AskAjay.ai

Reference

10 min read · ·

The GDPR Was Not Written for AI. It Governs It Anyway

Articles 22, 9, 6, 25 and 35 read for AI, the cases regulators have brought, and the duties the EU AI Act adds.

Ajay Pundhir
Ajay PundhirAI Strategist & Speaker
Share
Six translucent grey bands stacked in staggered rows like layers of sediment. The fourth band from the top is solid dark blue, and a thin gold upright marks its left edge.
Art: AskAjay.ai house pattern, 'Strata' variant

Key Takeaways

  • →The GDPR applies to any AI system that processes personal data of people in the EU, though it never names AI. That includes companies outside the EU that offer them goods or services or monitor their behaviour.
  • →Since the Court of Justice ruled in SCHUFA (7 December 2023), a credit agency's score can itself be an automated decision under Article 22 when the lender that receives it draws strongly on it. The ruling concerned credit; its reasoning may reach other bought-in scores.
  • →What a model infers counts. Data that reveals health, ethnicity or sexual orientation indirectly is special-category data under Article 9.
  • →The regulator cases against AI and algorithmic systems in this reference turned on the lawful basis for the data or on the rights of the people decided about. Several large fines are under appeal, and Italy's EUR 15 million fine on OpenAI was annulled in March 2026.
  • →The EU AI Act sits on top of the GDPR and leaves it in force. For Annex III systems the deployer duties apply from 2 December 2027. The Commission's proposal to amend the GDPR for AI is still a proposal.

The GDPR never uses the words "artificial intelligence" or "algorithm". It still governs every AI system that processes personal data of people in the EU, and five of its articles decide most of what such a system may do: 22, 9, 6, 25 and 35.

My view is that the privacy notice and the consent screen get most of the attention, while the larger exposure sits in three places regulators keep finding beside them: the lawful basis for the data a model is trained and run on, what the model infers about people, and the decisions it shapes. Since December 2023, a credit score that a lender draws strongly on can itself be the decision, and the reasoning may carry to other scores. This reference is for the executive who signs off an AI system that touches EU customers or staff: general counsel, the chief data officer, or the product owner who carries the risk. It is a practitioner's reading of public texts, not legal advice.

Location does not help. Under Article 3(2), the GDPR reaches a company outside the EU that offers goods or services to people in the EU or monitors their behaviour there.

Article 22: When Does a Score Become the Decision?

Article 22(1) gives a person "the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning him or her or similarly significantly affects him or her." The Court of Justice reads it as "a prohibition in principle", as do the EDPB-endorsed guidelines (WP251rev.01). It lifts where the decision is necessary for a contract, authorised by law with safeguards, or based on explicit consent. Under the contract and consent routes the company must offer at least human intervention, a chance to give a view, and a way to contest (Article 22(3)).

The ruling that widened the article is SCHUFA (C-634/21), of 7 December 2023. A lender refused a loan after a credit agency's negative score. The Court held that the score itself is "automated individual decision-making" where a third party "draws strongly on that probability value" to make, run or end a contract with the person. A narrower reading, it said, would leave "a lacuna in legal protection". The Court ruled on a credit score, but its reasoning was general, and Italy's regulator relied on it in a case about scoring delivery riders. Treat a hiring or fraud score bought from a vendor as one that may be the decision in law.

WP251rev.01 says a company "cannot avoid the Article 22 provisions by fabricating human involvement", and that oversight must be "meaningful, rather than just a token gesture", carried out by "someone who has the authority and competence to change the decision". One quick diagnostic is to ask how many outcomes the reviewer changed last quarter. Zero proves nothing on its own, since an accurate model may rarely need correcting, but it should prompt the WP251 questions: can the reviewer overrule, do they see all the relevant data, and do they have time to weigh it?

People subject to such decisions are owed "meaningful information about the logic involved" (Articles 13(2)(f), 14(2)(g) and 15(1)(h)). In Dun & Bradstreet Austria (C-203/22), on 27 February 2025, the Court said this means explaining "the procedure and principles actually applied". A formula is not enough; telling the person how a change in their data would have changed the result may be. Claimed trade secrets go to the regulator or court, which decides how much the person sees.

