GDPR and AI: the 5 articles that actually matter

GDPR has 99 articles. When you ship an AI feature, five carry almost all the weight: Article 6 (legal basis), Articles 13–14 (transparency), Article 22 (automated decisions), Article 25 (privacy by design), and Article 35 (DPIA). The other 94 still apply. The fines cluster around these five.

Three things shifted between 2024 and 2026. The CJEU ruled in CK v Dun & Bradstreet Austria (C-203/22) that the algorithmic formula on its own does not satisfy the right to explanation. The CNIL's June 2025 final recommendations made legitimate interest the practical default for AI training. And the Court of Rome annulled the Italian Garante's €15 million fine against OpenAI on 18 March 2026, with the full reasoning still unpublished as of April 2026. Each move changed how one of the five articles below operates in practice.

This is a reference page. Each section explains one article as it lands on an AI feature in 2026, with the case law that has actually moved.

Article 6: the legal basis you'll actually pick

GDPR has six legal bases. Three matter for AI features. Most teams pick legitimate interest.

Legitimate interest (Article 6(1)(f))

The CNIL finalised its position on 19 June 2025: legitimate interest is the appropriate basis for almost all AI training. The catch is the three-step test, and the CNIL has been explicit about what it expects to see documented.

1. A specific interest. "Improving our AI" fails. "Building a customer-support assistant that triages incoming tickets in our SaaS product" passes.
2. Necessity. Personal data must be required to achieve the stated interest. If synthetic or pseudonymised data gives you the same model quality, the necessity test fails.
3. Balancing. Data subjects' rights cannot be overridden. The CNIL's June 2025 focus sheet on web scraping adds concrete weight: respecting robots.txt and explicit opt-outs, excluding sites that have objected to scraping, deleting irrelevant collected data on a defined schedule, and offering a prior opt-out mechanism are all enumerated as expected mitigations.

I think legitimate interest is now the practical default for AI training in the EU, and the CNIL recommendations are the most useful operational guidance any EU regulator has put out. The Italian Garante's December 2024 €15 million fine against OpenAI was based partly on a missing legal basis for training. That fine was annulled by the Court of Rome on 18 March 2026; the full reasoning has not yet been published, so the precedential weight of the reversal is genuinely uncertain.

Consent (Article 6(1)(a))

Consent is poorly suited to AI training. It must be freely given, specific, informed, and revocable, and revocation from a trained model is technically near-impossible. For inference (the user prompts the system in real time), opt-in consent can work if it is genuine. Buried in terms of service, it is not.

Contract performance (Article 6(1)(b))

Works when the AI is the thing the customer signed up for: an AI writing assistant, an AI code review tool, an AI meeting summariser. It does not cover "we used your data to train our next model for other customers." Different purpose, different basis.

Articles 13–14: what "meaningful information about the logic" requires after Dun & Bradstreet

When you collect personal data from someone (Article 13) or obtain it from another source (Article 14), you must tell them specific things. For automated processing, that includes "meaningful information about the logic involved." The CJEU narrowed what counts as meaningful in CK v Dun & Bradstreet Austria (C-203/22), decided 27 February 2025.

The question was: what does a credit reference agency owe a person whose mobile contract was refused based on an automated credit assessment? The court answered:

• The algorithmic formula on its own does not satisfy the right.
• A detailed step-by-step description of the entire automated process does not satisfy it either, if the result is not concise and intelligible.
• What is required is a practical explanation that lets the individual understand why this specific decision was made about them.
• Trade secrets cannot be used as a blanket reason to deny access. If the controller objects, it must hand the disputed material to the supervisory authority or court, which then balances the rights case by case.

For training notices specifically (Article 14, when the data was scraped or repurposed), the CNIL's web scraping focus sheet requires advance information about the purpose and the right to object. You do not have to disclose every implementation detail. You do have to say what you took, why, and how to opt out.

Article 22: when human review stops being a defense

Article 22(1) gives individuals the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects.

Two conditions, both required:

1. The decision is solely automated. A human rubber-stamping the output without genuine review still counts as solely automated. The review must be substantive and capable of changing the outcome.
2. The decision produces legal or similarly significant effects. Credit denial, insurance pricing, hiring decisions, benefits eligibility, access to essential services, and exclusion from work opportunities all qualify. Most product recommendations and content personalisation do not.

