Was the AI model you use trained lawfully? A supplier evidence sheet from EDPB Opinion 28/2024 (training data sources, anonymity claims, legitimate interest) for small businesses and their accountants
What you get: a finished AI model supplier evidence sheet in three tabs: a model list with one row per AI tool you feed personal data into and a result formula that sorts each row into STOP, ASK, RECORD or LOW; a supplier question template you can paste into an email, built from the documentation list in EDPB Opinion 28/2024; and a legitimate interest note for the cases where you rely on Article 6(1)(f) GDPR. Four made-up example rows show how it fills in.
Short answer: if you deploy someone else's AI model and process personal data with it, the EDPB says supervisory authorities "should take into account whether the controller deploying the model conducted an appropriate assessment" to ascertain that the model "was not developed by unlawfully processing personal data", looking at criteria such as "the source of the data" and whether a supervisory authority or a court has found an infringement (Opinion 28/2024, para. 129). A supplier simply saying the model is anonymous is not enough: "a mere assertion of anonymity of the model is not enough to exempt it from the application of the GDPR" (para. 134). How deep you go depends on the risk (para. 130). This sheet records that assessment.
Why this matters for a small business
Under Article 5(2) GDPR the controller "shall be responsible for, and be able to demonstrate compliance". The EDPB applies this to AI models: "each controller should ensure the lawfulness of the processing it conducts and be able to demonstrate it", so authorities may look both at the developer and at "the controller that acquired the AI model and processes the personal data by itself" (para. 126). Where processing was unlawful, an authority may order corrective measures that can go as far as, depending on the facts and proportionality, "the erasure of the whole dataset used to develop the AI model and/or the AI model itself" (para. 114). If your workflows depend on one model, you want to know before that happens.
Two limits to keep in mind. The Opinion is addressed to supervisory authorities, and its observations "should be assessed on a case-by-case basis" (para. 112). And it deals with lawfulness (Article 5(1)(a) and Article 6 GDPR); other issues such as transparency or data protection by design are "not analysed in this Opinion" (para. 111).
Step 1: tab 1, model list (one row per AI tool or model)
| Column | What to write | Source |
|---|---|---|
| A. Tool and model | For example "Support chatbot, built on model X from developer Y". Ask the vendor which model sits underneath if it is not stated | |
| B. Personal data you put in | Yes or No. Customer chats, CVs, emails, call transcripts, client files all count | Art. 4(1) GDPR |
| C. Risk of your use | High (special category data, decisions about people, many people, children) or Normal. Decides how much evidence you ask for | Para. 130: degree of assessment "may vary depending on" the risks |
| D. Supplier claims the model is anonymous? | Yes, No or Not stated | Paras. 34, 43, 134 |
| E. Anonymity evidence received | Yes (documents covering tab 2, questions 4 to 7), Partial or None | Para. 58(a) to (f) |
| F. Training data sources described? | Yes or No. For general-purpose AI models, look for the public training content summary (column I) | Para. 129 ("the source of the data"); para. 58(c) |
| G. Finding by an authority or court? | Yes, No (you searched the developer's name plus "data protection authority" and "fine") or Unknown | Paras. 129, 130 |
| H. Data from a breach? | Yes, No or Unknown | Para. 130: "if the data originates from a personal data breach" |
| I. Public training summary link | Link, or N/A. Providers of general-purpose AI models must publish "a sufficiently detailed summary about the content used for training", using an AI Office template, since 2 August 2025 | AI Act Art. 53(1)(d), Art. 113(b) |
| J. Result | =IF(OR(G2="Yes",H2="Yes"),"STOP: get legal advice before further use",IF(B2="No","LOW: no personal data from you, keep a note",IF(AND(D2="Yes",E2<>"Yes"),"ASK: anonymity claimed without evidence",IF(OR(F2<>"Yes",G2="Unknown",H2="Unknown",AND(C2="High",E2<>"Yes")),"ASK: send tab 2 questions","RECORD: file the evidence, re-check at renewal")))) | Formula and labels are our working rule, not a legal test |
A high-risk AI system comes with an EU declaration of conformity, which must contain, where personal data is processed, "a statement that that AI system complies with Regulations (EU) 2016/679" (AI Act Annex V, point 5; Article 47). File it, but do not stop there: the EDPB says "such a self-declaration may not constitute a conclusive finding of compliance under the GDPR" (para. 131).
