GDPR and AI Training Data: What Companies Must Know

By Stefan, founderPillar guide · Privacy, compliance and security11 min read

If your training data contains information about identifiable people in the EU, the GDPR applies to it, from collection through labeling and training to deletion. That means you need a lawful basis for using the data, you must use only what you need, you must tell people how their data is used and any labeling vendor must work under a written data processing agreement.

The good news: most training projects can reduce their GDPR exposure a lot with a few early decisions, mainly removing personal data the model doesn't need before anyone labels it.

This guide explains the rules in plain words. It is not legal advice; for a specific project, talk to a data protection lawyer or your data protection officer.

When the GDPR Applies to Training Data

The GDPR applies whenever data relates to a person who is identified or identifiable. That covers far more than names and emails. In training data, personal data often hides in:

  • customer emails, chats and support tickets
  • call recordings and their transcripts
  • invoices, contracts and forms with names, addresses and signatures
  • photos and videos showing faces, licence plates or house numbers
  • free-text fields, where people write anything about themselves
  • prompts and conversations users had with your chatbot

Anonymised and pseudonymised data are not the same. This distinction decides whether the GDPR applies at all:

Pseudonymised data Anonymised data
What it means Direct identifiers are replaced (for example, a name becomes "Customer 4821"), but someone holding extra information could link it back Nobody can identify the person any more, using any means reasonably likely to be used
GDPR applies? Yes. It is still personal data. No. It falls outside the GDPR.
Typical use Reduces risk; a good safety measure Ideal when the task truly doesn't need personal details

Recital 26 of the GDPR sets the test: whether a person can be identified, taking into account "all the means reasonably likely to be used", including cost, time and available technology. In practice, true anonymisation is hard. Removing names but keeping a rare job title, a small town and a date of birth may still point to one person.

Practical techniques and their limits, are in How to Anonymize Data Before Sending It for Annotation.

A Lawful Basis for Training

Every use of personal data needs one of the six lawful bases in Article 6 of the GDPR. For AI training, two come up most often.

Lawful basis When it fits The difficulty
Consent You can ask people directly and clearly and they can refuse without losing anything Must be specific, informed and freely given and people can withdraw it at any time; hard for data you already hold
Legitimate interest You have a real business interest, the processing is necessary for it and it doesn't override people's rights You must document a balancing test and people can object

What European regulators have said. In December 2024, the European Data Protection Board adopted Opinion 28/2024 on personal data in AI models. Its main points, in short:

  • Legitimate interest can be a valid basis for developing AI models, but not by default. You must pass a three-step test: a legitimate interest, processing that is necessary for it and a balance that doesn't override people's rights.
  • People's reasonable expectations matter. Safeguards such as pseudonymisation, clearer information and an easy way to opt out can help tip the balance.
  • An AI model is only "anonymous" if extracting personal data from it, directly or through questions, is very unlikely. This is decided case by case.
  • If personal data was processed unlawfully during development, authorities can order corrective measures, which may affect the model built from it.

Reusing data you already have. Data collected for one purpose, such as answering support tickets, can't automatically be used for another, such as training a model. Article 6(4) requires a compatibility check: how related the new purpose is, what people would expect, the nature of the data and what safeguards apply.

The GDPR Principles, Applied to AI Data

Article 5 of the GDPR sets out the principles that every processing must follow. Here is what they mean in a training and labeling project.

Principle What it means for training data In practice
Purpose limitation Use data only for purposes you defined and told people about Write down what the dataset is for before collecting or labeling it
Data minimisation Use only the personal data you actually need Remove names, contacts and IDs the model doesn't need before labeling
Accuracy Keep personal data correct where it matters Correct or remove records you know are wrong
Storage limitation Keep personal data no longer than needed Set deletion dates for raw data, labeled sets and vendor copies
Integrity and confidentiality Protect data against misuse and leaks Access control, encryption, no downloads by annotators, audit logs
Transparency Tell people how their data is used Update privacy notices to mention AI training, in plain words
Accountability Be able to prove you comply Keep records: the lawful basis, assessments, contracts, deletion confirmations

Minimisation is the most useful principle in practice. Every personal detail removed before labeling is one fewer thing that can leak, one fewer access right to manage and often one fewer legal question. Ask of every field: does the model need this to learn the task?

Sensitive Data and Impact Assessments

Special categories of data need extra care. Article 9 of the GDPR lists data that is especially sensitive: health, racial or ethnic origin, political opinions, religious beliefs, trade union membership, genetic and biometric data and data about sex life or sexual orientation. Processing it is forbidden unless one of a short list of exceptions applies, such as explicit consent.

This matters for labeling because sensitive data turns up where you don't expect it: a support chat that mentions an illness, a photo that shows a religious symbol, a voice recording used to identify a speaker. If a task doesn't need it, remove or mask it before labeling. If it does, such as clinical notes for a medical model, get legal advice first and limit access to a small, trained team.

Data protection impact assessments (DPIAs). Article 35 requires a DPIA before processing that is likely to result in a high risk to people's rights, especially when using new technologies. Large-scale training on personal data, sensitive data, or data about vulnerable people often meets that bar. A DPIA describes the processing, checks that it is necessary and proportionate, lists the risks and sets out how you will reduce them. Even when not strictly required, writing one is a good way to think a project through.

Outsourcing Labeling: The Data Processing Agreement

When you send personal data to a labeling company, you are usually the controller (you decide why and how the data is used) and the vendor is a processor (it works on your behalf and on your instructions). The GDPR requires a written contract between you, often called a data processing agreement or DPA.

