Privacygevoelige data in AI: van mogen naar aantoonbaar controleren

Je moet kunnen aantonen welke persoonsgegevens je AI-systemen ingaan, hoe ze worden gebruikt, en welke controles je hebt ingesteld. Dit is niet langer een kwestie van toestemming vragen — het is een verplichting tot verificatie en transparantie.

An analysis of 17 August 2026 of how privacy safeguards for AI systems with personal data must now be demonstrable argues that regulators across jurisdictions have shifted the burden from permission to proof. The European Data Protection Board issued guidance on web scraping in generative AI on 8 July 2026, making clear that processing personal data in model training falls under GDPR and requires documented legal basis, transparency, data minimisation and controls on sensitive categories. In our assessment, this marks a structural change for any organisation using AI on confidential or personal information: you must now be able to show auditors and supervisors exactly what data entered your systems, where it went, and what safeguards operated at each step.

Wat heeft zich in juli 2026 veranderd?

Two separate regulatory moves converged in July 2026. The European Data Protection Board issued its guidance on web scraping and generative AI, establishing that personal data processing in AI training is subject to GDPR requirements — legal basis, transparency, data minimisation, and restrictions on special categories. In the same period, the European Commission published guidelines on transparency obligations for AI system providers and users. These confirmed that Article 50 of the AI Act took effect on 2 August 2026, requiring that people be informed when they interact with AI or encounter AI-generated content. The effect is the same in both cases: AI use cannot remain hidden inside a workflow.

This is not a European phenomenon alone. Singapore introduced AI-specific notification requirements on 20 July 2026 for organisations using personal data to train generative AI models. Users must be told which data types and purposes apply, and what opt-out options exist. The underlying concern is consistent across jurisdictions: sensitive personal data is difficult to remove once it enters model training.

Welke risico's ontstaan als je privacygevoelige data niet beperkt?

• Data leakage and memorisation — personal data reproduced or inferred from model outputs once training is complete.
• Lack of legal basis — processing personal data without documented lawful grounds under GDPR.
• Absence of user notice — deploying AI on personal information without informing the individuals affected.
• Loss of data minimisation — retaining or processing more personal data than necessary for the stated purpose.
• Uncontrolled downstream use — AI-generated outputs containing personal data used or published without further safeguards.

Wat moet je kunnen laten zien aan toezichthouders?

1. Document which personal data enters each AI workflow — record the data types, their source, and the lawful basis for processing them.
2. Record which AI model processes the data and for what purpose — maintain an audit trail linking data to model, model to use case, and use case to business justification.
3. Demonstrate data minimisation steps taken before AI processing — show what sensitive values were removed, masked, or replaced with synthetic equivalents.
4. Log user interactions with AI outputs — record when and how individuals were informed that they encountered AI-generated content or AI-mediated decisions.
5. Retain evidence of control and correction mechanisms — document how errors, hallucinations or policy violations were detected and remedied before output was used or shared.

Hoe verschilt dit van eerdere privacyregels?

Privacy regulation has long required legal basis, transparency and data minimisation. What has changed is the specificity demanded and the burden of proof. Under GDPR alone, an organisation could argue that it had obtained consent or identified a legitimate interest. Under the combined effect of the EDPB guidance and the AI Act transparency rules, you must now show the mechanism: which data, which model, which safeguards, which user notification, and which control step. Toezichthouders vragen niet om een intentieverklaring, maar om bewijs dat je weet wat er gebeurt en dat je grip houdt op de gegevens.

The AI Risk Management Framework published by NIST reinforces this shift. Organisations deploying generative AI on confidential or personal data must demonstrate appropriate safeguards, data minimisation and risk controls. Responsibility remains with the party deploying the AI, not the model provider or the infrastructure operator.

Welke architectuurkeuzes helpen je aan aantoonbaarheid?

Two architectural patterns support demonstrable control. The first is verification: routing a task through selected, independent AI models while making verification steps, corrections, disagreements and sources visible for inspection. This does not guarantee correctness or eliminate hallucinations, but it enables the user to retain control and to show auditors what happened. The second is data minimisation at the point of entry: replacing sensitive values in a document with synthetic, session-bound equivalents on EU infrastructure before AI processing occurs. The AI chain then analyzes the synthetic version; original values can be restored locally afterward. If the privacy control fails, the document is not sent — the workflow is fail-closed. Both patterns create an audit trail that satisfies the EDPB and Commission requirement: you can show which data was used, which AI interaction was made visible, and which control and minimisation steps were executed.

Tooling can make these steps visible and repeatable. What remains with you is the professional judgement: which data you process, which outputs you send downstream, and what you publish. Regulators will ask you to show your work. The systems you choose must let you do that.