The GDPR applies to AI training data whenever a licensed dataset contains personal data: the seller needs a legal basis to disclose it, the buyer needs one to train on it, and both must inform the people in it and secure any transfer outside the EU. The Court of Justice accepts disclosure for consideration on legitimate interests if strictly necessary (C-621/22, 2024-10-04), the EDPB holds that a trained model is not anonymous in all cases, and on 2026-03-18 the Tribunale di Roma upheld OpenAI's opposition to the Garante's EUR 15 million decision.
The rules come from Regulation (EU) 2016/679 (CELEX 32016R0679), in force since 2016-05-24 and applicable from 2018-05-25, and form the privacy layer of AI training data licensing. Rulings are as of November 9, 2026; the unadopted proposal COM(2025) 837 is kept apart from final law.
When the GDPR Applies to a Company Licensing Data for AI Training
The GDPR applies to a data deal when the dataset holds information about identifiable people, such as names in support tickets, voices in call recordings or authors of internal messages, and it applies to the seller and the buyer separately. Each disclosure needs an Article 6 basis, an Article 6(4) compatibility test, Article 9 handling, notices, usually an impact assessment, national employee rules (Article 88) and a Chapter V transfer tool.
Identifiability depends on who holds the data: in EDPS v SRB (C-413/23 P, 2025-09-04) the Court of Justice held that pseudonymized data are not personal in all cases and for every person, judged by the means each entity is reasonably likely to use. Content matters too: the Irish Data Protection Commission (DPC) stated in its AI Insights Report of September 2026 that a controller is unlikely to show that people expect their private messages to be used for AI training.
Seller, Buyer and Service Provider: Who Is Controller in a Data Sale
In a sale or license the seller decides to disclose and the buyer decides to train, so each pursues its own purpose, and the standard contractual clauses include a controller-to-controller module for that transfer. The buyer, as recipient of data it did not collect, informs the people in the set under Article 14.
Buyers describe the technical side as claims: Scale AI states that anonymization happens inside the data holder's environment before any data is extracted, and Mercor states that it covers extraction, anonymization and transfer costs (both checked 2026-11-09). A service provider's role depends on who sets the purposes and means of processing, which the data protection clauses of AI data licensing agreements fix.
The GDPR Steps in a Training Data Deal at a Glance
Eight GDPR steps decide whether personal data can go into a training data deal, and each row names the article, who acts and the page that covers it in depth.
The table follows deal order as of 2026-11-09; the source is Regulation (EU) 2016/679 unless the row names another.
| Step | Article | Who acts | What it requires | Source | Detail page |
|---|---|---|---|---|---|
| Legal basis | Art. 6(1)(f) | seller (disclosure), buyer (training) | lawful interest, strict necessity, balancing | eur-lex.europa.eu; C-621/22 | Legitimate interest for AI training data |
| Compatibility of reuse | Art. 6(4) | seller | five factors (a) to (e), unless consent or Union or member-state law | eur-lex.europa.eu | Purpose compatibility when reusing data |
| Special categories | Art. 9 | seller and buyer | an Art. 9(2) exception in addition to an Art. 6 basis | eur-lex.europa.eu; C-252/21 | Special category data in AI training |
| Notices | Arts. 13(3), 14 | seller before reuse; buyer within one month | information, or the Art. 14(5)(b) exemption with public information | eur-lex.europa.eu | GDPR notices before a data sale |
| DPIA | Art. 35(1) | each controller | impact assessment before processing likely to result in a high risk | eur-lex.europa.eu | DPIA for an AI data licensing deal |
| Employee data | Art. 88 | seller | national rules apply only if they meet Art. 88(2) | eur-lex.europa.eu; C-34/21 | Employee data under Article 88 |
| Transfer outside the EU | Arts. 44 to 49 | seller as exporter | DPF, SCCs or an Art. 49 derogation | eur-lex.europa.eu (32023D1795, 32021D0914) | Transferring training data to US buyers |
| Anonymity of the trained model | EDPB Opinion 28/2024 | buyer | extraction and query retrieval of personal data insignificant | edpb.europa.eu | When an AI model is anonymous |
The steps are cumulative: a legal basis does not replace a notice, and a notice does not replace a transfer tool.
A Legal Basis to Disclose and to Train: Article 6(1)(f) in Three Steps
Most data sales rest on legitimate interests under Article 6(1)(f), which allows processing necessary for an interest of the controller or a third party unless the people's interests or fundamental rights override it, in particular where a child is involved. The test covers disclosure and training, in three steps:
- Interest: a lawful interest, which need not be laid down by law (C-621/22).
- Necessity: the processing is strictly necessary for that interest.
- Balancing: the people's interests or rights do not override it.
