GDPR Right to Erasure and Machine Unlearning

Overview

This series explains how the GDPR right to erasure moves from a legal decision to actions across live systems, processors, recipients, backups, and, when relevant, learned models. It distinguishes deletion, restriction, anonymisation, and machine unlearning, then uses vision-language-model tests to examine alternate prompts, indirect context, and retraining. The central boundary is evidence: a changed record or quiet output supports only the scope and conditions actually mapped and tested, not universal proof that a model has forgotten.

Learning outcomes

GDPR erasure is a scoped legal decision, not proof that every associated record or derived state vanished.
Account closure, masking, restriction, backup control, and deletion produce different system states.
VLM unlearning should test alternate prompts and retraining while measuring retained and general capability.
A defensible claim names its object, scope, action, recovery coverage, and remaining uncertainty.

Keywords

GDPRErasureUnlearningRecoverabilityBackupsAnonymisationProcessorsRetentionVLM

Chapters

GDPR Right to Erasure vs Machine Unlearning

Scope, hidden copies, and recoverable associations

9 min · Scope, hidden copies, and recoverable associations

9 min

GDPR Right to Erasure: Grounds and Exceptions

How controllers decide, communicate, and document outcomes

8 min · How controllers decide, communicate, and document outcomes

8 min

GDPR Right to Erasure in Real Systems

Mapping scope, backups, recipients, and proof

10 min · Mapping scope, backups, recipients, and proof

10 min

Machine Unlearning in Vision-Language Models

Where representations, outputs, and trade-offs diverge

9 min · Where representations, outputs, and trade-offs diverge

9 min

Testing Machine Unlearning Robustness

Prompts, recovery attacks, and benchmark limits

9 min · Prompts, recovery attacks, and benchmark limits

9 min

From GDPR Erasure to a Defensible Forgetting Claim

Match each conclusion to the evidence it covers

8 min · Match each conclusion to the evidence it covers

8 min

GDPR Right to Erasure: Model Forgetting Needs Proof

A bounded conclusion across law, systems, and tests

10 min · A bounded conclusion across law, systems, and tests

10 min

Preview every chapter · Unlock full text, audio, and video from chapter 4

Plans

Ch.1-3 FreeCh.4-7 Premium

FAQ

Does deleting data prove that a model has forgotten it?

No. Deleting a source record does not by itself change learned parameters or prove that the model cannot recover the association.

What does the GDPR right to erasure require?

The GDPR right to erasure requires a controller to assess the processing, identify an applicable Article 17 ground, test relevant exceptions, and communicate a timely, demonstrable outcome. Restriction or lawful retention may apply instead for part of the data.

How can machine unlearning be tested for robustness?

Machine unlearning is tested by changing access routes: use original, paraphrased, and discriminative prompts, then indirect context and matched or related retraining. Measure target suppression with retained and general capability so broad failure is not mistaken for selective forgetting.

What makes a forgetting claim defensible?

A defensible claim names the object, legal scope, action, persistence paths addressed, tests performed, and evidence limits. It should state only what the mapped systems and tested model behaviour support.