due to, e.g., their historical marginalisation.”Footnote 103 In all those cases, FRIA is devised for recognised and visible groups, and non-iterative (Articles 27 (2 and 4) AI Act).
I operationalise group protection from the “continuous iterative process planned and run throughout the entire lifecycle”(Article 9 (2)) of an AI system. Article 9 (9) implements the risk management system considering the “intended purpose of high-risk AI system” on underaged and “other vulnerable groups.” Given the current wave of simplification,Footnote 104 it is unlikely that a future reform would broaden this requirement to other than “high-risk systems.” Scepticism and speculation aside, risk management systems should be required to include considerations of inferred groups, as defined in this paper. Recital 65 AI Act already contains the premises but could be expanded to cover not only “identifying foreseeable misuse” or “reasonably expected” outcomes. It should also cover adverse consequences with a rebuttable presumption of indirect liability upon deployer’s evidence that the purpose behind the inference was not to infringe fundamental rights or to adversely legally or economically affect a person, identified as part of an inferred group or the group as a whole.
The previous point and this one could be introduced in the AI Act after the 2028 Evaluation and review report of the Commission, pursuant to Article 112 AI Act. More immediate remedy, although not with the same result, would be to amend accordingly the technical requirements listed in Annex IV (3) mandating that deployers should provide information on how they ensure effective redress, in particular, to affected groups as described in the Regulation.
4. Expand the “data subject” in the GDPR
In order to achieve a more comprehensive protection, a modification of the definition of “data subject” in Article 4 GDPR should include both individuals and groups. This would acknowledge that violations can occur at a group level and would provide legal recourse for collective harms arising from group-based profiling or targeting. The definition should cover not only vulnerable subjects and groups but also such groups which are traditionally not recognised because of their unawareness and instability, as discussed previously. Practically, this change would require a profound reconceptualisation, or rather catharsis, acknowledging the reality and therefore with significant consequences.
The dichotomy between anonymous/personal data is collapsing. No more is the question “who are you?” but “in which modelled group do you fit?” Personal data is not just about the person it names; it is relational and can be externally produced, false and used to make decisions that affect other people. Therefore, future reforms should adapt the concept of the data subject by broadening the criteria for identifiability by relationalising it, not relativising it.Footnote 105 This would mean personal data would be such if there is a potential to relate to the same subject or another natural person.
5. Collective transparency and rights in the GDPR
The introduction of explicit mention to groups in the GDPR would presuppose a bigger reform, which would encompass the rights and obligations enshrined therein. So, to ensure that the intended outcome of the previous modifications is achieved, several Articles such as Articles 12–14 should include a transparency mechanism that ensures individuals included in and potentially affected by group targeting are informed. For example, the most straightforward way to do it is through Article 13 (2)(f), Articles 14 (2)(g) and 15 (1)(h). Before the inclusion of a more explicit group transparency provision, those provisions could be interpreted to provide information about “the logic involved,” which could include any algorithmic groups.
6. Obligations in the GDPR
An amendment to Article 25 requiring data controllers to take specific measures to protect group data protection might be useful. Controllers should be required to assess how data processing affects both individuals and groups, especially where group profiling is used to make automated decisions or generate inferred predictions about groups.
7. Collective redress mechanism in the GDPR
Extension and development of Article 80 is needed in order to allow groups to lodge complaints with supervisory authorities in cases of group-level data violations. This would enable a representative entity to act on behalf of a group, similar to class-action lawsuits, to address collective damages such as algorithmic discrimination or large-scale data breaches. This is different from the current provision as in the proposed one the defence and compensation would be in the name of a collective entity, and not in the name of a number of individuals.
8. Group data protection impact assessment in the GDPR
The previous amendments require a complex procedure to be carried out and may be better suited to a more detailed and ambitious reform. However, the text of Article 35 (4) GDPR mandates that the European data protection supervisor should adopt a list of the types of processing operations subject to a data protection impact assessment. It was done in 2019.Footnote 106 The document acknowledges multiple of the high-risk circumstances discussed here. The list of criteria could include (inferred) groups’ impact as a criterion as well. It would include not only any potential adverse effect on age, gender, health, social or economic groups but also any other unexpected or even unintended groups, emerging from automated analysis.
VIII. System sync between shell and soul
The suggested changes are nonetheless patches, if not wishes, unless a more ambitious reform of the current data protection system is undertaken. Without that system synch, EU law will keep regulating the Shell while the Ghost keeps deciding outcomes off-ledger.
Neither consent nor the notion of groups are new concepts. AI systems routinely designs groups of individuals, often unaware, and unidentifiable, while the GDPR and the AI Act still allocate rights and remedies to identifiable subjects. The result is a structural asymmetry where injunctions can be obtained without mandates, but compensation still requires identifiability and individual standing. That gap is not accidental because it is a legacy of an individual‑centric architecture that never anticipated AI‑generated groups of people.
Just as at the very beginning of the industrial revolution, Public law could not provide remedies to negative externalities caused by industry, because private bargaining was supposed to suffice. In hindsight, it seems insufficient, if not impracticable, without a further regulatory conceptual development. Just as European data protection is today, in the face of the Digital Revolution. The law already contains the pieces to fix it. The Parliament’s draft AI Act supplied the missing parts: definition, transparency/participation and a group‑complaint right. The 2028 reviews of both the GDPR and the AI Act are the obvious moments to re-address those concerns.
Absent those steps, the EU’s data protection framework will continue to protect only the visible, recognised already individuals, while leaving the AI‑constructed ghosts, the soul behind the cyborg shell, groups of people with no realistic path to a meaningful remedy. Recognising, defining, and empowering groups is not a conceptual revolution but an overdue adjustment to how AI already processes and harms people. This is why my hope is that this study contributes to the opening of new horizons in data protection and privacy by inspiring further research and discussions.
Open access
