Under Article 36, the supervisory authority must be consulted first where a high risk remains.
Regulation (EU) 2016/679 applies from May 25, 2018. The impact assessment is one step in the privacy layer of AI training data licensing. This page covers when it is required, who runs it, what it contains, prior consultation and templates, as of 2026-11-17.
When Article 35 Requires an Impact Assessment Before Data Is Licensed
An impact assessment is required whenever processing, in particular with new technologies, is likely to result in a high risk, and Article 35(3) names three cases where it is always required. One assessment may cover similar operations with similar high risks (Art. 35(1)).
- (a) systematic and extensive evaluation based on automated processing, including profiling, with legal or similarly significant effects
- (b) processing on a large scale of the special categories in Article 9(1) or of criminal conviction data in Article 10
- (c) large-scale systematic monitoring of a publicly accessible area
For a deal, a licensed set with health, biometric or other Article 9 items at scale falls under case (b) for both sides, as the page on special category data (Article 9) in AI training. Most company exports fall under the general Article 35(1) test, which the WP248 criteria below put into practice (editorial reading).
The Nine WP248 Criteria Read Against a Training Data License
The Article 29 Working Party's DPIA guidelines (WP248 rev.01, revised October 4, 2017) list nine criteria for high-risk processing, and the EDPB endorsed them on May 25, 2018. The table compares the nine criteria with a training data license, as of 2026-11-17, and its fourth column is an editorial reading.
| No. | Criterion (WP248 rev.01) | What the guidelines include | Bearing on a training data license (editorial reading) | Source |
|---|---|---|---|---|
| 1 | evaluation or scoring | profiling and predicting work performance, economic situation, health, preferences, behavior | applies where the buyer's model scores people | ec.europa.eu |
| 2 | automated decisions with legal or similar significant effect | Art. 35(3)(a) | rare in training; relevant to the buyer's deployment | ec.europa.eu |
| 3 | systematic monitoring | observing or controlling people, including publicly accessible areas | call or screen recordings collected by monitoring | ec.europa.eu |
| 4 | sensitive data or data of a highly personal nature | Art. 9 and Art. 10 data; also personal documents, emails, diaries, notes | email, chat and document exports | ec.europa.eu |
| 5 | data processed on a large scale | number of people, volume and range of items, duration, geographic extent | whole-company exports | ec.europa.eu |
| 6 | matching or combining datasets | from different purposes or controllers beyond reasonable expectations | the buyer combines the set with other sources | ec.europa.eu |
| 7 | data concerning vulnerable data subjects | children, employees, patients, others in an imbalance | staff communications, health records | ec.europa.eu |
| 8 | innovative use or new technological or organizational solutions | Art. 35(1), recitals 89 and 91 | training an AI model | ec.europa.eu |
| 9 | processing that prevents exercising a right or using a service or contract | Art. 22, recital 91 | rare in a license | ec.europa.eu |
WP248 says that "in most cases" a processing meeting two criteria requires a DPIA, that the more criteria are met the more likely the risk is high "regardless of the measures" planned, and that in some cases one criterion suffices.
An export of emails, tickets or chats for a third party's model typically touches criteria 4, 5, 6 and 8, and criterion 7 when staff data is included (editorial reading).
National Lists Under Article 35(4) and the CNIL's AI Sheet
Every supervisory authority must publish a list of processing that requires a DPIA (Art. 35(4)) and may publish a list of processing that does not (Art. 35(5)), using the EDPB consistency mechanism where processing concerns several member states (Art. 35(6)).
The French CNIL adopted its list in Délibération n° 2018-327 of October 11, 2018 (14 types, announced November 6, 2018), and the list is not exhaustive, so processing outside it can still need a DPIA under the WP248 criteria. Other member states' lists are not covered.
CNIL how-to sheet 5 of its AI recommendations (English summary dated 2026-01-05) "Carrying out a data protection impact assessment when necessary", says processing that meets at least two of the nine EDPB criteria "should be presumed" to require a DPIA, five of them especially relevant to development: sensitive data, large scale, vulnerable persons, crossed or combined datasets, and new technological solutions or innovative use.
A DPIA "must" be carried out where significant risks exist (misuse, breach, discrimination) even below two criteria, and is presumed necessary for AI Act high-risk systems using personal data. The guidance is non-binding, and the page on the CNIL guidance on AI training data covers it in full.
Seller and Buyer: Who Assesses Which Processing
Each controller assesses its own processing, so a seller that discloses personal data for AI training and a buyer that trains on it each need to decide whether their own operation requires a DPIA.
The controller carries out the assessment and seeks the advice of the data protection officer where one is designated (Art. 35(2)). WP248 adds that the controller stays responsible even if someone else performs it, and that joint controllers must set out who is responsible for each measure.
On the buyer side, CNIL sheet 5 recommends a general DPIA for the entire processing where the provider also controls deployment and knows the operational use; where that use is unclear, the provider assesses only development.
