Abstract
The development of artificial intelligence has profoundly reshaped the ways in which personal data are generated, processed, and retained, placing intelligent systems at the heart of debates on privacy and fundamental rights. This article examines, from a European Union legal perspective, the application of the General Data Protection Regulation (GDPR) to AI and assesses whether the principles and rights enshrined in European law—particularly the rights to erasure, to be forgotten, and to rectification—can be effectively exercised once information has been absorbed by machine learning models. The study examines the main legal and technical challenges arising from the nature of AI, which does not store data in a static form but transforms it into knowledge, thereby complicating its localisation, alteration, or deletion. It also analyses the relationship between the GDPR and the Artificial Intelligence Act (AIA), emphasising their complementary roles and the need to ensure coherence between the two regulatory frameworks. From a legal and ethical standpoint, the paper considers phenomena inherent to AI systems—such as hallucinations, algorithmic bias, and neurodata—to illustrate how they challenge essential principles such as accuracy, minimisation, and purpose limitation, and how they test the rights of individuals in contexts where information cannot truly be “forgotten”. Finally, it proposes alternative mechanisms, mitigation strategies, and emerging solutions aimed at preserving individuals’ effective control over their data in the algorithmic age, thereby reinforcing privacy protection and public trust in the responsible use of new technologies.
Keywords
1. Introduction
The rapid development of generative artificial intelligence (AI) systems is transforming multiple sectors — from healthcare and education to the judiciary and public administration — as well as numerous dimensions of everyday life. This expansion is driving such profound change that new needs are emerging regarding how individual rights should be protected.
A striking example of this swift uptake is ChatGPT, developed by OpenAI. According to data published by Reuters, 1 ChatGPT reached 100 million active monthly users in January 2023, barely two months after its launch, making it the fastest-growing consumer application in history. This unprecedented rate of adoption clearly shows that technological progress far outpaces the capacity of existing legal frameworks to adapt.
Within this context of accelerated expansion and progressive regulation, generative AI poses a crucial challenge: how to ensure that technological advancement remains compatible with the effective protection of personal data rights. The development of models trained on vast volumes of personal information raises major legal and ethical questions concerning individuals’ control over — and real ability to decide on — their own data within such environments.
These concerns have also drawn international attention. UNESCO’s Recommendation on the Ethics of Artificial Intelligence 2 highlights that such systems can affect human dignity, autonomy, and fundamental rights. Similarly, the Council of Europe 3 warns of the need to establish governance frameworks that prevent systemic risks to democracy and the rule of law, reinforcing the idea that AI is not merely a technical or legal challenge, but also a matter of global and social responsibility.
In the European context, it is essential to respect the set of norms and rights that shape the legal reference framework. Article 8 of the Charter of Fundamental Rights of the European Union 4 recognises data protection as an autonomous right, subject to supervision by an independent authority. Building on this principle, the regulatory system culminates in Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data (General Data Protection Regulation, GDPR), 5 and more recently in Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act, AIA). 6
This paper poses a central question: Can artificial intelligence truly “forget”? From a European Union legal perspective, answering this requires examining to what extent the principles and rights of data protection (The fundamental principles governing the processing of personal data are set out in Article 5 of the GDPR and include, among others, the principles of lawfulness, fairness and transparency; purpose limitation; data minimisation; accuracy; storage limitation; and integrity and confidentiality, alongside the concepts of privacy by design and by default. Likewise, the rights of individuals—traditionally referred to as ARCO rights (access, rectification, cancellation and objection)—were expanded under the GDPR to include the rights to data portability and to restriction of processing, forming what legal doctrine now refers to as the ARCO+ rights.) — particularly the principle of accuracy and the rights to erasure, rectification, and to be forgotten — can be effectively applied to AI systems. The study analyses the legal and technical constraints that shape the exercise of these rights and assesses whether the current European framework provides adequate mechanisms to safeguard personal information in the algorithmic age. Ultimately, this legal reflection seeks to determine the actual capacity of individuals to exercise their rights in automated environments, and whether those rights need to be reinterpreted in light of the new technological paradigm.
The novelty of this research lies in examining, through an integrated approach, how the technical limitations inherent to generative AI affect the practical effectiveness of the data protection rights and principles recognised under the European legal framework.
