DataPacts

Special Category Data (Article 9) in AI Training

Checked

At a glance

Under GDPR Article 9, personal data revealing racial or ethnic origin, political opinions, religious or philosophical beliefs or trade union membership, genetic data, biometric data for uniquely identifying a person, health data and data on sex life or sexual orientation may not be processed unless one of the ten Article 9(2) exceptions applies, in addition to an Article 6 legal basis, and this holds for a dataset licensed for AI training.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

Regulation (EU) 2016/679 entered into force on 2016-05-24 and applies from 2018-05-25 (Art. 99). The seller and the buyer each need an exception for their own processing.

As of 2026-11-17, this page covers the list, inferred data, the exceptions that fit a license, AI Act bias detection and follow-on duties, and keeps a pending proposal apart. It is the special-category layer of AI training data licensing.

Every rule below cites its source and date, as explained in how we source and date every rule.

The Article 9(1) List and Why One Sensitive Record Reaches the Whole Training Set

Article 9(1) names nine kinds of data, and the Court of Justice held that a set collected en bloc with at least one sensitive item is prohibited unless an Article 9(2) exception applies.

  • Data revealing racial or ethnic origin, political opinions, religious or philosophical beliefs or trade union membership
  • Genetic data (Art. 4(13))
  • Biometric data for uniquely identifying a person (Art. 4(14))
  • Data concerning health (Art. 4(15))
  • Data concerning sex life or sexual orientation

In Meta v Bundeskartellamt (C-252/21, Grand Chamber, 2023-07-04, para. 89), the Court held that a set containing sensitive and non-sensitive data, collected en bloc without separating the items at collection, is prohibited if it contains at least one sensitive item and no Article 9(2) derogation applies.

For a seller, the question is whether any record in the export holds such an item; free text, call audio and personnel files are places to check. Photographs are biometric data only when specifically processed for identification (Recital 51); the page on biometric data in AI training carries that depth.

Inferred and Indirect Sensitive Data: C-184/20, C-252/21 and C-21/23

Data that only allow a sensitive fact to be inferred are special-category data too, and the Court of Justice has said so three times. In OT (C-184/20, Grand Chamber, 2022-08-01), data liable to disclose sexual orientation indirectly fall under Article 9(1).

In C-252/21 (paras. 69 to 70) and C-21/23 (para. 87), the prohibition applies whatever the controller's aim and whether or not the information is correct.

In Lindenapotheke (C-21/23, Grand Chamber, 2024-10-04), order data for pharmacy-only medicines are health data even without a prescription, and even if it is only probable that the product is for the customer (operative point 2; paras. 83 and 90).

Data count where they can reveal health status "by means of an intellectual operation involving collation or deduction".

Read for AI, a training set with names beside purchases, diagnoses in support chats or membership lists can be special-category data even if no field is labeled "health". No Court of Justice ruling on Article 9 in AI training is recorded as of 2026-11-17.

Which Article 9(2) Exceptions Can Carry a Training Data License

Article 9(2) lists ten exceptions, and each is strict: the Court reads them narrowly (C-252/21, para. 76), and none replaces the Article 6 legal basis a seller and a buyer also need. That layer is usually legitimate interest (Article 6(1)(f)) for AI training data or consent.

The table compares the ten exceptions of Article 9(2) with a training data license, as of 2026-11-17, and its middle column is an editorial reading, not a holding.

Point Condition in the text (short) Bearing on an AI training data license (editorial reading) Source
(a) explicit consent for specified purposes, unless law bars lifting the ban main route where people can be asked; withdrawal possible (Art. 7(3)) eur-lex.europa.eu
(b) employment, social security and social protection law obligations, authorized by law or collective agreement tied to employer duties, not a sale eur-lex.europa.eu
(c) vital interests where the person cannot consent not a licensing purpose eur-lex.europa.eu
(d) not-for-profit bodies with a political, philosophical, religious or union aim; members only; no disclosure outside without consent bars disclosure to a buyer without consent eur-lex.europa.eu
(e) data manifestly made public by the data subject own public data only (C-252/21; OLG Köln interim) eur-lex.europa.eu; publications.europa.eu
(f) legal claims or courts acting judicially not a licensing purpose eur-lex.europa.eu
(g) substantial public interest on the basis of Union or member state law basis cited for AI Act Art. 4a bias detection (recital 9 of Reg. (EU) 2026/1744) eur-lex.europa.eu
(h) preventive or occupational medicine, diagnosis, care, health systems, under law or contract with a health professional, with Art. 9(3) secrecy care purposes, not training a third party's model eur-lex.europa.eu
(i) public interest in public health under law with specific safeguards public-health law, not a private license eur-lex.europa.eu
(j) archiving in the public interest, scientific or historical research or statistics, per Art. 89(1), under Union or member state law needs such a law; research status of training not settled (Guidelines 1/2026 draft) eur-lex.europa.eu

Three exceptions come up in AI data deals and get their own sections below. Points (b), (c), (d), (f), (g), (h) and (i) are tied to purposes a commercial license does not have (editorial reading).