Consent of the people in the set is the alternative basis.
Disclosing Personal Data for Payment: What CJEU C-621/22 Allows
The Court of Justice held on October 4, 2024 that disclosing personal data for consideration to serve a commercial interest can rest on Article 6(1)(f), but only if strictly necessary and not overridden by the people's interests. The case, KNLTB v Autoriteit Persoonsgegevens (Ninth Chamber), concerned a Dutch sports federation disclosing members' data for consideration. The judgment does not concern AI, but licensing customer records to an AI buyer for a fee is the same kind of disclosure, so the test applies by analogy.
How the EDPB Assesses AI Model Development (Opinion 28/2024)
EDPB Opinion 28/2024, adopted on December 17, 2024 at the Irish supervisory authority's request, sets out how supervisory authorities assess legitimate interest for developing and deploying AI models and what follows when development was unlawful. It is soft law that does not bind courts, and its corrective measures include erasure of the training dataset and/or the model (paragraph 114).
Two records on Meta's training show the test in practice: the Oberlandesgericht Köln accepted on May 23, 2025, on summary review, that AI training can be a legitimate interest (15 UKl 2/25), and the DPC recorded on May 21, 2025 Meta's legitimate interest assessment, DPIA and compatibility assessment before its planned EU training start of May 27, 2025.
Staff Records in the Set
Employee emails, chats and HR files follow the same articles, but Article 88 lets member states and collective agreements add more specific rules, which are national. The Court of Justice held in C-34/21 (March 30, 2023) that national rules lacking the measures of Article 88(2) must be disregarded unless they form an Article 6(3) legal basis. Those rules sit on the member-state pages and in employee data and AI training rules by jurisdiction.
Selling Data Collected for Another Purpose: The Article 6(4) Factors
Company data was collected for service, sales or employment, so licensing it for AI training is further processing, and unless it rests on consent or Union or member-state law the seller must test whether the new purpose is compatible under Article 6(4):
- (a) any link between the original purpose and AI training;
- (b) the context of collection, in particular the relationship between the people and the seller;
- (c) the nature of the data, in particular special categories (Article 9) and criminal data (Article 10);
- (d) the possible consequences for the people;
- (e) appropriate safeguards, which may include encryption or pseudonymization.
The EDPB's Guidelines 01/2025 state that pseudonymization's risk reduction may support both this test and the Article 6(1)(f) balance, though only a consultation version (2025-01-17 to 2025-03-14) exists as of November 9, 2026. Meta's compatibility assessment of May 2025 is one documented case of this test.
Health, Biometric and Other Sensitive Items in a Licensed Set (Art. 9)
Data revealing health, genetic data, biometric data used to identify a person, racial or ethnic origin, political opinions, religious or philosophical beliefs, trade union membership or sex life and sexual orientation may not be processed unless an Article 9(2) exception applies, on top of an Article 6 basis. Such items can sit in support tickets, call recordings and HR files.
The Grand Chamber held in Meta v Bundeskartellamt (C-252/21, July 4, 2023) that a set collected en bloc with at least one sensitive item is prohibited absent an Article 9(2) derogation, and that the exception for data made public (Article 9(2)(e)) needs the person's explicit intention to make them public, as Opinion 28/2024 quotes in paragraph 17. A seller without an exception filters out such items or excludes the affected records.
Who Must Inform the People in the Data, and When
The seller must inform people before it uses their data for the new purpose (Article 13(3)), and the buyer that receives personal data must inform them under Article 14 within a reasonable period, at the latest one month after obtaining it. Three provisions set the timing and the exit:
- Article 13(3): the seller informs on the new purpose before further processing.
- Article 14(3)(a): the buyer informs within one month at the latest.
- Article 14(5)(b): where informing each person is impossible or disproportionate, the buyer takes appropriate measures instead, including making the information public.
In C-413/23 P the Court of Justice held that the duty to inform people about recipients is assessed at collection from the controller's perspective, irrespective of pseudonymization for the recipient. On objections, Meta states that objections to its AI training are honored at any time (page dated 2026-03-27), while noyb's letter of May 14, 2025 says Meta limits their effect to objections made before training starts. The two statements contradict each other.
Personal, Pseudonymized or Anonymous: How Far the GDPR Reaches Into a Dataset
Anonymous data falls outside the GDPR and pseudonymized data does not automatically, and since September 4, 2025 the Court of Justice judges pseudonymized data from the means each recipient is reasonably likely to use. The consultation version of EDPB Guidelines 01/2025 says the opposite: data attributable with additional information are personal even if another person holds it. EDPB Guidelines 02/2026 on anonymization, adopted on July 7, 2026 for consultation (to October 30, 2026) and not final as of November 9, 2026, set out:
- three criteria: no record isolation, no linkage and no inference;
- assessment from each relevant entity's perspective;
- no new assessment for datasets found anonymous under WP29 Opinion 05/2014.