2. Literature review
The legal and doctrinal literature on the ARCO+ rights constitutes a consolidated body of scholarship within the field of data protection.7,8 These rights have undergone progressive regulatory and jurisprudential development in Europe, evolving from Directive 95/46/EC to their current formulation under the GDPR. Within this framework, the right to erasure has reached its most complete expression, accompanied by its specific manifestation as the right to be forgotten. 9
The most significant judicial milestone in this regard is Google Spain SL and Google Inc. v Agencia Española de Protección de Datos (AEPD) and Mario Costeja González, 10 which recognised the right to be forgotten in the digital environment. Numerous scholars have emphasised the importance of distinguishing this right from that of erasure, despite their apparent equivalence under Article 17 and Recital 65 of the GDPR. While the former concerns the technical deletion of data within a specific processing activity, the right to be forgotten focuses on limiting the public dissemination of information, primarily through the de-indexing of search engine results.11,12
In contrast to this well-established doctrinal framework, the literature examining the interaction between these rights and AI systems remains at an early stage. Recent studies converge on the idea that machine learning models do not store individual records but rather integrate patterns derived from large volumes of data, which complicates the application of the right to erasure in its classical sense.13–15
More recent research on large language models (LLMs) indicates that the technical limitations inherent to such architectures prevent the effective erasure or rectification of personal data. 16 Within this context, partial solutions such as machine unlearning have been proposed,17,18 although the literature warns that these techniques have yet to achieve the complete removal of a data trace. 19
To date, case law has not directly addressed the application of the rights to erasure and to be forgotten in relation to AI systems, although the first proceedings are beginning to emerge before European data protection authorities.
3. Regulatory framework
Any analysis of data subjects’ rights in relation to AI systems must begin with an examination of the applicable legal framework, particularly the provisions that determine how personal data should be processed in automated environments.
In the European context, the main reference framework consists of the GDPR and the AIA. A joint analysis of these two instruments allows for a clearer understanding of the real scope of personal data protection in the age of AI and helps assess the extent to which their respective objectives are compatible in practice—especially regarding the exercise of individuals’ rights vis-à-vis automated systems.
Before undertaking such analysis, it is worth recalling that the application of these frameworks displays certain specificities within public administrations, where legal and organisational requirements introduce particular nuances that merit separate consideration (In the context of public administrations, the right to erasure cannot always be exercised on the same terms as in the private sector, as it is subject to the exceptions set out in Article 17(3) of the GDPR, as well as to legal obligations regarding data retention, archiving, and transparency. In Spain, this framework is further reinforced by specific legislation on archival management and records administration. Given its complexity and scope, a detailed analysis of this differentiated regime lies beyond the scope of the present study and is reserved for future research focusing on the public sector, where the tension between data retention and erasure acquires particular significance.).
3.1. The applicability of the GDPR to AI systems
In order to analyse the relevant legal framework, a preliminary question must be addressed: whether AI systems in fact fall within the scope of the GDPR. Only if they process personal data will the rights and obligations set out in that Regulation apply.
This issue was, for a time, a matter of controversy—particularly following the position adopted by the Hamburg Data Protection Authority (HmbBfDI) (Der Hamburgische Beauftragte für Datenschutz und Informationsfreiheit.), which argued that information used to train an AI model should not be regarded as personal data if it could not reasonably be retrieved or associated with an identifiable individual. The authority maintained that the parameters of such a model, even when trained on personal data, “should not necessarily be considered personal information.” This interpretation sought to dissociate model outputs from their original data sources, on the premise that AI systems do not “store or recall information about real individuals”. 20 It was based on the assumption that those parameters do not directly enable identification of the persons whose data were used during training.
However, in practice, both the data used for training and many of the generated outputs are undeniably linked to identified or identifiable individuals. Accepting this position would mean that AI “hallucinations”—the generation of false information about real persons—would fall outside the protection of the GDPR, leaving the rights to erasure, rectification, and restriction of processing without effective coverage and creating an unacceptable gap in legal safeguards. This dilemma has gained particular relevance in the recent case Schrems v OpenAI, 21 which questions the technical impossibility of rectifying or erasing inaccurate information generated by a language model.
At present, the controversy appears to have been largely resolved. There is now broad consensus, both among legal scholars and European supervisory authorities, that AI systems do process personal data. The European Data Protection Board, in its Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models 22 , holds that the mere fact that a model has been trained on personal data is sufficient to qualify such activity as processing within the meaning of the GDPR, unless effective anonymisation can be demonstrated.
This position is supported by recent technical research. Carlini et al. 23 demonstrated that large language models can reproduce fragments or inferences of personal information contained in their training data, even without retaining literal copies. Similarly, they concluded that generative AI can indirectly recreate personal data from statistical patterns, reinforcing the need to regard training as a form of data processing. This interpretation reflects the very nature of machine learning: models do not retain data as discrete records but incorporate statistical patterns derived from them, thereby reproducing parts of the original information indirectly.
AI also broadens the scope of what may be considered personal data. In a hyperconnected environment, virtually any human activity can be transformed into processable information—what we say, search for, buy, observe, or how long we sleep. Even our gestures, pauses in speech, or variations in breathing can become measurable variables. Beyond traditional data, AI systems are capable of inferring emotions, analysing behavioural patterns, and even anticipating decisions.