Developing a foundation model or general-purpose system requires a DPIA when personal data are processed, and sharing it may help users comply.
On data acquired from a third party, the CNIL says the holder that did not collect it for AI training runs the compatibility test, the re-user checks the set is not manifestly unlawful, and an agreement between them is recommended.
For the seller, the practical reading is to ask the buyer for a DPIA summary or a statement that one was done, and to record its own decision, including a decision that no DPIA was needed, because EDPB Opinion 28/2024 lists such decisions among the documents an authority checks (editorial reading). Both sides' checks are part of due diligence in an AI data deal.
What the Assessment Must Contain
Article 35(7) sets four minimum elements, and WP248 Annex 2 turns them into a checklist a regulator can use to judge whether a DPIA is complete. The table compares them with what a deal adds, as of 2026-11-17 (fourth column: editorial reading).
| Element | Article | What WP248 Annex 2 checks | Deal-specific item to record (editorial reading) | Source |
|---|---|---|---|---|
| Systematic description of the processing and purposes, incl. the legitimate interest pursued | 35(7)(a) | nature, scope, context, purposes; data, recipients, storage period; assets | which records go to which buyer, for which training purpose, for how long | legislation.gov.uk; ec.europa.eu |
| Necessity and proportionality | 35(7)(b) | purpose, lawfulness (Art. 6), minimization, storage limits, rights, processors, Chapter V transfers, prior consultation | why each field is needed; transfer tool | legislation.gov.uk; ec.europa.eu |
| Risks to rights and freedoms | 35(7)(c) | sources, impacts, threats, likelihood and severity of illegitimate access, modification, disappearance | extraction from the model, misuse of the set (CNIL list) | legislation.gov.uk; ec.europa.eu; cnil.fr |
| Measures to address the risks | 35(7)(d) | measures determined per risk | de-identification, access limits, deletion terms | legislation.gov.uk; ec.europa.eu |
| DPO advice and people's views | 35(2), 35(9) | advice sought; views sought where appropriate | staff or customer representatives where appropriate | legislation.gov.uk; ec.europa.eu |
| Review | 35(11) | periodic review, at least when risk changes | new buyer, new purpose, new data | legislation.gov.uk; ec.europa.eu |
WP248 calls the DPIA "a continual process, not a one-time exercise", and Article 35(11) requires a review at least when the risk changes. Publication is not required, though a summary "could foster trust" (WP248).
AI-Specific Risks and Measures Regulators Expect in the Assessment
The CNIL lists eight risks specific to AI that a DPIA should take into account, and several start in the training set a seller delivers.
- confidentiality of data extractable from the system
- misuse of the training set by staff or after a breach
- automated discrimination from bias introduced in development
- false content about a real person, in particular with generative AI
- automated decisions driven by automation or confirmation bias
- loss of control over data published online
- AI-specific attacks such as data poisoning
- systemic and serious ethical risks of deployment
CNIL example measures are security (homomorphic encryption, secure execution environments), minimization (synthetic data), anonymization or pseudonymization (differential privacy) and protection from the outset (federated learning).
EDPB Opinion 28/2024 calls DPIAs "an important element of accountability" for AI models and leaves them outside its own scope (para. 17), yet it lists "any information relating to DPIAs" among the documents an authority checks when a model is claimed to be anonymous (para. 58(a)).
The full test is on the page on when an AI model is anonymous under EDPB Opinion 28/2024.
Prior Consultation Under Article 36 When High Risk Remains
A controller must consult the supervisory authority before processing when its DPIA shows a high risk that its own measures do not mitigate (Art. 36(1)).
If the authority considers that the processing would infringe the GDPR, it gives written advice within up to eight weeks of the request, and it may extend that period by six weeks because of the complexity of the processing.
It must tell the controller of any extension, with reasons, within one month of receiving the request, and the periods can be suspended until the authority has the information it asked for (Art. 36(2)).
The controller's request must include the following (Art. 36(3)):
- the responsibilities of the controller, joint controllers and processors
- the purposes and means of the processing
- the measures and safeguards provided
- the contact details of the data protection officer
- the data protection impact assessment itself
- any other information the authority requests
Consultation comes first, so a seller with a high residual risk cannot disclose, and a buyer cannot train, until it ends (editorial reading). WP248 says the full DPIA goes to the authority. The steps that come next are covered under delivering licensed data to an AI buyer.
