The EU AI Act and Training Data: What the Rules Require

By Stefan, founderPrivacy, compliance and security5 min read

The EU AI Act sets two main rules for training data. Providers of high-risk AI systems must follow documented data governance practices, covering how data was collected and prepared, including annotation and labeling and how it was checked for bias and gaps (Article 10). Providers of general-purpose AI models must publish a summary of the content used for training (Article 53). After the 2026 "Digital Omnibus" amendments, the main high-risk obligations apply from 02.12.2027 or 02.08.2028, depending on the system.

For anyone labeling data, the practical message is simple: keep a clear record of how your data was labeled, by whom, under which rules and with what quality checks.

This article explains the rules in plain words. It is not legal advice; check your specific case with a lawyer.

Who the Rules Apply To

The AI Act sorts AI by risk and the duties depend on both the risk level and your role.

Category Examples Training data duties
Prohibited practices Social scoring, certain manipulative or exploitative systems Banned outright
High-risk systems AI used in hiring, credit scoring, education, essential services and safety components of regulated products Data governance under Article 10
General-purpose AI models Large models that can be used for many tasks Technical documentation and a public training content summary under Article 53
Transparency cases Chatbots, AI-generated content Disclosure duties, not training data rules
Minimal risk Spam filters, most business tools No specific training data duties

Roles matter. Most training data duties fall on the provider, the company that develops the system or model and puts it on the market. A labeling company working for a provider doesn't carry those duties itself, but its work becomes part of the provider's evidence. That is why providers may ask their vendors for documentation of how data was labeled.

The GDPR still applies alongside the AI Act whenever training data contains personal data. See GDPR and AI Training Data.

High-Risk Systems: Article 10 Data Governance

Article 10 sets out what providers of high-risk AI systems must do with their training, validation and testing data.

1. Follow data governance practices (paragraph 2). These must cover:

  • design choices and where the data came from and how it was collected
  • data preparation operations, which the Act lists as "annotation, labelling, cleaning, updating, enrichment and aggregation"
  • the assumptions about what the data measures and represents
  • an examination of possible biases that could affect health, safety or fundamental rights
  • measures to detect, prevent and reduce those biases
  • gaps or shortcomings in the data and how they are addressed

2. Meet quality criteria (paragraph 3). Datasets must be "relevant, sufficiently representative and to the best extent possible, free of errors and complete" in view of the system's intended purpose, with appropriate statistical properties for the people it will be used on.

3. Fit the real setting (paragraph 4). Data must reflect the geographical, contextual, behavioural or functional setting where the system will be used. A hiring model trained only on one country's CVs may not meet this in another.

4. Sensitive data for bias detection (paragraph 5). Special categories of personal data may be processed to detect and correct bias, only when strictly necessary and with strict safeguards, including access controls and deletion afterwards.

For labeling, the words that matter most are "annotation, labelling" and "free of errors": the Act names labeling as part of data governance and expects its quality to be managed and documented.

General-Purpose AI Models: Article 53

Providers of general-purpose AI models, such as large language models offered to others, have their own duties under Article 53, which have applied since 02.08.2025. They must:

  • keep technical documentation on training, testing and evaluation, available to the EU AI Office and national authorities
  • give downstream companies that build on the model the information they need about its capabilities and limits
  • put in place a policy to respect EU copyright law, including rights holders' opt-outs
  • publish a "sufficiently detailed summary" of the content used for training, using a template from the AI Office

The European Commission published that template, guidelines on who counts as a general-purpose model provider and a voluntary Code of Practice ahead of the August 2025 start date.

What it means for human-labeled data: fine-tuning, preference and evaluation data written and labeled by people is part of the model's training story. Providers need to know where it came from and how it was produced, so vendors should be able to describe their sources, workforce and process clearly.

Key Dates and Penalties

The AI Act applies in stages. The 2026 Digital Omnibus amendments, published and in force since July 2026, pushed back the main high-risk obligations.

The main high-risk duties now start in December 2027 · EU AI Act dates after the 2026 Digital Omnibus
StepApplies from
Act enters into force01.08.2024
Prohibitions apply02.02.2025
GPAI model duties02.08.2025
AI-content transparency02.12.2026
High-risk (Annex III)02.12.2027
High-risk (Annex I)02.08.2028

High-risk systems listed in Annex III, such as those used in hiring, credit and education, must comply from 02.12.2027; high-risk systems in regulated products under Annex I from 02.08.2028. Duties for general-purpose AI models have applied since 02.08.2025.

Penalties under Article 99:

Breach Maximum fine
Prohibited practices €35 million or 7% of worldwide annual turnover, whichever is higher
Most other obligations, including providers' data governance duties €15 million or 3%, whichever is higher
Incorrect or misleading information to authorities €7,50 million or 1%, whichever is higher

For small and medium-sized companies, including start-ups, the cap is whichever of the two amounts is lower.

What This Means for Labeling Projects

If you build a high-risk system or a general-purpose model, your labeling process is part of what you may need to show. A practical checklist:

  • Keep the annotation guidelines, with every version and change log
  • Record who labeled the data: in-house team or vendor and how annotators were selected and tested
  • Keep the quality numbers: accuracy on gold items, agreement between annotators, review results
  • Document how the sample was chosen and check it represents the people and settings where the system will be used
  • Look for bias and gaps: compare labels and accuracy across groups, languages and regions
  • Record any personal data used, the lawful basis under the GDPR and the safeguards
  • Ask your labeling vendor for this information as part of each delivery

Labeling You Can Document

AnnoTrue delivers every batch with a quality report, the guideline version used and a record of how the work was checked, so you have what you need for your own compliance file. Start with a free pilot.

Sources