Among the emerging categories of particularly sensitive information that challenge the right to erasure are neurodata, understood as data obtained directly from brain activity or from biometric signals capable of revealing emotions, intentions, or mental states. 24 Such information, closely linked to the most intimate sphere of the individual, not only falls within the definition of personal data but also raises new ethical and legal concerns regarding cognitive autonomy and mental freedom.
For all these reasons, denying that AI systems process personal data is untenable, both technically and legally. AI systems are fuelled by human information that reflects aspects of identity, behaviour, and decision-making. To assume otherwise would be to claim that AI can learn without knowing—a contradiction that is both logical and legal.
3.2. The General Data Protection Regulation and the challenge of an AI that cannot forget
Once it is established that AI systems do process personal data, it becomes necessary to examine the legal framework governing such processing. The GDPR constitutes the cornerstone of the protection of individual rights in the digital environment and therefore serves as the essential point of departure for analysing how its principles and safeguards apply to AI.
The GDPR is structured around a set of fundamental principles, together with the rights granted to data subjects—commonly referred to as ARCO+ rights—which ensure that technology does not define individuals without affording them the ability to intervene. Although the Regulation was conceived as a technologically neutral instrument—applicable to any system processing personal data regardless of the technology employed—its application to AI raises significant challenges. Consequently, AI systems must comply with the principles, requirements, and obligations of the GDPR, although the means of ensuring their effective implementation differ substantially from the traditional environments for which the Regulation was designed.
The regulatory model of the GDPR was developed on the premise that personal data can be identified, located, modified, or erased when necessary. AI systems—particularly large language models—do not operate in this way. They do not store data in discrete records; rather, they transform information into statistical patterns and representations of knowledge, effectively diffusing the original data throughout the model.
This structural difference explains why many of the rights recognised by the GDPR—particularly those of rectification and erasure—encounter considerable obstacles when applied to systems that, by their very nature, cannot “unlearn” what has already been integrated. The technical difficulty of deleting or modifying information embedded within a model does not, however, constitute a legal exemption.
It is important to emphasise that the GDPR does not include any exception specifically applicable to AI systems. Under Article 2, any processing of personal data—irrespective of the method or technology employed—falls within its material scope. Thus, although the Regulation was not originally designed with AI in mind, its provisions apply in full. What changes is not the legal framework itself, but the technical and organisational conditions required to give it effect.
The following sections analyse how these tensions manifest in some of the most relevant principles and rights, and what implications arise from their application in a context where systems learn, generalise, and retain information in an almost irreversible manner.
3.2.1. When AI cannot forget: The limits of the General Data Protection Regulation in relation to erasure, forgetting and data truth
Among the principles and rights recognised by the GDPR, some acquire particular significance in relation to AI systems, as they reveal the gap between legal safeguards and the practical possibility of enforcing them in environments where information is learned but cannot be forgotten.
The first of these—forming the core of this study—is the right to erasure, set out in Article 17 of the GDPR. This right enables data subjects to request the deletion of their personal data under certain circumstances, for example when the data are no longer necessary for the purposes of processing or when consent has been withdrawn. Although not an absolute right—since the Regulation itself provides for exceptions, particularly for reasons of public interest or legal obligation—its restorative value is undeniable: it represents a means through which individuals may regain a measure of control over their personal information, especially in the face of persistent and opaque automated processing.
In practice, however, its application to AI systems raises substantial challenges. While in traditional systems erasure entails deleting a record or destroying a file, in generative models the information is integrated into learning structures. These systems do not store data in static form; instead, they incorporate them through patterns and correlations during the training process. As a result, deleting the original data does not remove the imprint it has left on the model or the knowledge derived from it.
This is the real challenge of the right to erasure in the age of AI: ensuring effective deletion once data have already been transformed into learning. Technical proposals such as machine unlearning seek to provide a response, but they remain experimental and cannot yet offer legal certainty that a model has genuinely “forgotten” personal information.18,19
Closely related to the right to erasure is the right to be forgotten, which, following the interpretation of the Court of Justice of the European Union in Google Spain 10 , is primarily aimed at limiting the public dissemination of personal information—for instance, by de-indexing search engine results—rather than ensuring its material deletion. In the context of artificial intelligence, this right could be understood as restricting the access to or use of certain personal data within model-generated responses, rather than their complete removal. However, this possibility encounters a fundamental technical difficulty: language models generalise and infer information, making it extremely complex to identify or isolate personal data, while any attempt to block or rectify it risks affecting the integrity of the model as a whole.
