DPIA for acquiring an AI training dataset: a buyer-side template

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

When a licensed training dataset triggers a DPIA

A DPIA is mandatory when processing "is likely to result in a high risk" to individuals, and Article 35(1) names new technologies as a factor to weigh [2]. Article 35(3) adds cases that always need one, including large-scale processing of special categories of data [2]. The Article 29 Working Party guidelines (WP248 rev.01, endorsed by the EDPB) set out nine screening criteria, among them sensitive data, large scale, matching or combining datasets, vulnerable data subjects and innovative technology, and treat processing that meets two of them as usually needing a DPIA.

A licensed corpus for model training almost always hits "innovative use" and "large scale", and frequently "matching or combining datasets" when you merge it with your own logs or other licensed sets. Article 35(4) lets each supervisory authority publish a list of processing that requires a DPIA [2], so check your lead authority's list; several name AI or new technology combined with another criterion.

Record the screening decision even when you conclude no DPIA is needed. A documented screening supports the Article 5(2) accountability duty [2].

What the buyer must document that the supplier cannot

The buyer's DPIA covers the buyer's processing, so the supplier's privacy paperwork is evidence for it, not a substitute. Article 35(7) sets the minimum content: a systematic description of the processing and purposes, an assessment of necessity and proportionality, an assessment of risks to data subjects, and the measures to address them [2]. Only the buyer can describe the training purpose, the model architecture, who will query the model, and whether weights will be released.

EDPB Opinion 28/2024 makes three points that shape AI DPIAs [1]:

• Model anonymity is case by case. A model trained on personal data is anonymous only if the likelihood of extracting personal data, directly or via queries, is insignificant, and the controller must be able to show that [1]. See our breakdown of EDPB Opinion 28/2024 for data buyers.
• Legitimate interest needs the three-step test. Purpose, necessity and balancing must be documented, including data subjects' reasonable expectations [1]. Our page on documenting legitimate interest for licensed personal data covers the balancing record.
• Unlawful upstream processing can follow the model. The opinion addresses how unlawful processing during development can affect later deployment [1], so the DPIA should record what you checked about the supplier's collection.

Summaries of the opinion list source selection, data minimization, governance, testing for extraction and documentation as measures supervisory authorities will look for [3]. Treat that list as the spine of your mitigation section. Later commentary also discusses developments after the opinion that bear on training assessments [4].

Supplier evidence to attach to the DPIA

Attach the supplier's de-identification and provenance evidence as annexes, and record what each one does and does not prove. A redaction report proves which fields were transformed; it does not prove free-text notes are clean. A HIPAA Expert Determination letter covers a specific recipient and environment and may not travel with your use.

Request these as a minimum:

1. Data inventory and schema: table and field names, record counts, date ranges, and which fields hold direct identifiers (name, email, phone, account number) or quasi-identifiers (ZIP, birth date, job title).
2. Collection context: the original purpose, the notices or consents shown, and the supplier's lawful basis for disclosing to you. See consent and notice records.
3. De-identification method and results: tools used, entity types targeted, sample QA results with miss rates. Our de-identification evidence package checklist lists the documents.
4. Rights attestation: ownership and consents, signed. Use a data rights attestation template.
5. Transfer route: if EU or UK personal data sits in a US-held dataset, the mechanism (adequacy, DPF certification, SCCs) and the transfer impact assessment. See transfer mechanisms for US-sourced data.

NIST SP 800-188 cautions that traditional de-identification has inherent limitations compared with formal privacy methods such as differential privacy [5]. Your DPIA should say which kind the supplier used and quantify residual re-identification risk rather than assume Recital 26 anonymity [2].

The DPIA template, section by section

The template below follows Article 35(7) and adds the AI-specific fields that Opinion 28/2024 and WP248 point to [1][2]. Fill it per dataset, and per material change in purpose.

Illustrative example: invented to show structure; it does not describe an available dataset.

Risks specific to training on licensed records

The risks a DPIA must score for training differ from those for ordinary analytics, because the model can carry personal data forward after the dataset is deleted. Score at least these:

• Memorization and extraction. Rare strings, such as account numbers or unique signatures in email bodies, can be regurgitated. Document canary insertion or extraction testing; see memorization and extraction risk.
• Redaction misses in free text. Agent notes, transcripts and attachments defeat field-level masking. Record the measured miss rate, not the tool's marketing figure; see PII redaction pipelines and how to measure misses.
• Linkage. Combining a licensed set with your own CRM or product telemetry can re-identify pseudonymized records, which is the WP248 "matching or combining" criterion.
• Purpose creep. A dataset licensed for an internal triage model reused for a public chatbot changes the risk and may exceed the license; tie the DPIA purpose to the license's permitted uses, as set out in your data license term sheet.
• Rights handling. Erasure and objection requests may reach you after training; the DPIA should state whether you can filter, retrain or neither.

For high-risk AI systems under the EU AI Act, Article 10 adds data governance duties for training, validation and testing sets, including examination for biases [6]. As of October 2026, Regulation (EU) 2026/1744 reportedly moved the Annex III high-risk start date to 2 December 2027. If you place a general-purpose model on the EU market, the AI Office's public training-content summary template, published 24 July 2025, also asks you to describe data sources [7], so keep your DPIA source descriptions consistent with that summary.

Keeping the DPIA current through the license term

A DPIA is a living document: Article 35(11) requires review when the risk represented by the processing changes [2]. Re-open it when you receive a new data delivery, add a supplier, change the model's deployment surface, decide to release weights, or receive an incident report such as personal data found in a supposedly de-identified file. Our response playbook for personal data found in a licensed dataset shows what to log when that happens.

Watch the legal baseline too. As of October 2026, the GDPR amendments in the Commission's Digital Omnibus proposal remain proposals, not law [8], so EDPB Opinion 28/2024 is still the reference for assessing models [1].

How SourceX supports the evidence side of a DPIA

SourceX sources operational datasets from US companies on request and manages licensing, so DPIA inputs are tied to a specific dataset and license. Every dataset is rights-reviewed for ownership and consents and delivered under a license defining records, uses, term and delivery. Personal details such as names, emails, phones and account numbers are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. Diligence materials on source, rights, preparation and allowed use are prepared per dataset, which gives your DPIA annexes a starting point. You can describe the data your team needs to SourceX; the DPIA itself remains the controller's own assessment. For a quick view of which regimes may apply, try the privacy rules checker, and see the privacy and de-identification hub for related guides.

Building the DPIA for a licensed training dataset

SourceX looks for US businesses that hold the data you describe, and nothing is contracted until a supplier agrees. Delivery runs through private, access-controlled workflows only after an executed agreement and supplier approval. Start a data request for your training dataset.