Blog

Irish DPC warns personal data can be hard to remove from an AI model after training

The Irish DPC says personal data is hard to remove from AI model parameters after training. What does that mean for erasure requests and retraining?

· By

A thin open folder with loose sheets beside a thick, tightly bound closed folder on a wooden desk in daylight.
The Irish DPC notes that personal data is hard to remove from a trained AI model, unlike from the source dataset.Image: IamVera.ai — original editorial illustration

Not automatically. The Irish Data Protection Commission states that personal data is hard to remove from an AI model after training, but retraining is only required when Article 17 GDPR applies and weaker measures such as source removal, output filtering or blocking do not eliminate the risk. You must be able to demonstrate in advance which measure is effective.

In its guidance on AI, large language models and data protection, the Irish Data Protection Commission (DPC) describes a practical problem: once processed in a model, personal data can be hard to remove from the model. This puts the exercise of GDPR rights, including the right to erasure, under pressure.

For controllers, this shifts the core question. It is no longer only about whether you can remove a record from a data store, but about whether you can demonstrate that the information concerned also does not resurface from the trained model or the model output. See our broader overview of AI privacy and GDPR questions for the context in which this plays out.

What exactly does the Irish Data Protection Commission state about erasure in trained AI models?

The DPC warns that personal data can be hard to remove from a model after model training. Its guidance says that the ability to uphold rights such as access and erasure effectively should be considered at the earliest stage of development, which supports putting procedures in place to handle such requests. In its guidance, the regulator stresses that developers must consider transparency, erasure and storage limitation already at the design stage, rather than looking for an erasure mechanism only after training.

The DPC also welcomed the EDPB opinion on personal data in the development and deployment of AI models, in which the different processing phases are assessed separately. This means that training, deployment and the handling of data subjects' rights must be assessed as distinct processing phases, with appropriate attention to the legal basis and safeguards for each phase.

Why is removing data from the training data not the same as removing it from the model?

Because there are three different actions at play that are often conflated. Removing data from a source dataset is something other than preventing that data from being used again, and that is in turn something other than demonstrably removing or rendering unusable information that is already in the trained model.

The distinction the DPC draws between the source set and the trained model has a direct consequence. In our assessment, this creates an accountability challenge that goes beyond a cleaned-up database: controllers should assess and document whether personal data may persist through memorisation or model output, and whether the selected measure gives effective effect to the request. The three actions can be separated as follows:

  • Source removal — deleting data from the dataset used for training.
  • Preventing reuse — ensuring that deleted data does not end up in a subsequent training round again.
  • Removal from the model — demonstrating that the information is no longer reproducible from parameters or output, whether through filtering, blocking, model adjustment or retraining.

That last point touches on the question of why token IDs in a language model can be personal data and when training data falls outside the GDPR after the EDPB test of whether training data is truly anonymous.

When does retraining an AI model become legally necessary?

Retraining is not automatic. The right to erasure under Article 17 of the GDPR is not an absolute right: it applies only under conditions, such as where the data is no longer needed, consent has been withdrawn without another legal basis, the processing was unlawful or an objection prevails. So it must first be established that erasure is due.

Our reading of the sources: once the Article 17 threshold is met, the controller should assess whether personal data may still persist in parameters or output and whether available alternatives effectively address the request. Depending on that assessment, retraining or model adjustment may come into view. A defensible decision requires a documented weighing of, among other things:

  • the model type and the category of personal data;
  • the memorisation and extraction risk, in other words the chance that the model returns the data verbatim;
  • what is still visible in the model output;
  • which technical removal options exist and how effective they are;
  • the proportionality of retraining against alternatives such as output filtering or blocking.

What does this mean for directors, lawyers and CISOs who deploy AI?

Our analysis: because personal data can persist in model parameters after training and the DPC distinguishes source removal from removal from the model, a controller can no longer dispose of an erasure request with a cleaned-up database alone; it is therefore advisable to record for each AI system which technical measure actually carries out a request and how that is validated. Because the DPC stresses that developers must consider the rights process already at the design stage, an organisation that sets up this process only after training runs the risk that it cannot, in practice, carry out a due erasure. Our analysis is therefore that the board can, when procuring or developing a model, include as a condition that the supplier contractually describes or demonstrates the removal route and a validation test. Because training, deployment and the handling of rights must be assessed as distinct processing phases, our analysis is that organisations should make the model identity, the version of the training dataset and the scope of each request separately traceable, in so far as this is necessary to be able to account for the decision-making and the exercise of rights. Because Article 17 is not an absolute right, a lawyer who automatically treats every request as a retraining obligation gets needlessly bogged down in costs and proportionality discussions; the lawyer should therefore first test the legal basis for each request and document that judgement in writing before a technical measure is chosen.

In short, a sensible organisation now does the following:

  1. Record for each model which personal data was used and how a request is traced.
  2. Test every erasure request first against the conditions of Article 17 before you choose a measure.
  3. Assess whether source removal, output filtering or blocking suffices, and justify why or why not.
  4. As a recommended decision-making approach, reserve retraining or model adjustment for cases where the documented assessment shows that lighter measures do not effectively address the request.
  5. Document the decision, the measure carried out, the validation test and the human approval.

Sources and references

  1. AI, Large Language Models and Data ProtectionData Protection Commission · 2026-05-23
  2. Irish Data Protection Commission welcomes EDPB opinion on the use of personal data in the development and deployment of AI modelsData Protection Commission · 2026-05-23
  3. Regulation (EU) 2016/679 (General Data Protection Regulation) - Article 17EUR-Lex · 2016-04-27

Sources: The article draws on the Irish Data Protection Commission's guidance on AI and data protection, its response to the EDPB opinion and Article 17 of the GDPR via EUR-Lex.

← All articles in this topic ← All articles