Closely linked to the rights of erasure and forgetting are the principle of accuracy (Article 5 (1) (d) GDPR) and the right to rectification (Article 16 GDPR), which ensure that personal data are accurate, up to date, and complete—preventing decisions from being based on incorrect information and allowing data subjects to correct inaccuracies where they occur. In AI systems, this principle takes on particular relevance: models not only process the data they are given but also infer, predict, and generate new information. Thus, when the underlying data contain errors or the model produces “hallucinations”, the result may be the creation of false information about a real person, violating the principle of accuracy and, in many cases, rendering the right to rectification practically unenforceable.
Ultimately, both the rights to erasure and rectification share the same dilemma: the GDPR is premised on the ability to identify, locate, and modify personal data, whereas AI models diffuse such data across learning networks that cannot be selectively undone. This mismatch between legal design and technical reality calls into question the practical effectiveness of the rights enshrined in the Regulation.
Added to this is the requirement of data protection by design and by default (Article 25 GDPR), which obliges controllers to implement protective measures from the earliest stages of system development. This means that, before deploying any model that processes personal data—including AI models—controllers must integrate the principles established by the Regulation and demonstrate that the system can effectively ensure the exercise of data subjects’ rights.
If an AI model cannot rectify, erase, or guarantee the accuracy of the information it generates, it fails to meet the minimum standards of the European data protection framework and should therefore not operate in the European market. As emphasised by the European Data Protection Board, 22 compliance with these obligations cannot depend on the technical limitations of developers, since the protection of fundamental rights constitutes a precondition—not a consequence—of the deployment of any technology.
In sum, the GDPR provides the essential framework of safeguards governing the processing of personal data, yet its application to AI exposes profound tensions between law and practice. Despite these challenges, the Regulation remains fully applicable: AI systems must comply with its obligations, even when doing so requires rethinking how those obligations can be implemented within machine learning environments.
3.3. The interaction between the AI and data protection regulatory frameworks
The Artificial Intelligence Act (AIA) constitutes the second key reference framework in this study, owing to its central role in regulating the development and use of AI systems within the European Union. Its objective is to ensure that such systems are developed, placed on the market, and used in a manner that is safe, reliable, and respectful of fundamental rights and the values of the Union.
The AIA maintains a close connection with the GDPR, as expressly recognised in Recital 10, which states that it should facilitate the exercise of rights guaranteed under Union law in the field of personal data protection. This interrelation is reinforced in Recitals 48, 67, and 69, which incorporate data protection into the notion of safety and emphasise that privacy must be safeguarded throughout the entire life cycle of an AI system.
In particular, Recital 48 underlines that the impact of an AI system on fundamental rights—including the protection of personal data—is a determining factor in classifying it as high-risk, thereby triggering enhanced obligations relating to control, traceability, documentation, and human oversight. This link between technological risk and the protection of rights confirms that, within the AIA framework, personal data protection is conceived as an essential element of both security and the ethical governance of AI.
The AIA does not replace or modify the GDPR; rather, it reinforces it from a complementary perspective, consolidating the notion that compliance with data protection principles and rights is indispensable to fostering trust in AI and ensuring its compatibility with the European model of fundamental rights. Consequently, AI is not exempt from compliance with the GDPR: all principles and rights recognised therein—including the ARCO+ rights—apply fully to AI systems insofar as they involve the processing of personal data. The AIA does not redefine these rights; it presupposes and requires their observance as a fundamental condition for the responsible development of the technology.
Notwithstanding their coherence in matters of data protection, the interaction between the GDPR and the AIA may give rise to certain practical challenges, particularly regarding the exercise of data subjects’ rights in environments involving multiple actors and automated systems. One area requiring special attention concerns accountability and governance. The GDPR establishes a structure based on the roles of controller and processor, whereas the AIA introduces additional actors—such as providers, importers, distributors, and deployers—each bearing specific obligations throughout the AI system’s life cycle. This expansion does not necessarily lead to a dilution of responsibility, but it does introduce a more complex chain of shared obligations, which may hinder the effective exercise of rights such as erasure or rectification, especially when coordination among different technological agents is required. The real challenge lies not so much in determining who should be held accountable, but in ensuring that the deletion or correction of personal data can be effectively realised across the entire AI ecosystem—a task that demands sufficient technical traceability and genuine cooperation among all actors involved.
Ultimately, the GDPR and the AIA must be understood as complementary rather than competing legal frameworks, designed to operate jointly in the protection of fundamental rights. However, for this relationship to be effective, technology must be capable of translating legal principles into real and verifiable technical solutions.