Impact Assessments in Regulators' AI Training Files
Three regulator records on AI training name the impact assessment, and the table lists them oldest first, as of 2026-11-17.
| Date | Authority | Matter | What it says about the DPIA | Status | Source |
|---|---|---|---|---|---|
| 2024-09-12 | Data Protection Commission (Ireland) | inquiry into Google Ireland Limited, PaLM 2, s.110 Data Protection Act 2018 | whether an Art. 35 assessment preceded processing EU/EEA personal data | inquiry announced; no reference number public; no decision recorded | dataprotection.ie |
| 2024-11-27 | Garante (Italy) | provvedimento n. 741, warning to GEDI on its OpenAI content deal (2024-09-24) | the DPIA did not adequately analyze the legal basis | formal warning (art. 58(2)(a)) | garanteprivacy.it |
| 2025-05-21 | Data Protection Commission (Ireland) | statement on Meta AI training | legitimate interest assessment, DPIA and compatibility assessment listed among measures reviewed | statement, not an approval | dataprotection.ie |
The Irish inquiry asks whether an assessment was done at all, with no decision recorded. The Italian authority found that a publisher's assessment did not adequately analyze the legal basis for licensing its archive, and issued a warning, not a fine. The Meta statement is not an approval.
Regulator Templates and the Pending EU-Wide Lists
The GDPR prescribes no method for a DPIA, only the Article 35(7) elements, so regulators publish their own frameworks.
WP248 Annex 1 lists examples from authorities: Germany's Standard Data Protection Model (2016), the Spanish AEPD's EIPD guide (2014), the CNIL's PIA (2015) and the UK ICO's 2014 PIA code of practice. The CNIL offers open-source PIA software and three AIPD guides (method, models, knowledge bases).
At EU level, the EDPB consulted on a "Template for Data Protection Impact Assessment" from 2026-04-14 to 2026-06-09, and its consultations page lists no final version as of 2026-11-17. The Commission's proposal COM(2025) 837 of November 19, 2025 would replace Article 35(4) to (6) so that the EDPB prepares proposals for EU lists and a common template and methodology, and the European Parliament lists procedure 2025/0360(COD) as "Awaiting committee decision" (re-checked 2026-11-17).
A proposal has no legal effect; see proposed GDPR changes in the EU Digital Omnibus.
Assessing Before Licensing: Benefits, Risks, Downsides and Alternatives, Including Not Licensing
A DPIA costs time before any license fee arrives, but skipping one where it is required is itself a breach, and licensing less or nothing remain valid choices.
Benefits
- shows the risk and the measures before data leaves the company (Art. 35(7))
- documents a decision an authority may ask for
- can be shared in summary with a counterparty
Risks
- under Article 83(4)(a) as adopted in 2016, fines up to EUR 10,000,000 or 2 % of total worldwide annual turnover, whichever is higher, for breaches of Articles 25 to 39
- a regulator can open an inquiry into a missing DPIA, as the Irish authority did for PaLM 2
- a DPIA that does not analyze the legal basis is cited against the deal, as in the Italian warning
Downsides
- prior consultation can add up to 14 weeks, plus any suspension
- review whenever the risk changes
- national lists differ between member states
Alternatives
- license only anonymous or non-personal data
- remove special category and staff records to drop criteria 4 and 7 (editorial reading)
- narrow the set to a small sample
- not licensing at all, as set out in alternatives to selling data to AI labs, including not selling
Beyond the Impact Assessment: The Other GDPR Steps in a Data Deal
The impact assessment is one of eight GDPR steps in a training data deal. The sequence is set out in GDPR rules for AI training data, and related pages follow.
- Legal basis: legitimate interest (Article 6(1)(f)) for AI training data
- Reuse test: purpose compatibility (Article 6(4))
- Telling people: GDPR notices before a data sale
- Staff records: employee data under GDPR Article 88
- Transfers, where the transfer impact assessment is a different test: transferring training data to US buyers
- Risk-lowering measures: PII removal and de-identification before licensing
- Regulator records: Data Protection Commission oversight of AI training and the GEDI and OpenAI archive deal
- Other regimes: UK GDPR after the Data (Use and Access) Act 2025 and, for a privacy impact assessment outside the EU, Quebec Law 25 and AI training data
How This Page Is Sourced and Dated
Every rule here cites the Official Journal text of Regulation (EU) 2016/679, the WP248 rev.01 guidelines as endorsed by the EDPB, or a regulator's own document with the date it was read.
See how we source and date every rule. This page is general information about GDPR Articles 35 and 36, not legal advice for any specific dataset or deal.
Frequently Asked Questions
Is a DPIA part of the GDPR?
Yes, Article 35 of Regulation (EU) 2016/679, applicable since May 25, 2018, requires an impact assessment before high-risk processing, and Article 36 adds consultation of the authority where high risk remains.
When must a DPIA be conducted for AI training data?
Before the processing starts, whenever it is likely to be high risk: always in the three Article 35(3) cases, and in most cases where two WP248 criteria are met.
Does a company selling data need its own DPIA if the buyer has one?
Usually yes for its own disclosure, because each controller assesses its own processing; a buyer's DPIA covers the buyer's training, and joint controllers split responsibility.
How long does prior consultation with a supervisory authority take?
Up to eight weeks for written advice, extendable by six weeks, and suspended while the authority waits for information it requested (Art. 36(2)).