AI and data protection: what the GDPR means for AI systems

Legal framework

How does the GDPR regulate the use of AI?

The GDPR regulates every use of AI in which personal data is processed – that is, names, email addresses, IP addresses or biometric characteristics.

As soon as an AI system reads, stores, analyses or outputs such data, data protection law applies in full. That affects the training phase and productive operation alike.

Which legal bases are applicable?

In practice, companies usually base AI processing on two grounds: consent under Article 6(1)(a) GDPR or legitimate interest under Article 6(1)(f) GDPR.

Consent often fails because of technical reality – data once embedded in a neural network can hardly be deleted selectively, even though data subjects may withdraw their consent at any time. Legitimate interest goes considerably further but hits hard limits with sensitive data and minors.

Does the GDPR also apply to AI training?

Yes. Under data protection law, the training phase is nothing other than processing under Article 4(2) GDPR and needs the same legal basis as any other data process.

If existing data is repurposed for training, the principle of purpose limitation applies – later use for model training is only permissible if the new purpose is compatible with the original one.

When does the AI Act replace the GDPR?

Never. Recital 63 of the AI Act makes clear that the regulation creates no independent legal basis for the processing of personal data.

Companies must comply with both sets of rules in parallel: the GDPR for data protection, the AI Act for the product safety of the AI system.

Data protection risks

What data protection risks arise from AI?

The use of AI creates particular data protection risks through black-box decisions, large-scale web scraping for training data and the subsequent change of purpose for existing data in the company.

These three risk areas are among the central AI risks and fundamentally distinguish AI processing from classic IT.

Black box and lack of transparency

Deep learning models learn autonomously and deliver results whose origin even the developers can no longer fully reconstruct. This black-box problem collides directly with the transparency requirement of Article 5(1)(a) GDPR.

Data subjects have a right of access under Article 15 GDPR, and the information obligations under Articles 13 and 14 require a comprehensible description of the processing. The solution is explainable AI – models that make their decisions auditable.

Web scraping and training data

Many foundation models were trained on billions of pieces of content found online, including personal data of people who never consented to processing. Subsequent consent is practically impossible with such datasets.

For companies this creates a follow-on risk: anyone using an externally trained model inherits its legal legacy. The data processing agreement under Article 28 GDPR only partly covers this.

Change of purpose for existing data

Customer data once collected for contract processing or customer service cannot simply be used to train an internal AI model. Article 5(1)(b) GDPR requires a compatibility assessment.

The context of the original collection, the type of data and the possible consequences of further processing must be examined. Result: in many cases, data subjects must be informed anew or a fresh legal basis is needed.

Principles

Which GDPR principles apply to AI systems?

The same five principles apply to AI systems as to any other data processing: lawfulness, purpose limitation, data minimisation, transparency and accuracy of the data.

Article 5 GDPR bundles these principles and serves as the yardstick for every processing decision in the AI lifecycle – from data collection to model output.

Lawfulness and purpose limitation

Every processing needs one of the six legal bases from Article 6 GDPR, and the purpose of processing must already be specified and legitimate at the time of collection.

For AI projects that means: anyone collecting training data documents purpose and legal basis before the first data point – not once the model is built.

Data minimisation with large training datasets

AI models learn better with more data, but data protection law demands the opposite. The solution lies in technology: anonymised and synthetic training data, federated learning and so-called privacy enhancing technologies (PETs).

The goal is not the absolute smallest amount of data but to limit processing to the minimum necessary for the purpose.

Transparency and rights of access

Under Articles 13 and 14 GDPR, companies must inform about the use of AI, purpose, legal basis and any automated decisions before processing. Article 15 GDPR guarantees data subjects a right of access that also covers the logic behind AI decisions.

The privacy policy must name AI systems explicitly and explain which inputs are processed and which data flows to external providers.

Accuracy and data quality

Article 5(1)(d) GDPR requires accurate and up-to-date data – a principle of particular relevance for AI hallucinations. Inaccurate training data produces inaccurate models, and wrong model outputs about people can constitute data protection violations.

Article 10(3) AI Act doubles this obligation for high-risk systems with its own quality requirements for training, validation and test datasets.

2026 update

How do the AI Act and the GDPR interact from August 2026?

The AI Act complements the GDPR and does not replace it – from 2 August 2026, the central obligations for high-risk AI systems apply.

Companies must comply with both sets of rules in parallel, for the same systems and often for the same data processing. This dual compliance is the central governance topic of 2026.

What applies from 2 August 2026?

From this date, the rules for high-risk AI systems under Article 6 AI Act are fully applicable. These include risk management, data quality documentation, human oversight, technical documentation and conformity assessment before market launch.

Initial rules – in particular the ban on certain AI practices and the AI literacy obligation – have already applied since 2 February 2025.

Which risk classes does the AI Act define?

The AI Act distinguishes four classes: unacceptable risk (prohibited), high risk (fully regulated), limited risk (transparency obligations) and minimal risk (no obligations).

High-risk systems are typically found in recruitment, lending, law enforcement, educational assessment and critical infrastructure. For companies, correct classification is the first step of any AI compliance.

What fines apply?

Violations of the ban on certain AI practices are punished under Article 99 AI Act with up to €35 million or 7 percent of global annual turnover. For violations of other obligations, the maximums are €15 million or 3 percent of annual turnover.

GDPR fines under Article 83 remain unaffected and can be imposed in addition – both sanction systems are cumulative.

AI literacy – mandatory since February 2025

Article 4 AI Act obliges providers and deployers to ensure sufficient AI competence among their employees. This obligation has been applicable since 2 February 2025.

That means: every person who works with an AI system professionally needs demonstrable knowledge of how the system used works, its limits and its risks. Training must be documented and fit the respective use case.

Implementation

How do companies use AI in a legally compliant way?

Legally compliant AI use starts with a complete AI inventory, relies on binding data protection impact assessments for high-risk applications and requires trained employees at all levels of the hierarchy.

Implementation is not a one-off project but an ongoing compliance process with clear responsibilities.

GDPR-compliant?

Whether a tool counts as GDPR-compliant depends on four factors: EU hosting or standard contractual clauses, a data processing agreement under Article 28 GDPR, model training on input data that can be switched off, and a documented deletion process.

In practice, the right tool selection and clean contract design decide on lawfulness – an AI agency for automation checks these criteria project by project and documents them.

When is a data protection impact assessment mandatory?

Article 35 GDPR requires a DPIA where there is likely to be a high risk to the rights and freedoms of natural persons. The supervisory authorities keep positive lists that regularly prescribe a DPIA for systematic evaluation or new technologies.

AI projects almost always meet at least two of these criteria at the same time. Without a completed DPIA, high-risk AI processing may not start.

Which technical measures are standard?

At the technical level, four measures have become established: anonymisation and pseudonymisation of training data, end-to-end encryption in transit and at rest, role-based access controls and regular audit logs of model access.

For particularly sensitive applications, federated learning, differential privacy and synthetic training data are added.

What AI inventory do companies need?

A clean AI inventory lists every system in use with purpose, data categories, legal basis, provider, risk class under the AI Act and DPIA status. Without this inventory, neither the AI Act obligation to conformity assessment nor the GDPR accountability obligation can be met.

In practice, such inventories only come about when someone systematically goes through the departments – many AI tools today run unofficially via business units.

FAQ

Frequently asked questions about AI and data protection