4. The ethical dimension of AI: Bias, hallucinations, and the protection of personal rights
Control over personal data—embodied in rights such as erasure, rectification, and the right to be forgotten—is not merely a technical or legal matter; it is, fundamentally, a practical expression of dignity, truth, and personal autonomy. In the context of AI, these rights represent the individual’s capacity to maintain control over their personal information in environments where systems learn, infer, and make decisions based upon it.
Both the GDPR and the AIA go beyond the imposition of legal obligations: they translate fundamental ethical values into enforceable norms that seek to balance technological progress with human rights. They are therefore not external to ethics, but rather its legal realisation within the digital domain.
As Floridi 25 observes, the ethical challenge of artificial intelligence lies in ensuring that technology develops for humanity, not at its expense. From this perspective, the law—through instruments such as the GDPR and the AIA—transforms foundational values into concrete legal duties, enabling technological innovation to advance without endangering human dignity or individual freedom. When these principles are violated, it is not only the law that is breached; public trust in technology and in the institutions that regulate it is also eroded.
Thus, ethics and law converge towards a shared purpose: to ensure that AI systems uphold the values that safeguard human dignity. Ethical tensions emerge precisely when these principles and rights cease to be fulfilled or lose their practical meaning in environments where the knowledge generated by AI cannot easily be “forgotten.”
4.1. Imperfect AI: Ethical and legal challenges of forgetting in the algorithmic age
Having established the legal framework guiding the responsible development of AI, this section examines how those principles are tested in practice, with particular attention to the challenges surrounding the effective implementation of the rights to accuracy, erasure, and forgetting. The interaction between AI and personal data is not confined to a theoretical plane; it manifests through various phenomena and data typologies that expose the technical and regulatory limitations of ensuring the full exercise of data subjects’ rights and maintaining public trust in technology. By way of illustration, three particularly significant examples are considered: hallucinations, algorithmic bias, and neurodata. Without attempting an exhaustive analysis, this section offers an overview of the dilemmas posed by these situations in a context where AI not only processes information but also interprets, infers, and transforms it.
4.1.1. Hallucinations
Hallucinations represent one of the most problematic behaviours of generative language models. They occur when AI “imagines” what never existed, generating false yet plausible information. These “inventions” do not derive from stored data but from erroneous statistical inferences presented by the model as factual. An illustrative example is ChatGPT citing non-existent judicial decisions or attributing statements to real individuals 26 —a phenomenon also documented in academic research showing that large models tend to fill gaps in their knowledge with plausible fabrications. 27 As Christakis 28 notes, such hallucinations are not mere technical faults but legally relevant events, as they can generate and disseminate false information about real persons, calling into question the effectiveness of data subject rights and evidencing the need to adapt their exercise to the technical limitations of AI.
From the perspective of the GDPR, these behaviours violate the principle of accuracy (Article 5 (1) (d)) and undermine the rights to rectification (Article 16) and erasure (Article 17), since in many cases data subjects cannot—although they should be able to—delete or correct information that the system has inferred or fabricated. The Schrems v OpenAI 21 case reflects this tension: the system generated inaccurate personal data about the claimant and refused to erase or rectify them. Such errors reveal that AI not only remembers what it has learned but sometimes imagines what never existed—and that although the legal framework provides mechanisms to guarantee accuracy and remedy, their effective implementation is particularly complex in systems that were never designed to forget.
4.1.2. Algorithmic bias
Algorithmic bias constitutes one of the greatest challenges to the effective protection of individual rights in AI systems. Bias arises when the data used to train a model are not neutral but reflect existing prejudices, inequalities, or behavioural patterns in society. AI learns from such data and, in doing so, may reproduce or amplify discrimination—even unintentionally. These biases can manifest across diverse domains—from recruitment and credit scoring to law enforcement—disproportionately affecting certain social groups. A paradigmatic example is ShotSpotter, an AI-based gunshot detection system used in the United States, which was criticised for concentrating surveillance in predominantly African-American neighbourhoods and producing false positives that led to wrongful arrests.
From the GDPR perspective, bias directly affects the principles of accuracy, fairness, and data minimisation, and calls into question the validity of personal inferences that may be erroneous or harmful. It also impacts the rights to rectification and erasure, as individuals should be able to demand the correction or deletion of inaccurate conclusions about them. The problem, however, is that unlike an explicit data point, a bias cannot easily be “deleted”: it is embedded within the model’s patterns and correlations.
Hence, the right to erasure or rectification in such systems must be understood not merely as the deletion of data, but as the capacity to unlearn unfair inferences. If a system falsely associates a person with a risk, a profile, or a behaviour, that statistical imprint should also be capable of being erased. Otherwise, the model retains a form of discriminatory memory that distorts truth and erodes equality.