Article 9: What the Model Infers Counts

Article 9(1) prohibits processing data revealing racial or ethnic origin, political opinions, beliefs or trade union membership, and genetic, identifying biometric, health and sex-life or sexual-orientation data, unless an Article 9(2) condition applies on top of an Article 6 basis.

The word that matters for AI is "revealing". In OT (C-184/20), on 1 August 2022, the Court held that data "liable to disclose indirectly" a person's sexual orientation is special-category data, including where the link takes "an intellectual operation involving deduction or cross-referencing". In Meta v Bundeskartellamt (C-252/21), on 4 July 2023, it applied the same reading to browsing and app data that "allows information falling within one of those categories to be revealed". A model that predicts a health condition from purchases, or ethnicity from postcode and name, is doing that deduction at scale. So the audit has to cover what the model can infer, as well as the columns that go in.

Bias testing often needs the very data Article 9 restricts. Since 27 July 2026, a new Article 4a of the AI Act, inserted by Regulation (EU) 2026/1744, lets providers of high-risk systems process special-category data where "strictly necessary" for bias detection and correction. Deployers of high-risk systems get the same permission. So do other AI systems and models, where the biases are likely to harm health, safety or fundamental rights. Six conditions apply, including that synthetic or anonymised data would not do and that the data is deleted after correction. It replaces the old Article 10(5).

Article 6: Which Lawful Basis Holds for Training and Use?

Article 6(1) lists six lawful bases. For commercial AI the live choice is usually consent, contract or legitimate interests, with no hierarchy among them (EDPB). The basis has to fit each stage: collection, training and use.

  • Consent. People can withdraw it at any time, and withdrawal does not make earlier processing unlawful (Article 7(3)). No ruling I have found says what withdrawal means for a model already trained on their data.
  • Contract. The Court allows it only where processing is "objectively indispensable for a purpose that is integral to the contractual obligation" (C-252/21). On that ground Ireland's regulator fined Meta EUR 390 million in January 2023 for behavioural advertising; Meta has appealed.
  • Legitimate interests. The EDPB's Opinion 28/2024 (adopted 17 December 2024) applies a three-step test to AI models: a legitimate interest, processing necessary to pursue it, and no overriding interest of the people whose data is used. It must be documented, and people can object under Article 21(1).

The same Opinion matters to anyone who buys a model. A regulator may order a company to erase a dataset or a model built on unlawfully processed data, subject to proportionality. A deployer is expected to check how the model it uses was developed. And a model is not anonymous just because its owner says so: the chance of extracting personal data from it, or of getting it out through queries, has to be "insignificant".

On 19 November 2025 the Commission proposed a "Digital Omnibus" that would add an Article 88c naming AI development and operation as a legitimate interest, and rewrite Article 22. On 6 October 2026 Parliament's procedure file showed it still in committee, with no vote taken, and nothing had been adopted. It is separate from the Digital Omnibus on AI, Regulation (EU) 2026/1744, which amends the AI Act and has been in force since 27 July 2026.

Articles 25 and 35: Design It In, Assess It First

Article 25 requires data protection by design and by default: measures such as pseudonymisation decided "at the time of the determination of the means for processing", and by default only the personal data "necessary for each specific purpose". For an AI team, feature selection, retention of training sets and access to embeddings are design decisions with legal weight, judged against a "state of the art" the EDPB's guidelines call a "dynamic concept".

Article 35 requires a data protection impact assessment "prior to the processing" wherever processing, "in particular using new technologies", is likely to create a high risk. It is mandatory for "a systematic and extensive evaluation of personal aspects" based on automated processing with legal or similarly significant effects. The EDPB-endorsed guidelines (WP248rev.01) give nine criteria, among them scoring, automated decisions, sensitive data, large scale, combined datasets and innovative technology. "In most cases" two of them mean an assessment is needed, and an AI system that scores people and decides about them already meets two. Article 35(11) requires a review when the risk changes, as after retraining that alters what the model does.