A model trained on personal data, Opinion 28/2024 states, cannot in all cases be considered anonymous: extraction of personal data and obtaining it through queries must both be insignificant, judged by all means reasonably likely to be used.
Delivering Personal Data to a Buyer Outside the EU: Chapter V Tools
Personal data delivered to a buyer outside the EU needs a Chapter V tool: an adequacy decision, appropriate safeguards such as standard contractual clauses, or a narrow derogation (Articles 45, 46 and 49). For a US buyer the routes are:
- EU-US Data Privacy Framework (DPF): Commission Implementing Decision (EU) 2023/1795 of July 10, 2023 (CELEX 32023D1795) covers organizations on the DPF List, so a transfer to a buyer on the list needs no specific authorization (Article 45(1)) and no SCCs.
- Standard contractual clauses: Commission Implementing Decision (EU) 2021/914 of June 4, 2021 (CELEX 32021D0914), with a controller-to-controller module that fits a sale; other Article 46 tools need enforceable rights and effective remedies.
- Article 49(1) derogations for specific situations, such as explicit consent or necessity for a contract.
As of November 9, 2026 the DPF stands: the General Court dismissed Latombe v Commission (T-553/23) on September 3, 2025, and the appeal C-703/25 P, brought on October 31, 2025, is pending.
What Regulators and Courts Have Decided on Personal Data in AI Training
Regulators and courts have tested AI training under the GDPR in a handful of decisions, and none of them concerns a company selling its own records to an AI buyer. The table lists nine records, newest first, as of 2026-11-09.
| Date | Authority or court | Matter | Outcome as stated by the source | Source | Page |
|---|---|---|---|---|---|
| 2026-09 | Irish DPC | AI Insights Report | about 180 AI products and services engaged 2021 to 2025 (DPC's own count); unlikely that people expect private messages to be used for AI training | dataprotection.ie | n/a |
| 2026-03-18 | Tribunale di Roma | OpenAI v Garante, judgment n. 4153/2026 | upheld OpenAI's opposition to Garante decision n. 755 of 2024-11-02 (EUR 15 million, six-month information campaign); Garante temporarily removed the decision from its website; earlier urgent limitation 2023-03-30 | garanteprivacy.it (docweb 10085432, 10085455) | Italian Garante v. OpenAI |
| 2026-03-27 / 2025-05-14 | Meta; noyb (statements) | effect of objections to Meta AI training | Meta: objections honored at any time; noyb: Meta limits their effect to objections before training starts; noyb states it complained to 11 authorities on 2024-06-06 | about.fb.com; noyb.eu | n/a |
| 2025-09-04 | Court of Justice | EDPS v SRB, C-413/23 P | pseudonymized data not personal in all cases and for every person; General Court judgment T-557/20 set aside, case referred back | eur-lex.europa.eu (62023CJ0413) | EDPS v SRB (C-413/23 P) |
| 2025-09-03 | General Court | Latombe v Commission, T-553/23 | action against the DPF decision dismissed; appeal C-703/25 P pending | publications.europa.eu (62023TJ0553) | n/a |
| 2025-05-23 | OLG Köln | Verbraucherzentrale NRW v Meta, 15 UKl 2/25 | interim injunction application rejected; AI training can be a legitimate interest on summary review; DPC statement of 2025-05-21 gives 2025-05-27 as Meta's planned EU training start | nrwe.justiz.nrw.de; dataprotection.ie | Meta AI training on EU user data |
| 2024-10-04 | Court of Justice | KNLTB v Autoriteit Persoonsgegevens, C-621/22 | disclosure for consideration may rest on Art. 6(1)(f) if strictly necessary and not overridden | publications.europa.eu (62022CJ0621) | n/a |
| 2023-07-04 | Court of Justice (Grand Chamber) | Meta v Bundeskartellamt, C-252/21 | a set collected en bloc with one sensitive item needs an Art. 9(2) derogation | publications.europa.eu (62021CJ0252) | n/a |
| 2023-03-30 | Court of Justice | Hauptpersonalrat, C-34/21 | national employee rules that miss Art. 88(2) are disregarded | publications.europa.eu (62021CJ0034) | n/a |
The Garante's notice says the court upheld OpenAI's opposition, not whether the decision was set aside in whole or in part.
Two rows record positions, not decisions: the DPC report and the Meta and noyb statements.
Copyright claims run on a separate track, in the tracker of AI copyright lawsuits over training data.
