AI Training Data and GDPR Deletion Rights: Practical Compliance When Data Has Already Been Trained

Somewhere in the weights of a large language model, there may be traces of personal information about millions of people. They did not consent to being part of a training dataset. They cannot easily identify which information was collected. And if they submit a deletion request under GDPR’s Article 17 right to erasure, neither they nor the company they are asking can be certain whether compliance is technically possible — or what compliance even means when the personal information in question is not a row in a database but an infinitesimally small influence distributed across billions of numerical parameters.

This is the machine unlearning problem, and it represents one of the most technically challenging and legally unsettled compliance issues facing AI companies operating in Europe or serving European users. This post addresses what GDPR’s right to erasure actually requires in the context of AI training data, what technical approaches currently exist and their limitations, how US state laws have extended and complicated deletion obligations, and what practical approaches regulators currently accept.


What GDPR Article 17 Actually Says

Article 17 of the General Data Protection Regulation grants individuals — “data subjects” in GDPR terminology — the right to obtain from a data controller the erasure of personal data concerning them without undue delay. Several grounds trigger the right, but the most relevant in the AI training data context are: (1) the personal data are no longer necessary in relation to the purposes for which they were collected or otherwise processed; (2) the data subject withdraws consent on which the processing was based and there is no other legal ground for processing; and (3) the data subject objects to processing pursuant to Article 21 and there are no overriding legitimate grounds for the controller’s processing.

GDPR Article 17 also identifies exceptions. Erasure is not required when processing is necessary for compliance with a legal obligation, for reasons of public interest, for archiving purposes in the public interest, scientific research, or historical research purposes, or for the establishment, exercise, or defense of legal claims.

The fundamental problem is that Article 17 was written in the context of structured databases and defined record-keeping systems. Erasure in that context is conceptually clear: locate the record, delete it, confirm deletion. In the context of AI models, none of those steps translates directly.


Why AI Training Data Is Different

When personal information is used to train a machine learning model, it is not stored in the model in a retrievable form. The training process uses the data to adjust the model’s weights — billions or trillions of numerical parameters — through a process of repeated optimization. When training is complete, the training data itself is (typically) not retained in the model. What remains is the cumulative effect of the training data on those weights, distributed across the entire model in a way that cannot be reversed to isolate and remove the influence of any particular data point.

The European Data Protection Board (EDPB) addressed this directly in Opinion 28/2024 (December 2024), its landmark guidance on personal data processing in AI models. The EDPB made several key findings that reject arguments AI developers have sometimes made:

AI models are not automatically anonymous. Some developers have argued that because training data is not stored within the model, the model itself does not contain personal data and therefore GDPR does not apply to the trained model. The EDPB rejected this position. The EDPB’s view is that a model can retain personal information from training data — for example, when a model can reproduce specific information about identifiable individuals — and that the mere absence of record-level storage does not render a model anonymous under GDPR’s definition of anonymization.

AI developers can be data controllers. The EDPB confirmed that organizations that train AI models on personal data are data controllers under GDPR, subject to the full range of data subject rights including the right to erasure.

The question of whether a model contains personal data is a factual inquiry. Not every model trained on data that includes personal information will itself contain personal data in a legally meaningful sense. The EDPB suggested that factors including whether the model can reproduce identifiable information on request, whether it was fine-tuned on personal data in ways that allow reproduction, and the nature and volume of the training data influence the analysis.

The practical upshot is that GDPR compliance obligations — including the right to erasure — can attach to trained AI models, not just to the training datasets themselves.


The Technical Challenge: Machine Unlearning

The technical research community has worked to develop methods for “machine unlearning” — procedures by which a trained model can be made to “forget” the influence of specific training data without retraining the model from scratch. The challenge is substantial.

Exact unlearning requires full retraining. The only method guaranteed to completely remove the influence of a specific data point is to retrain the model from scratch on a training dataset that excludes that data point. For large foundation models, full retraining is extremely expensive — the pure energy costs for training a major large language model can run into several million dollars — and is operationally impractical as a response to individual deletion requests.