Explicit Consent Under Article 9(2)(a) for Licensed Records

Explicit consent for one or more specified purposes lifts the ban (Art. 9(2)(a)), unless Union or member state law says the person cannot lift it. Consent must be freely given, specific, informed and unambiguous (Art. 4(11)), and withdrawal must be as easy as giving it (Art. 7(3)).

The German supervisory authorities (DSK, 2024-09-11) allow special-category customer data, such as health data, in an asset deal only with informed explicit consent (Art. 9(2)(a), Art. 7). Steps are in customer consent before licensing data for AI.

Consent given for a service rarely names AI training by a third party. A withdrawal reaches records already licensed, which buyer terms must handle, for example through a deletion clause. Member states may add conditions for genetic, biometric and health data (Art. 9(4)), not listed here.

Data "Manifestly Made Public" Under Article 9(2)(e): The Person's Own Choice Only

Article 9(2)(e) covers only data the person concerned has manifestly made public, so it never covers sensitive data about other people.

In C-252/21, the Court limited the exception to data made public by the data subject (para. 75), requires an intention "explicitly and by a clear affirmative action" to make data accessible to the general public (para. 77), and asks for a choice made beforehand, on settings selected with full knowledge, for an unlimited number of persons (para. 82; operative point 3). EDPB Opinion 28/2024 quotes this (para. 17).

In an interim ruling that rejected an injunction against Meta's training on public posts, the Higher Regional Court of Cologne (OLG Köln, 15 UKl 2/25, 2025-05-23) held that the exception covers data users put about themselves in public accounts or public posts (Rn. 114), not third parties' data in those posts (Rn. 115). The record is on the page about Meta AI training on EU user data.

For third-party data, the court assumed, subject to a possible Court of Justice reference in the main proceedings, that the Article 9(1) ban needed "activation" by the third party's request to remove the data (Rn. 116). That is one interim court's view and not settled law.

Research Under Article 9(2)(j): Union or Member State Law and Article 89(1) Safeguards

The research exception in Article 9(2)(j) is not self-executing: it needs a Union or member state law that is proportionate and sets specific safeguards, and processing must follow Article 89(1). Article 89(1) calls for safeguards, data minimization and pseudonymization where the purpose allows.

Whether training counts as scientific research is judged under the EDPB's Guidelines 1/2026, a consultation version and not final as of 2026-11-17. The six indicators sit with the research presumption beside purpose compatibility (Article 6(4)).

Bias Detection Under the AI Act: Article 4a Replaced Article 10(5)

The AI Act lets providers of high-risk AI systems process special categories exceptionally and only where strictly necessary for bias detection and correction, and since Regulation (EU) 2026/1744 that rule sits in a new Article 4a, while Article 10(5) was deleted. The regulation entered into force on 2026-07-27, and Article 4a(1) sets six conditions.

  • (a) Other data, including synthetic or anonymized data, would not work.
  • (b) Technical limits on re-use, state-of-the-art security and pseudonymization.
  • (c) Access controls and documentation.
  • (d) No transmission, transfer or other access by other parties.
  • (e) Deletion once the bias is corrected or the retention period ends.
  • (f) Records of processing state why it was strictly necessary.

Article 4a(2) extends the rule to other AI systems and models and to deployers, without creating a duty to run bias detection. Recital 9 calls bias detection a substantial public interest tied to GDPR Article 9(2)(g).

Because of condition (d), the rule supports a developer's own processing, not a sale of sensitive data to it (editorial reading). Model providers' EU AI Act Article 53 duties are a separate set of rules.

Duties That Follow Once Special Categories Are in the Set

A dataset with special categories brings extra GDPR duties for whoever processes it on a large scale, and higher fines for breaches of Article 9.

  • An impact assessment for large-scale processing of special categories (Art. 35(3)(b)).
  • A data protection officer where core activities are such processing (Art. 37(1)(c)).
  • A professional secrecy condition for health purposes under 9(2)(h) (Art. 9(3)).
  • As of 2026-11-17, fines up to EUR 20,000,000 or 4 % of worldwide annual turnover, whichever is higher, for breaches of the basic principles including Article 9 (Art. 83(5)(a)).

The impact assessment for a deal has its own page, the DPIA for an AI data licensing deal.

What Regulators, Courts and Buyers Have Written About Sensitive Items in Training Sets

Few authorities have addressed special categories in a sold or licensed training set, and the table lists every position recorded here, oldest first, as of 2026-11-17.