The 2023 SCHUFA ruling (C-634/21, 7 December 2023) expanded the reach. SCHUFA only calculated the credit score; the bank made the lending decision. The CJEU held that when the bank rejects "in almost all cases" where the score is poor, the scoring itself counts as the Article 22 decision. That dramatically widened the article's scope. Profiling and scoring relied on by downstream decision-makers now sit inside it.

When the safeguards apply, you must offer:

• the right to obtain human intervention (a real person who can reach a different result)
• the right to express a point of view
• the right to contest the decision

The Italian Garante fined Foodinho €2.6 million in 2021 for an algorithmic rider-management system that excluded couriers from work without any avenue to contest. It was the first Article 22 enforcement in employment. It remains the cleanest example of what "no genuine human in the loop" looks like in practice.

Article 25: privacy by design inside an AI pipeline

Article 25 has two prongs. By design: technical and organisational measures must be embedded into the system from the architectural level, not patched in after launch. By default: only the personal data necessary for each specific purpose is processed, and the most privacy-protective settings are the defaults the user sees first.

For an AI pipeline this typically lives in four architectural choices:

• Pseudonymise where the model can tolerate it. Most embedding and retrieval steps work fine on tokenised customer IDs.
• Minimise the prompt context. Sending the whole user record when only the last three orders are needed is the most common Article 25 failure in production AI pipelines, and the easiest to fix.
• Set retention by default. If conversation logs and vector store entries sit forever because nobody set a TTL, the burden falls on you to justify why that is "necessary".
• Treat the trained model as not anonymous unless you have proven otherwise. The EDPB's Opinion 28/2024 (December 2024) is the framework. The likelihood of direct extraction and extraction via queries must both be insignificant for the model to qualify as anonymous. Most production models do not meet that bar without specific mitigations. The companion article on vector embeddings as personal data walks through the technical evidence.

Article 35: the DPIA you almost certainly owe

A DPIA is required before processing begins when the processing is "likely to result in a high risk to the rights and freedoms of natural persons." The EDPB published nine criteria. Hitting two or more usually makes a DPIA mandatory.

1. Evaluation or scoring (profiling, prediction)
2. Automated decision-making with legal or significant effects
3. Systematic monitoring
4. Sensitive data (Article 9 special categories)
5. Large-scale processing
6. Matching or combining datasets
7. Vulnerable subjects (children, employees, patients)
8. Innovative use of new technologies
9. Processing that prevents the exercise of a right or use of a service

Most AI features hit at least three: innovative technology, evaluation or scoring, large-scale processing. If your feature profiles users or makes decisions about them, four or five.

A DPIA must contain a systematic description of the processing and its purposes, an assessment of necessity and proportionality, an assessment of risks to data subjects' rights, and the measures planned to address those risks. The practical DPIA check walks through the decision in more detail.

The Dutch DPA fined Clearview AI €30.5 million in September 2024. Among the violations: no DPIA for large-scale processing of biometric data. The Clearview decision is now the reference fine for "shipped sensitive AI without an impact assessment."

Under the EU AI Act, high-risk systems also require a Fundamental Rights Impact Assessment. The FRIA complements the DPIA, it does not replace it. Public bodies, private entities providing public services, and deployers using AI for credit scoring or insurance pricing will need both.

The five articles, mapped to a use case

Use case	Art. 6	Art. 13–14	Art. 22	Art. 25	Art. 35
Chatbot (info only)	LI or contract	Disclose AI, explain logic	Usually no	Minimise data	If at scale
Recommendation engine	LI (careful balancing)	Disclose personalisation logic	If significant effects	Opt-in defaults	Yes (profiling + scale)
RAG system	LI for corpus data	Inform about data in knowledge base	If outputs drive decisions	Pseudonymise corpus	If personal data at scale
AI agent (tool use)	LI per data category	Disclose agent capabilities	Likely yes (autonomous actions)	Least privilege access	Almost certainly
Credit / insurance scoring	LI or consent	Explain scoring logic (C-203/22)	Yes (SCHUFA)	Privacy-preserving scoring	Mandatory
HR / recruitment AI	LI (document carefully)	Explain AI involvement and criteria	Yes for rejected candidates	Bias testing, human review	Mandatory

LI = legitimate interest.