Approximate unlearning methods exist but have limitations. Several approximate unlearning methods have been developed that modify model weights to reduce the influence of target data without full retraining. These include gradient-based methods (which reverse the training update associated with a particular data point), influence function approaches, and data poisoning-based methods. A notable advance in September 2025 came from researchers who proposed “source-free unlearning,” which operates without the original source data by using a surrogate dataset to guide model modification, achieving performance comparable to full retraining at a fraction of the computational cost.

However, approximate unlearning methods raise a fundamental legal question: whether the reduction in influence they achieve constitutes “erasure” within the meaning of Article 17. The EDPB has not definitively answered this question. The concern is that approximate methods may reduce but not eliminate the influence of training data, and it is not clear that a partial reduction satisfies a legal standard of erasure.

Membership inference attacks complicate verification. Even after unlearning procedures are applied, adversarial techniques called “membership inference attacks” can sometimes determine whether a particular data point was part of the original training set. If these attacks can identify the data point’s influence post-unlearning, the unlearning was arguably incomplete. The practical standard for what “completeness” requires is not yet settled by regulators or courts.


What Regulators Currently Accept: The Practical Compliance Landscape

Given the technical impossibility of exact unlearning at scale, regulators — primarily European data protection authorities (DPAs) — have taken pragmatic positions while pushing companies toward better practices. As of 2025 and 2026, the following practical approaches are generally understood to be relevant to a defensible compliance posture, though no formal regulatory safe harbor has been established:

Pre-training data hygiene is the first line of defense. The most defensible position is to avoid having personal data in the training dataset to begin with. Systematic scrubbing of training data to remove or pseudonymize personal information before training begins — or filtering data sources to exclude categories of personal information that are likely to create identifiability risks — reduces both the training data deletion burden and the model reproduction risk. The EDPB has indicated that appropriate technical and organizational measures applied to training data before training can support a determination that the resulting model does not contain personal data in a legally meaningful sense.

Lawful basis and purpose limitation at the training data collection stage. GDPR requires that personal data be collected on a lawful basis and used only for specified, explicit, and legitimate purposes. AI training is only a lawful purpose if either consent was obtained for that purpose, or a legitimate interest or other basis applies and was properly assessed before collection. Many AI developers have trained on data scraped from the internet under a claimed legitimate interests basis; the validity of that basis remains disputed and varies by jurisdiction and the nature of the data involved. Companies with clear, documented lawful bases and purpose limitation at the point of collection are better positioned to respond to deletion requests by demonstrating either that the right to erasure does not apply (because processing falls within an exemption) or that the data was collected for a purpose that now no longer exists.

Differential privacy techniques at training time. Differential privacy is a mathematical framework that introduces calibrated noise into the training process in a way that provides provable bounds on the amount of information any individual’s data contributes to the trained model. Models trained with sufficiently strong differential privacy guarantees can support a technical argument that no individual’s personal data is meaningfully present in the model — though the strength of the privacy guarantee trades off against model performance. Regulators have indicated that differential privacy is a relevant technical measure, even if they have not specified required parameters.

Deletion of training datasets and suppression of model outputs. Even if the trained model cannot be modified to fully unlearn specific training data, responding meaningfully to a deletion request may include: (1) confirming deletion of the personal data from any retained copy of the training dataset; (2) suppressing or filtering the model’s output to prevent reproduction of identifiable personal information associated with the requesting individual; and (3) documenting these steps in a written response to the data subject. While this approach does not constitute full erasure under GDPR, regulators have in practice focused enforcement on companies that take no steps in response to deletion requests rather than on companies that document reasonable technical measures.

EDPB Coordinated Enforcement Framework (2025). In March 2025, the EDPB launched its annual Coordinated Enforcement Framework (CEF) action specifically focused on the right to erasure. Thirty data protection authorities across Europe plus the European Data Protection Supervisor contacted controllers across sectors, opened investigations, and conducted fact-finding on how organizations handle deletion requests. This enforcement action signals that regulators are actively investigating the gap between formal GDPR rights and actual organizational compliance — and that AI companies are not exempt from scrutiny.


US State Law Extensions of Deletion Rights to AI Training Data

The European framework is the most developed, but US state privacy laws have also created deletion rights that can apply to AI training data, and several states have gone further than GDPR in addressing the AI-specific context explicitly.