Date Author Document Position on special categories Status Source
2023-07-04 Court of Justice (Grand Chamber) C-252/21, Meta v Bundeskartellamt en bloc set with one sensitive item prohibited absent a derogation (para. 89); 9(2)(e) needs explicit choice (paras. 77, 82) judgment publications.europa.eu (62021CJ0252)
2024-09-11 DSK (German supervisory authorities) asset-deal resolution special-category customer data, such as health data, transfer only with informed explicit consent (Art. 9(2)(a), Art. 7) authorities' resolution datenschutzkonferenz-online.de
2024-11-27 Garante (Italy) provvedimento n. 741, warning to GEDI on its OpenAI content deal archives hold special-category and judicial data; processing could breach Arts. 9, 10, 13, 14 and Chapter III formal warning (art. 58(2)(a)) garanteprivacy.it
2024-12-17 EDPB Opinion 28/2024 quotes C-252/21 on 9(2)(e) (para. 17) adopted, soft law edpb.europa.eu
2025-05-23 OLG Köln 15 UKl 2/25 (interim) 9(2)(e) only for users' own public data; third-party data: ban assumed to need a removal request (Rn. 114 to 116) interim, injunction rejected nrwe.justiz.nrw.de
2026-11-12 (terms date; checked 2026-11-17) micro1 (buyer) Data License Master Terms, s.5.7 PHI and biometric identifiers excluded unless the Order Form allows; program page claims micro1 de-identifies sensitive information itself buyer's own terms (claim) micro1.ai

The rows show authorities asking for explicit consent or warning, and some buyers excluding the categories by contract; micro1 also states that it de-identifies sensitive information itself (claim, checked 2026-11-09).

SimpleClosure, a shutdown service that handles sales for closing companies and is not a buyer, states that AssetHub excludes sensitive employee, compensation, health and personal information (checked 2026-11-09). No fine for Article 9 in a training sale is recorded.

Pending, Not Final: Proposed Article 9(2)(k) for Residual Sensitive Data

The Commission's proposal COM(2025) 837 of 2025-11-19 would add an Article 9(2)(k) for special categories that remain in data used to develop and operate AI, and as of 2026-11-17 the European Parliament lists the procedure (2025/0360(COD)) as "Awaiting committee decision".

A separate proposed Article 9(2)(l) concerns biometric verification under the person's sole control, not training. Relying on the proposal has no legal effect today. The text and procedure are on the page about proposed GDPR changes in the EU Digital Omnibus.

Licensing a Set That May Hold Special Categories: Benefits, Risks, Downsides and Alternatives, Including Not Licensing

A dataset with sensitive items can sometimes be licensed, but each route carries cost and risk, and filtering the items out or not licensing remain valid choices.

Benefits

  • Explicit consent gives a clear Article 9(2)(a) route where people can be asked.
  • Removing sensitive records can take a set outside Article 9 (C-252/21).
  • Some buyers exclude the categories by contract, which narrows the check.

Risks

  • Inferred data count, so a label-based filter can miss them (C-21/23).
  • A withdrawal of consent reaches licensed records.
  • Article 9(2)(e) fails for third parties in public posts.
  • A model trained on the data may itself hold personal data; see when an AI model is anonymous.
  • As of 2026-11-17, fines up to EUR 20,000,000 or 4 % of worldwide annual turnover (Art. 83(5)(a)).

Downsides

  • Explicit consent is slow to collect at scale.
  • Filtering free text and audio takes time and lowers set value.
  • National conditions under Article 9(4) differ between member states.

Alternatives

  • Remove or de-identify the sensitive records before licensing.
  • License only records without special categories.
  • Route health data through the European Health Data Space permit system once it applies.
  • Not licensing at all, with the full list on the page about alternatives to selling data to AI labs, including not selling.

Beyond Article 9: Faces and Voices, Health Records and the Other GDPR Steps

Article 9 is one step among the GDPR rules for a training data deal. The parent page sets it beside the legal basis, notices and transfers: GDPR rules for AI training data, step by step.

How This Page Is Sourced and Dated

Every rule here cites the Official Journal text of Regulation (EU) 2016/679 or Regulation (EU) 2026/1744, a judgment by case number and date, or a regulator's or buyer's own document with the date it was checked. This page is general information about GDPR Article 9, not legal advice for any specific dataset or deal.

Frequently Asked Questions

Can health data be used to train AI under the GDPR?

Only with an Article 9(2) exception, such as explicit consent, on top of an Article 6 basis, and data that merely allow health to be inferred count (C-21/23, 2024-10-04).

Does sensitive data posted publicly fall under Article 9(2)(e)?

Only data the person made public themselves, by a clear choice to reach an unlimited audience (C-252/21, paras. 77, 82), and never data about others.

What if only some records in a dataset are sensitive?

A set collected en bloc with one sensitive item is prohibited without a derogation (C-252/21, para. 89), so sellers remove such records or find an exception.

Can AI developers process special category data to detect bias?

Yes, exceptionally, under AI Act Article 4a, inserted by Regulation (EU) 2026/1744, where strictly necessary and the data are not passed to other parties.