Pending, Not Final: What COM(2025) 837 Would Change
The Commission's Digital Omnibus proposal COM(2025) 837 of November 19, 2025 would amend several GDPR rules on AI training, and as of November 9, 2026 the European Parliament lists it as "Awaiting committee decision", so it is not law. A joint ITRE-LIBE draft report followed on June 22, 2026 and amendments on July 27, 2026; the EDPB-EDPS Joint Opinion 2/2026 of February 10, 2026 is only an opinion.
The table sets each proposed change beside the current final rule, as of 2026-11-09 (procedure 2025/0360(COD)).
| Proposed change | Current final rule | Status | Source |
|---|---|---|---|
| Art. 4(1): sentence making personal data relative to the entity | definition unchanged; C-413/23 P case law | proposal | COM(2025) 837 |
| Art. 9(2)(k): residual special-category data in AI development and operation | Art. 9(2) exceptions only | proposal | COM(2025) 837 |
| Art. 88c: legitimate interest for AI development and operation | Art. 6(1)(f) | proposal | COM(2025) 837 |
| Art. 13(4): notice exemption for clear, circumscribed, non-data-intensive relationships, not where data go to recipients or third countries | Arts. 13, 14 | proposal | COM(2025) 837 |
| Art. 41a: pseudonymized data | none | proposal | COM(2025) 837 |
Every other rule on this page is final text or a judgment. Procedure detail sits on the page on the EU Digital Omnibus and its GDPR changes.
Benefits, Risks, Downsides and Alternatives for a Seller Under the GDPR, Including Not Selling
The GDPR lets a company license personal data for AI training when each step holds, but every step costs time, and licensing only non-personal data or not licensing at all remain valid choices.
Benefits:
- Disclosure for consideration can rest on Article 6(1)(f) without each person's consent (C-621/22).
- Anonymous data falls outside the GDPR, which affects what company data is worth to AI companies.
- Transfers to DPF-listed US buyers need no SCCs.
Risks:
- Fines up to EUR 20,000,000 or 4 % of total worldwide annual turnover, whichever is higher (Article 83(5), Regulation (EU) 2016/679).
- Erasure of the dataset or the model (Opinion 28/2024, paragraph 114).
- A breach during delivery must be notified within 72 hours where feasible (Article 33(1)).
Downsides:
- Notices to each person, or public information (Article 14(5)(b)).
- An impact assessment before high-risk processing (Article 35(1)).
- The pending appeal C-703/25 P can change the DPF route.
Alternatives:
- License only anonymous or non-personal data.
- Exclude special-category items and private messages.
- Wait for the outcome of COM(2025) 837, which may be adopted, amended or dropped.
- Do not sell or license the data at all.
Options including keeping the data are compared under alternatives to selling data to AI labs.
Beyond This Page: GDPR Articles in Depth and the Wider EU Rules
The GDPR is one layer of EU law on a data deal, and each article above has its own page, while copyright, contract and AI Act duties sit on the EU and seller pages. The full set of EU law for AI training data adds the AI Act and the Data Act.
- AI training data law by jurisdiction
- Sell or license company data to AI companies
- The buyer's duties: EU AI Act Article 53
- EU Data Act and AI training data
- Anonymization vs pseudonymization for AI training data
- PII removal and de-identification
- Employee notice and consent
- UK GDPR after the Data (Use and Access) Act 2025
- Switzerland rules for licensing data
- Privacy law and AI training data compared
How This Page Is Sourced and Dated
Every rule here cites the Official Journal text, a court judgment or a regulator's own page, with the date it was checked. DataPacts Editorial Team checked each item on November 9, 2026, following the rules on how we source, date and check every figure. This page is general information about the GDPR, not legal advice for any specific deal.
Frequently Asked Questions
Does the AI Act replace the GDPR for AI training data?
No: the AI Act (Regulation (EU) 2024/1689) adds duties for providers of general-purpose AI models from August 2, 2025, while the GDPR still governs any personal data in the training set. Legal basis, notices and transfers stay GDPR questions.
Can a company sell customer data for AI training under the GDPR?
Yes, if a legal basis, a compatibility test, notices and transfer safeguards all hold. Disclosure for consideration can rest on legitimate interests where strictly necessary (C-621/22), and special-category items also need an Article 9(2) exception.
Is an AI model trained on personal data anonymous?
Not in all cases. EDPB Opinion 28/2024 treats a model as anonymous only where extracting personal data and obtaining it through queries are both insignificant.
What fine can a GDPR breach in a data sale bring?
Up to EUR 20,000,000 or 4 % of total worldwide annual turnover, whichever is higher, under Article 83(5) of Regulation (EU) 2016/679. Fines can be challenged: in March 2026 the Tribunale di Roma upheld OpenAI's opposition to the Garante's decision.