California: CCPA/CPRA deletion rights and ADMT regulations. Under the California Consumer Privacy Act (CCPA) as amended by the California Privacy Rights Act (CPRA), California residents have the right to request deletion of their personal information from businesses that collect it, including information used to train AI models. The right applies to the information collected, not to the trained model per se — but if the business retains a copy of the training dataset that includes a consumer’s personal information, deletion from that dataset is required.

Going further, California’s Generative AI Training Data Transparency Act (AB 2013), effective January 1, 2026, requires publicly accessible generative AI systems trained on personal information to publish disclosures describing the categories of training data used. This transparency obligation does not directly create a deletion right against trained models, but it increases accountability and creates a factual record that can support deletion claims.

California’s finalized ADMT regulations, effective January 1, 2026 for risk assessments and January 1, 2027 for consumer rights, require businesses using automated decision-making technology to make significant decisions about consumers to provide notice, an opt-out right, and in some cases a right to human review. While not a deletion right, these requirements create compliance interdependencies with data retention practices.

Colorado’s revised AI Act. Colorado’s Governor signed an amended AI Act into law on May 14, 2026, effective January 1, 2027. The Act requires deployers of automated decision-making technology used in consequential decisions to provide consumers with notice, an explanation of adverse outcomes, and information about consumer rights. Developers must furnish documentation about systems’ intended uses, data inputs, and known limitations.

Texas and other states. Several other states — including Texas, Virginia, and Connecticut — have enacted comprehensive privacy laws with deletion rights that can apply to personal data used in AI training. Texas, notably, does not have an exemption structure identical to GDPR’s, and the applicability of state-law deletion requests to trained model weights involves legal questions that remain unresolved.


Practical Compliance Framework for AI Companies

Given the technical and legal complexity, what should AI companies actually do? The following framework reflects the current state of regulatory expectations:

Map your training data. Maintain a record of every dataset used to train each model, including data sources, categories of personal information present, and the legal basis under which the data was collected. This inventory is the starting point for any deletion request and for demonstrating GDPR Article 30 compliance (records of processing activities).

Establish a deletion request process. Designate responsibility for receiving, tracking, and responding to deletion requests. The response to a deletion request should document: what training datasets containing the requestor’s data exist, whether those datasets have been deleted or the data removed; what, if any, technical measures were applied to the trained model; and why any applicable exception (research, legal claims, etc.) applies if erasure of the trained model is not being performed.

Invest in pre-training data hygiene. For future training runs, invest in systematic scrubbing of personal information from training datasets before training begins. Document the procedures used. Technical measures taken before training are generally more defensible and practically more effective than unlearning measures applied after the fact.

Evaluate differential privacy feasibility. For applications where model accuracy tolerances permit, evaluate whether differential privacy can be incorporated into the training process. The privacy-utility tradeoff is real, but differential privacy provides the strongest available technical basis for arguing that a trained model does not contain individuals’ personal data.

Document technical impossibility where applicable. Where a deletion request cannot be fully satisfied technically — for example, because the requested unlearning from a large foundation model would require retraining at prohibitive cost — document that fact in the response and explain what alternative measures are being taken. Regulators have been more focused on companies that take no responsive steps than on companies that honestly document technical limitations and take reasonable partial measures.

Monitor the evolving guidance. The EDPB, national data protection authorities, and US state agencies are actively developing more specific guidance on AI training data and deletion rights. The regulatory framework in this area is moving quickly and companies should have compliance counsel monitoring developments.


Conclusion

The tension between GDPR’s right to erasure and the technical reality of AI model training is real, and it will not be resolved by pretending the tension does not exist. The EDPB has confirmed that AI developers are data controllers, that trained models can contain personal data, and that deletion rights apply. The technical community is developing better unlearning methods. Regulators are conducting active enforcement.

For AI companies, the most defensible position is one built on pre-training data hygiene, documented lawful bases, systematic deletion request processes, and honest communication with regulators about technical limitations. The companies that face the most regulatory exposure are those that treat deletion requests as inapplicable to AI models as a categorical matter — a position that is increasingly untenable under current regulatory guidance.


This post is for general informational purposes only and does not constitute legal advice. Reading this post does not create an attorney-client relationship. If you have questions about your specific situation, consult a qualified attorney.



Leave a Reply