Data-subject rights meet machine learning
Lesson 4 of 6 in Privacy and Data Protection Law for AI: GDPR from Scratch.
Chapter III of the GDPR gives every person a toolkit of enforceable rights against whoever holds their data. Each one is straightforward against a database — and genuinely hard against a trained model. That gap is where much of AI privacy practice now lives.
Access (Art 15) — "show me what you have on me"
A copy of the personal data, the purposes, the recipients, and — for automated decision-making — meaningful information about the logic involved. For AI companies the hard question is scope: does access cover what the model would generate about the person? The EDPB’s ChatGPT taskforce expected providers to handle access requests covering both training data and outputs.
Rectification (Art 16) — "correct what is wrong"
Inaccurate personal data must be corrected. The noyb complaints put the dilemma on record: when a model hallucinates a false birth date or a fabricated crime, filtering the output may not satisfy a right to have the data corrected. No regulator has accepted "the weights make it impossible" as a legal answer — technical impossibility tends to become the controller’s problem, not the data subject’s.
Erasure (Art 17) — the right to be forgotten
Deletion when the data is no longer necessary, consent is withdrawn, or processing was unlawful. Against a trained model, true erasure may mean machine unlearning — an unsolved research problem at scale — or retraining. This is why deletion-aware design (dataset lineage, per-source indexing, retention discipline) is now an AI engineering requirement, not a nice-to-have.
Portability (Art 20), restriction (Art 18), objection (Art 21)
Portability: your provided data in machine-readable form. Restriction: freeze processing during disputes. Objection matters most for AI: where processing rests on legitimate interests, the person can object, and for direct marketing the objection is absolute. Opt-out mechanisms for AI training — like those Meta was required to offer for its 2024–25 training plans — are Article 21 in action, and Opinion 28/2024 treats an effective opt-out as a mitigation that strengthens the legitimate-interests balance.
Regulators have converged on a working position you should be able to recite. The EDPB ChatGPT taskforce report (May 2024) set common lines on lawful basis for scraping, transparency, and accuracy for LLMs. Opinion 28/2024 then answered the three structural questions: models trained on personal data are not automatically anonymous; legitimate interests can work for training, on a documented three-part test; and a model built on unlawfully processed data can poison downstream deployment — unless the deployer properly assessed the model or the model is genuinely anonymised. Note the incentive that last clause creates: proving anonymisation, or curating lawful data, becomes a commercial asset when your enterprise customers must document their own diligence.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.