Please delete! What the right to be forgotten means for AI models

The “right to be forgotten” under Article 17 of the GDPR obliges companies to permanently delete personal data upon request – e.g., in the event of withdrawal of consent, fulfillment of purpose, or unlawful processing.

Relevance for AI systems:
Many companies use personal data to improve algorithms, for example:

• for behavioral analysis in e-commerce
• for chatbots with personalized responses
• in applicant scoring systems

But if this data has to be deleted later, what happens to the AI model that was trained on it?

Problem: The GDPR does not distinguish between a classic database and a trained model. This leads to uncertainty – because even “learned knowledge” can be personal if it is traceable.

An AI model “learns” by processing training data. This results in model weights or vector representations that are not structured like classic data sets and cannot be deleted.

Example 1 – Chatbot trained with CRM data:
A chatbot is trained with real customer inquiries from the CRM, including names, problems, and order numbers. Even if the original data is deleted, the patterns remain in the model.

Example 2 – Applicant management with AI:
HR software uses applicant profiles to improve scoring. If an applicant is deleted, the model can continue to use their language style or criteria patterns – indirectly personal data.

Risk: If this data can be reconstructed later or is used as a significant factor in decisions, this constitutes a data protection violation.

When a request for deletion is made, the question arises: Does the AI model need to be retrained or adapted?

The GDPR does not provide a clear answer, but according to supervisory authorities, the following applies:

If the trained model stores personal information or makes it reproducible, the model is also subject to the obligation to delete.

Real-world example:
Meta has already been criticized by data protection authorities because training data with personal content could no longer be completely deleted in LLMs (Large Language Models). There is a growing expectation that AI systems should be able to respond to data deletion requests in a transparent and reversible manner.

Typical problems in companies:

Model was trained with live customer data, but no log exists
No versioning → it is not known which data is contained in which model

Deletion request also affects data that was relevant for training purposes

This is where AI engineering and data protection come together – there are technical options for enabling the right to be forgotten in practice:

1. Machine unlearning

Targeted removal of individual training data from a model without complete retraining.

• Works well with simple models (e.g., decision trees)
• More difficult with deep neural networks

2. Differential privacy

The model is trained in such a way that no conclusions can be drawn about individual data.

• Used by OpenAI, Google, Apple
• Well suited for aggregated data sets, but less so for individual contexts

3. Retraining with deletion list

The model is retrained regularly – without deleted data.

• Very reliable
• But: complex, expensive, only practical for large companies

4. Federated learning

Data remains decentralized – models learn locally and are merged centrally.

• Advantage: deletions can be implemented locally
• Complex to integrate

Tip: For companies with a lot of data deletions, a “model-centric” deletion process with automated audit trails and versioning is worthwhile.

The obligation to delete data does not end with CRM or the cloud—it increasingly affects algorithmic systems as well. That is why companies need:

1. Privacy by design for AI systems

Clarify early on: Which data may be used?
Pseudonymize or aggregate training data as much as possible

2. Model documentation

Who trained what with which data?
Which models are in use? Which version contains which data?

3. Process for deletion requests

Interdisciplinary between data protection, IT, and data science
Automated checks to see if models are affected
Model retraining or replacement if necessary

4. DPIA (Data Protection Impact Assessment)

Especially for sensitive models (HR, health, behavior, risk scoring)
Must assess whether data subject rights are technically enforceable

The right to be forgotten poses real challenges for AI developers and companies - but it can be solved. With documented data flows, technical “unlearning,” and clear responsibilities, the AI world can also become GDPR-compliant.

Companies building AI models today must consider reversibility - otherwise they risk expensive retraining, fines, and damage to their image.