Article 28(3) lists what that contract must say. The processor must:

  1. process the data only on your documented instructions, including on transfers outside the EU
  2. make sure everyone with access is bound by confidentiality
  3. apply appropriate security measures (Article 32)
  4. use sub-processors only with your authorisation, under the same obligations
  5. help you respond to people exercising their rights
  6. help you with security, breach notification and impact assessments
  7. delete or return all personal data at the end of the service
  8. give you the information needed to show compliance and allow audits

For labeling projects, add the practical details: who the annotators are and where they work, whether they can download or copy data, which tool is used and where it is hosted, how long data is kept and how deletion is confirmed.

Annotators are part of the chain. Whether they are employees or contractors, they must be bound by confidentiality and see only the data their task needs. Ask your vendor how this is enforced, not just whether it is promised. See Data Security in Outsourced Labeling: Questions to Ask a Vendor.

Annotators Outside the EU: International Transfers

Labeling teams are often spread across many countries. When people outside the European Economic Area can access personal data from the EU, the GDPR's rules on international transfers (Chapter V) come into play and regulators generally treat remote access as a transfer.

Transfers are allowed in three main ways:

Route How it works
Adequacy decision The European Commission has found that the country protects data to an EU-equivalent standard, so data can flow as within the EU. The United Kingdom and Switzerland are examples. For the United States, the EU-US Data Privacy Framework, adopted on 10.07.2023, covers only US companies that have joined it.
Standard contractual clauses (SCCs) Model contract clauses adopted by the Commission, the current set on 04.06.2021. They are signed with the recipient, together with an assessment of the destination country's laws and any extra safeguards needed.
Exceptions (Article 49) Narrow cases such as explicit consent for a specific transfer. Not suitable for regular, repeated labeling work.

The simplest safeguard is still minimisation. If the data is properly anonymised before labeling, transfer rules don't apply to it. If it isn't, ask your vendor which countries its annotators work from and which transfer route covers each.

Restricting by location is also an option. For sensitive projects, many companies require that annotators work only from the EU or from countries with an adequacy decision. A good vendor can limit a project to eligible countries.

People's Rights and What They Mean for Datasets

The people whose data sits in your training set keep their rights under the GDPR. Plan for these requests before they arrive, because they are much harder to handle once data has been copied, labeled and trained on.

Right What people can ask What it means for a training project
Information (Articles 13 and 14) To be told how their data is used Privacy notices must mention AI training in clear words
Access (Article 15) A copy of their data and how it is processed You need to find a person's records across raw and labeled datasets
Erasure (Article 17) Deletion of their data, in many situations Deleting from datasets is possible; removing what a trained model learned is much harder
Objection (Article 21) To stop processing based on legitimate interest You need a way to exclude their data from future training
Rectification (Article 16) Correction of wrong data Corrections should reach the dataset, not just the source system

Make data findable. Keep a link between dataset records and their source, stored separately and securely. Without it, you can't answer access or deletion requests reliably.

Pseudonymise early. Replacing direct identifiers before labeling protects people while keeping the ability to find and remove their records when asked.

How the GDPR Fits with the EU AI Act

The GDPR protects people's personal data. The EU AI Act regulates AI systems themselves, based on their risk. They overlap on training data and both can apply to the same project.

For providers of high-risk AI systems, such as those used in hiring, credit scoring or education, Article 10 of the AI Act requires data governance practices for training, validation and testing data. These explicitly cover data preparation operations, including annotation and labeling, as well as checks for bias and for gaps in the data.

The AI Act also allows processing special categories of personal data to detect and correct bias, under strict safeguards. The 2026 "Digital Omnibus" amendments widened this beyond high-risk systems, while keeping a strict-necessity test, pseudonymisation, access limits and deletion requirements.

Timing. Under the Digital Omnibus, published and in force since July 2026, the main obligations for high-risk systems listed in Annex III now apply from 02.12.2027 and for those covered by Annex I from 02.08.2028. The GDPR, by contrast, already applies in full.

What this means for labeling: document how your data was labeled, by whom, under which guidelines and with what quality checks. A good labeling partner gives you that record with every delivery. More detail is in The EU AI Act and Training Data: What the Rules Require.

Checklist Before You Send Data for Labeling

Work through this list before any annotator sees your data. Most items take minutes; skipping them can take months to fix.

Purpose and lawful basis

  • The purpose of the dataset is written down
  • The lawful basis is chosen and documented (with a balancing test, if legitimate interest)
  • Privacy notices tell people their data may be used for AI training
  • A DPIA is done, or you have recorded why it isn't needed

The data itself

  • Personal data the task doesn't need is removed or masked
  • Sensitive (special category) data is removed, or its use is legally cleared
  • Remaining identifiers are pseudonymised where possible
  • You can trace records back to their source to answer access or deletion requests

The vendor

  • An NDA and a data processing agreement meeting Article 28 are signed
  • You know which countries annotators work from and the transfer route for each
  • Annotators are bound by confidentiality and can't download or copy data
  • Storage location, access rules and retention period are agreed in writing

After the project

  • Raw data and vendor copies are deleted on schedule
  • Deletion is confirmed in writing and the confirmation is kept
  • A record of how the data was labeled and checked is stored with the dataset

How AnnoTrue Protects Your Data

We sign an NDA and a data processing agreement before we see any data. Data is stored in the EU, annotators work inside your tool or ours without downloading anything, access is limited to the people on your project and projects can be restricted to annotators in eligible countries. When the project ends, we delete your data and confirm it in writing.

Sources