Breaches of Articles 25 and 35 fall in the lower fine tier, up to EUR 10 million or 2 percent of worldwide turnover (Article 83(4)). Lawful-basis, Article 9 and rights breaches (Articles 12 to 22) fall in the upper tier, up to EUR 20 million or 4 percent (Article 83(5)). The caps are ceilings, and regulators fine failures together: the Garante's single EUR 5 million fine on Foodinho covered Articles 6, 22(3), 25 and 35 among others. The DPIA is also where a company sets out the processing, any legitimate interest it relies on, necessity and risks (Article 35(7)), and WP251 expects it to record the degree of human involvement, so it is where the upper-tier questions are answered first.

Where Regulators Have Brought Cases

Cases against AI and algorithmic systems, each traced to the regulator's own decision or announcement:

CaseWhat it turned onRegulator, date, fine
Foodinho (Glovo), rider algorithms

Automated rider scores, order allocation and account blocking with no right to human review, to give a view or to contest (Article 22(3)). The decision cites SCHUFA.

Italy, Garante, decision of 13 November 2024, EUR 5 million

Credit card refusals, unnamed company

Credit card applications refused automatically despite good credit, and customers not properly informed or answered when they asked for their data. The company accepted the fine.

Hamburg regulator, reported 30 September 2025, EUR 492,000

Clearview AI, facial recognition

Scraping faces from the web to build a biometric database with no lawful basis, plus failures on transparency and people's rights. The Italian, Greek and Dutch decisions also found an Article 9 breach.

Italy, Greece and France, 2022, EUR 20 million each; Netherlands, 2024, EUR 30.5 million

OpenAI, ChatGPT

Training on personal data without first identifying a lawful basis, transparency, no age checks, and an unreported breach. The Court of Rome upheld OpenAI's challenge in judgment 4153/2026, published 18 March 2026.

Italy, Garante, decision of 2 November 2024, announced 20 December 2024, EUR 15 million

Luka Inc., Replika chatbot

No lawful basis identified, an inadequate privacy notice and no age checks. The Garante opened a separate inquiry into training in May 2025. Under appeal.

Italy, Garante, decision of 10 April 2025, EUR 5 million

Meta, behavioural advertising

Contract is no lawful basis for behavioural advertising, plus transparency. Under appeal.

Ireland, DPC, announced 4 January 2023, EUR 390 million

LinkedIn, behavioural analysis

Consent, legitimate interests and contract all failed as a basis for behavioural analysis and targeted advertising. Under appeal.

Ireland, DPC, announced 24 October 2024, EUR 310 million

These seven are a selection: AI or algorithmic systems with a published decision, not a census. Meta and LinkedIn concern advertising algorithms and are here for what they say about lawful basis. In six the regulator found a lawful-basis failure, in two people could not get a human review or a proper answer, and in all seven people were not properly informed. None needed a new law for AI.

One case outside AI matters to anyone running models on US infrastructure: Ireland's EUR 1.2 billion fine on Meta in May 2023 for transfers under Article 46(1), also under appeal. The EU-US Data Privacy Framework of 10 July 2023 now covers certified US companies. The General Court upheld it in Latombe (T-553/23) on 3 September 2025, and an appeal is pending (C-703/25 P).

Where the EU AI Act Adds Duties

The AI Act leaves the GDPR in force. Article 2(7), as amended, says the Act "shall not affect" it, without prejudice to Articles 4a and 59. A company deploying a high-risk system answers to both. Where they meet, by duty, with AI Act duties applying to Annex III systems from 2 December 2027 unless a row says otherwise:

DutyWhat the AI Act adds, and from whenUnder the GDPR
Impact assessment

A fundamental rights impact assessment for public bodies, private providers of public services, and deployers who score credit or price life and health insurance (Article 27). It may reuse DPIA sections.

A DPIA before high-risk processing (Article 35).

Human oversight

Oversight designed into the system (Article 14), assigned to people with "the necessary competence, training and authority" (Article 26(2)).