Even when a bias reflects a “real” correlation, it is unethical to turn that correlation into a decision criterion. Statistical truth does not equate to justice: AI should not replicate the inequalities it encounters, but contribute to correcting them.
4.1.3. Neurodata
Among the various ethical dilemmas raised by AI in relation to the rights recognised under the GDPR, neurodata arguably represent the most sensitive frontier. Such data hold enormous potential in fields such as medical research, cognitive neuroscience, and the treatment of neurodegenerative diseases; however, they also entail unprecedented risks, as they open the possibility of accessing, collecting, sharing, and manipulating information derived from the human brain 24 .
If guaranteeing the right to erasure is already complex in relation to traditional personal data, the challenge becomes even greater when what is at stake are representations of the human mind itself. The prospect of deleting information that AI has learned from brainwaves, emotions, or thoughts raises novel questions about the limits of individual control. Neurodata are not merely an advanced privacy issue; they constitute a direct challenge to the principles of data minimisation, purpose limitation, and explicit consent—the very pillars of the GDPR and an expression of respect for human dignity.
Freedom of thought—that intimate sphere which John Milton championed in the seventeenth century when he wrote, “Thou canst not touch the freedom of my mind” 29 —remains the foundation of personal autonomy and critical reasoning, and must never be subjected to algorithmic training logics, commercial interests, or decisions made without the individual’s free, unambiguous, and informed consent. Protecting neurodata therefore means not only preventing their misuse, but ensuring that individuals retain control over them: the ability to decide, to rectify, or to erase the information that emerges from their own minds. Ultimately, the right to erasure here acquires its deepest meaning: safeguarding the individual’s sovereignty over their cognitive identity, ensuring that knowledge derived from the human mind does not become an external commodity but remains an extension of personal freedom.
In sum, the ethical challenges associated with AI transcend the technical sphere and reach the very core of human rights and freedoms. Collectively, they reinforce the urgency of consolidating safeguards that ensure individuals retain effective control over their personal information and of developing alternative mechanisms that make the practical exercise of rights—such as erasure and rectification—truly possible.
5. The Schrems case and the challenges of the right to Be forgotten in the age of AI
As has been anticipated throughout this study, the difficulties examined are not merely theoretical but have an evident practical translation. A paradigmatic example is Schrems v OpenAI, 21 currently pending before the Austrian Data Protection Authority. This case confronts a generative AI system—specifically, ChatGPT, one of the most widespread large language models (LLMs)—with the obligations arising under the GDPR. Its analysis is of particular relevance, as it illustrates in practice the limits of the European regulatory framework when applied to technologies that, by their very nature, learn, infer, and reproduce personal information in ways that are not always controllable. It therefore exposes structural challenges to the effective application of the rights to erasure, rectification, and forgetting, which lie at the heart of this paper.
The proceedings were initiated by Max Schrems, an Austrian lawyer and privacy activist known for his role in the landmark cases Schrems I 30 and Schrems II, 31 which led to the annulment of the Safe Harbour and Privacy Shield data transfer frameworks between the European Union and the United States. A central figure in the defence of privacy in Europe, Schrems’ career reflects a consistent conviction: that data protection law must not remain a theoretical construct, but be enforced effectively against major global technology actors.
In the 2024 case, the complaint was lodged by Schrems and the NGO NOYB (
This response amounts to an admission that the system does not allow for the effective exercise of rights recognised under the GDPR—particularly the rights of access (Article 15), rectification (Article 16), and erasure (Article 17)—as well as the principles of accuracy, transparency, and accountability (Article 5). It also reveals non-compliance with the principle of data protection by design and by default (Article 25 GDPR), since the system was deployed in the European market without a prior impact assessment or adequate analysis of risks to individuals’ rights.
In legal terms, the Schrems case demonstrates that OpenAI has acknowledged the technical impossibility of meeting its obligations under the GDPR while nevertheless maintaining the system’s operation in the European market. This situation exposes a fundamental contradiction: if a technology cannot guarantee the fundamental rights enshrined in existing law, its commercialisation or deployment should be legally restricted until compliance can be ensured. The issue here is not a regulatory vacuum, but rather a failure to enforce the framework already in place.
From an ethical standpoint, the problem extends beyond formal compliance. What this case demonstrates is that the failure to uphold data protection principles in AI models is not a technical deficiency but a violation of basic human rights—namely dignity, informational autonomy, and control over one’s digital identity. Ultimately, Schrems v OpenAI stands as a symbol of the current imbalance between the accelerated development of AI and the law’s capacity to ensure that individuals retain real power over their data.
The case has also reignited debate on how AI-generated and inferred data should be treated. In this regard, Häuselmann 32 provides a crucial contribution by analysing the right to rectification of such data—those not provided directly by the data subject but generated or deduced algorithmically. According to his argument, these data should be considered personal whenever they relate to an identified or identifiable person, meaning that erroneous inferences may be subject to rectification or erasure under Article 16 GDPR. Applied to the Schrems case, this implies that it is not sufficient to delete an entry or block a data point within the system: effective correction requires neutralising the inaccurate information learned by the model, through verifiable review mechanisms or editing of the embedded knowledge.
Taken together, these doctrinal reflections position Schrems v OpenAI as a highly significant legal and symbolic precedent. It not only tests the possibility of rectifying data generated or inferred by an AI model but also questions the real viability of the right to erasure—or “to be forgotten”—within systems that learn from personal data but cannot easily unlearn them. Moreover, it underscores the urgent need to establish control and oversight mechanisms capable of translating the GDPR’s guarantees into effective solutions to the technical challenges posed by AI.
As with the landmark Google Spain10 decision, this case may mark a turning point in the interpretation of the GDPR’s principles and of the rights to rectification, erasure, and forgetting in the context of generative AI. Its outcome will carry substantial legal and symbolic weight, potentially laying the groundwork for future European jurisprudence on digital rights and data protection in relation to AI—consolidating the principle that no technology stands above the rights of individuals.
6. Mechanisms and strategies for protecting privacy in the age of AI
The challenges analysed throughout this paper—as illustrated by the Schrems v OpenAI case 21 —lead to a clear conclusion: the current legal framework still faces serious obstacles in guaranteeing the effective enforcement of individuals’ rights against AI systems. These difficulties stem not so much from gaps in regulation as from the technical and structural limitations inherent to machine learning models, which in many cases prevent the correction or erasure of personal data without compromising the global knowledge acquired.
The GDPR contains no technological exemptions from compliance with its principles and obligations, and its full application constitutes an essential condition for any system to operate lawfully within the European sphere. Nonetheless, while technology does not yet provide complete solutions, it is vital to advance complementary measures that mitigate risks and allow individuals to maintain a degree of control over their privacy. These initiatives do not replace the rights established by law but reinforce them, while institutions, controllers, and developers adapt their systems to comply effectively with the guarantees imposed by the European legal order.
Among the most relevant technical and organisational measures, the European Data Protection Board 22 identifies anonymisation, pseudonymisation, blocking, and restriction of processing, as well as impact assessments, as key tools to strengthen control over personal information and to prevent misuse. The Board nonetheless emphasises that their effectiveness depends on their implementation from the design phase of a system and on ensuring the impossibility of re-identification or unauthorised reuse of data.
In the context of generative AI, however, such safeguards are palliative rather than structural solutions and face significant limitations: the very architecture of machine learning models makes it difficult to isolate or selectively remove a single data point without altering the model’s overall knowledge base. This tension—clearly demonstrated in Schrems v OpenAI 21 —reveals a practical contradiction with the Regulation itself, which precisely requires the possibility of granular intervention in personal data.
Some recent approaches draw inspiration from the logic of the right to be forgotten, seeking to limit the dissemination of personal information without necessarily deleting its original source. Applied to AI, this would mean preventing models from reproducing certain personal data in their outputs, even if such data were used during training. However, this restriction remains technically insufficient: blocking one data element without affecting the rest of the learned knowledge is still unfeasible, confirming the need to move towards deeper structural solutions.
Among these emerging avenues, machine unlearning stands out as a promising technical approach. It seeks to enable a model to “unlearn” previously incorporated information without requiring complete retraining. Although still at an early stage and not yet guaranteeing verifiable deletion in all cases, it represents a promising line of research for making the exercise of the right to erasure and data protection by design more feasible. As noted by Bourtoule et al. 18 and Ginart et al., 33 this technique could, in the future, allow for the selective neutralisation of undesired learnings in deployed models—bringing technical practice closer to the requirements of Article 17 GDPR.
In any event, these strategies—whether palliative or emergent—do not exempt controllers or developers from full compliance with the obligations laid down in the GDPR. Their value lies in serving as a bridge towards a scenario of genuine compatibility between technology and law, where authorities, controllers, and technological actors can effectively guarantee erasure, rectification, and authentic control over personal data in the age of AI.
6.1. Digital literacy and shared responsibility in the age of AI
Given the practical difficulties of exercising rights such as rectification, erasure, or forgetting in relation to large language models (LLMs), the effective protection of privacy requires progress on a complementary front: digital literacy and education as a means of fostering social awareness. In many cases, the best way to guarantee the right to erasure is not to delete data, but to prevent it from ever being generated or disseminated. Digital education thus becomes the first line of defence in data protection. Ultimately, the safest erasure is the one that is never needed.