Human intervention, a chance to give a view and to contest, where the contract or consent exception is used (Article 22(3)).

Telling people

Disclose that people are dealing with an AI system unless it is obvious (Article 50(1)), since 2 August 2026. Deployers of Annex III systems tell people they are subject to them (Article 26(11)).

Tell people about automated decision-making and its logic (Articles 13 to 15).

Explanation

An explanation of the AI system's role and the main elements of the decision (Article 86), where Union law does not already give one. The text dates it from 2 August 2026, but it rests on the Annex III classification that applies from 2 December 2027, so the start is contested.

"Meaningful information about the logic involved" (Article 15(1)(h)), read in Dun & Bradstreet Austria.

Special-category data

May be processed for bias detection and correction under six conditions (Article 4a). In force since 27 July 2026.

Prohibited unless an Article 9(2) condition applies.

High-risk duties for AI built into regulated products under Annex I apply from 2 August 2028. The EU AI Act After the Omnibus covers the Act in full, and the Regulation Tracker keeps the dates current. For a non-EU comparison, The AI Law the UAE Doesn't Have (Yet) sets the UAE's rule on automated decisions against Article 22.

What to Do This Quarter

  1. List every system whose output someone relies on to decide about a person. Credit, hiring, pricing, fraud blocks, eligibility. After SCHUFA, include the scores you buy from vendors. The vendor risk reference covers the contract side.
  2. Record the lawful basis for each stage. Collection, training and use, separately. Add the Article 9 condition wherever the model could infer a special category, and test its outputs as well as its inputs.
  3. Check the human review. Name who can change an outcome, and count how often they have. If they never have, check their authority, data and time before you rely on the review.
  4. Write the explanation before anyone asks. Which data was used, how, and what change would have altered the result.
  5. Run the impact assessment before launch, and again after a change in risk. For Annex III systems, build it so the fundamental rights assessment can reuse it. If you deploy someone else's model, ask how its training data was obtained. If you downloaded its weights, read the licence too: Open Weight Is Not Open Source, and Nobody Read the Licence covers what those licences ask.

Not legal advice. Confirm specifics with counsel.

Questions readers ask

Does a human review step take an AI decision outside GDPR Article 22?

Only if the review is real. The Article 29 Working Party's guidelines on automated decision-making, endorsed by the EDPB, say a company "cannot avoid the Article 22 provisions by fabricating human involvement" and that the reviewer needs "the authority and competence to change the decision". The Court of Justice went further in SCHUFA: a credit score counts as the decision when the lender "draws strongly" on it. In WP251's words, someone who routinely applies the system's output "without any actual influence on the result" leaves it a decision based solely on automated processing.

Can a company train an AI model on personal data under legitimate interests?

It can, if it passes the three-step test the EDPB set out for AI models in Opinion 28/2024: a legitimate interest, processing that is necessary for it, and no overriding interest of the people whose data is used. The test has to be documented, and people keep the right to object under Article 21(1). The EDPB's draft guidelines on web scraping for generative AI (July 2026, open for comment to 30 October 2026) apply the same test to scraped training data. The Commission's proposal to name AI development as a legitimate interest in a new Article 88c is not law.

Does a GDPR DPIA count for the EU AI Act?

In part. Deployers of high-risk systems must use the provider's instructions to carry out their DPIA (AI Act Article 26(9)). Some deployers, including public bodies and those scoring credit or pricing life and health insurance, also owe a fundamental rights impact assessment under Article 27. Since 27 July 2026, Regulation (EU) 2026/1744 lets that assessment cross-refer to the DPIA or reuse its sections. For Annex III systems both duties apply from 2 December 2027.


Ajay Pundhir
Ajay Pundhir

Senior AI strategist helping leaders make AI real across four continents. Forbes Technology Council member, IEEE Senior Member.

Let's Talk

Ajay's views, from fifteen-plus years in the field. Not legal or compliance advice. See full disclaimers →
Published by AI Exponent LLC

Get The Brief

Original thinking on AI strategy, governance, and durability. Every other Tuesday.