In reality, even individuals considered “digitally literate” often share personal data through online applications and services without understanding the full extent of its processing—driven by immediate incentives such as discounts, convenience, or entertainment. This lack of awareness sustains an ecosystem in which personal information becomes a form of currency, often with irreversible consequences. 2
In this context, both public and private institutions play a decisive role—not only in ensuring compliance with the GDPR but also in promoting training programmes that strengthen a culture of privacy and critical thinking in the use of AI-based technologies. Likewise, developers and those responsible for new European AI models must assume their share of ethical responsibility, incorporating from the design stage the principles of transparency, explainability, and minimisation.
European legislation already requires that these values be integrated at every stage of technological development. However, as this study has shown, compliance remains far from universal. Therefore, while progress continues towards stronger legal and technical solutions, education and awareness become the most powerful tools to preserve individuals’ informational autonomy and to prevent technological innovation from eroding the rights that underpin human dignity.
If AI cannot yet forget, we must learn to remember our responsibilities. This calls for a cultural shift: understanding that every interaction with technology leaves a trace, and that the protection of fundamental rights depends not solely on rules or algorithms but on the ethical commitment of those who design, regulate, and use them.
These reflections do not replace the obligations imposed by law, but they serve as a reminder that the ethical use of AI begins with individual and collective awareness—acting with responsibility, prudence, and respect for human dignity—within a framework of shared social commitment among all actors, including users themselves.
7. Conclusions
Generative artificial intelligence has tested the very foundations of the European model of data protection, revealing that today’s challenges do not stem from a regulatory vacuum but from the gap between law and technology. The GDPR and the AIA together form a coherent and robust framework, yet their effectiveness ultimately depends on whether systems are designed and monitored in accordance with their underlying principles.
The problem does not lie in the existence of rights, but in their practical effectiveness. Large language models demonstrate that the principle of accuracy and the rights to rectification, erasure, and forgetting are constrained by technical architectures that were never designed to forget. A persistent tension emerges between what the law requires and what technology allows—and that very tension transforms technical limitations into genuine legal challenges.
The Schrems v OpenAI case 21 illustrates this vividly: a system capable of generating false information about a person continues to operate without offering effective mechanisms for correction or deletion, exposing a profound gap between the legal framework and technological reality. This gap cannot be bridged by relaxing legal requirements, but rather by strengthening accountability and oversight.
In this context, the role of data protection authorities is decisive. A troubling imbalance remains between real risk and supervisory focus: while resources are often devoted to sanctioning minor breaches or incidents of limited social relevance, the structural challenges posed by AI systems—with potential impact on millions of citizens—frequently go unanswered. The logic of data protection cannot centre on managing anecdotal irregularities; it must instead address the systems that define the very architecture of information and knowledge in the digital age.
Reorienting institutional action towards these challenges is not merely a matter of efficiency but also of justice. If authorities concentrate their efforts on the peripheral rather than the structural, there is a risk of consolidating a formalistic model of data protection—one disconnected from technological reality and the needs of society. Europe requires a forward-looking approach to data protection, one that strengthens its capacity to respond to the risks generated by innovation.
From a European standpoint, and with a view to future regulatory developments, the analysis undertaken makes it possible to anticipate that the interaction between artificial intelligence and data protection will continue to occupy a central place on the European legal agenda. The practical difficulties observed in applying the GDPR to AI systems have, in recent months, prompted an intense doctrinal and institutional debate concerning the adequacy of the existing framework for technologies that were not envisaged at the time of its adoption. In this context, the very design of the GDPR—characterised by broadly framed principles and by the absence of an exhaustive regulation of specific technical measures—renders it a potentially suitable instrument for adaptation to emerging technological realities, allowing artificial intelligence to be accommodated within its normative structure. This reflective process is likely to result in amendments to the Regulation itself, as well as in reinterpretations and the development of complementary regulatory instruments, the practical articulation of which will make it possible to assess whether such flexibility effectively contributes to the preservation of fundamental rights protection or, conversely, leads to a gradual relaxation of the safeguards underpinning the European model.
From a broader perspective, the challenge should be understood not so much as a need to dilute legal requirements or to patch technical shortcomings, but rather as the imperative to ensure that machine-generated knowledge does not erode the human capacity to decide what should be preserved and what should be forgotten. The question, therefore, is not whether fundamental rights can be adapted to artificial intelligence, but whether artificial intelligence can be developed in a manner that respects those rights. The law must learn to engage in dialogue with technology and to inspire solutions that make forgetting possible where this is recognised and required by the legal order.
AI should not possess more memory than the limits imposed by law and ethics. The challenge is not to teach machines how to forget, but to ensure they remember only what human dignity can bear.
Footnotes
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.