Step 2: tab 2, supplier question template
Paste these into one email to the vendor. Questions 4 to 7 follow the documentation the EDPB says authorities should ideally see when a model is claimed to be anonymous (para. 58).
| # | Question to the supplier | Source |
|---|---|---|
| 1 | Which AI model or models does your service use, and who developed each one? | Para. 125 (roles of each actor) |
| 2 | Describe the sources of the training data, including any public web data and any third-party datasets. If the model is a general-purpose AI model, please send the link to its public training content summary. | Para. 129; para. 58(c); AI Act Art. 53(1)(d) |
| 3 | Has any data protection authority or court found that personal data was processed unlawfully to develop the model? Did any training data come from a personal data breach? | Paras. 129, 130 |
| 4 | If you say the model is anonymous: please share the DPIA information (or the decision that none was needed) and any advice from your DPO. | Para. 58(a), (b) |
| 5 | What technical and organisational measures did you take to reduce the likelihood of identification, for each source of training data? | Para. 58(c), (d) |
| 6 | How was the model tested against attacks such as regurgitation of training data, membership inference and model inversion: by whom, when, how and with what results? | Para. 55; para. 58(e) |
| 7 | What documentation do you give deployers about these measures and about residual risks? | Para. 58(f) |
| 8 | If personal data in the model is erased or restricted, will you tell us? | Para. 127; Art. 17(1)(d), Art. 19 GDPR |
Keep the reply, or a note that no reply came, as the evidence for the row. The EDPB adds that passing tests against known attacks "can only be evidence for the resistance to those attacks" (para. 55).
Step 3: tab 3, legitimate interest note
If your own use of the model relies on legitimate interest (Article 6(1)(f) GDPR), the EDPB says the unlawfulness of the initial processing "should be taken into account as part of the legitimate interest assessment" (para. 132; see also para. 123). Add three lines to your assessment for each model:
| Line | What to record | Source |
|---|---|---|
| Technical limits | Filters or access limits built into the model that you cannot switch off and that may stop personal data being disclosed | Para. 132 |
| Legal facts | What is known about the nature and severity of any unlawful processing in development (from column G) | Para. 132 |
| Expectations and risks | Whether the people affected would expect your use, and the risks to the people whose data was used for training | Paras. 123, 132 |
If a supplier shows that the model was anonymised and your use of it does not process personal data from training, the EDPB considers the GDPR would not apply to that operation, and the lawfulness of the personal data you then collect in deployment "should not be impacted by the unlawfulness of the initial processing" (paras. 134, 135). Your own processing still has to meet the GDPR on its own terms.
Four made-up example rows
| Tool | Key answers | Result |
|---|---|---|
| Bookkeeping firm: AI reads supplier invoices (contact names on them) | Personal data Yes, risk Normal, sources described, no finding found, breach No | RECORD: reply filed, re-check at renewal |
| Dental clinic: chatbot answering patient questions | Personal data Yes, risk High, vendor says "our model is anonymous", no documents | ASK: anonymity claimed without evidence |
| Web shop: AI writes product descriptions from the shop's own catalogue | No personal data put in | LOW: note kept |
| Recruiter: CV screening tool whose developer has a published authority decision about its training data | Finding Yes | STOP: legal advice sought, tool paused |
These rows are made up for illustration and do not describe any real business or supplier.
This sheet is a working aid, not a legal opinion or proof of compliance. References are to EDPB Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models, adopted on 17 December 2024 (paras. 34, 43, 55, 58, 111 to 135), to Regulation (EU) 2016/679 (Articles 4(1), 5(1)(a), 5(2), 6(1)(f), 17(1)(d), 19, 24) and to Regulation (EU) 2024/1689 (Articles 47, 53(1)(d), 113(b) and Annex V point 5), as read on EUR-Lex. EDPB opinions are not binding on the courts; the Court of Justice gives the final interpretation. The columns, result labels, formula and question wording are our own working rules. Checked on 28 October 2026. Spot an error in this guide? Write to mininglife7@gmail.com.
For accountants, tax advisers and bookkeepers
If a client is asked about its AI suppliers in a customer or bank questionnaire, tab 1 gives a documented answer per tool, and tab 2 gives the client a ready email to send to each vendor. Use the same sheet for the AI tools in your own practice, which handle client personal data every day. Stay inside your professional rules and refer legal questions on.
Keep the result next to your AI tool list
The EURO AI Register free edition runs in your browser with no account and no upload. It covers your first 10 AI systems and 10 evidence records, the screening questions, duties by role with dates, and a printable Article 4 AI-literacy record. List each tool once, then attach the supplier's reply as the evidence record.
When you need more than 10 systems, the EURO AI Register membership starts at $19 a month and adds unlimited systems, rule updates and email help. Neither tool is legal advice or proof of compliance.
Partner programme opening soon
Partner programme opening soon: join the free waitlist. Planned terms are 50% of each referred sale (75% once you have 5+ paying customers), single level, paid through Gumroad only from revenue received.
Free start: list your AI tools in the EURO AI Register free edition, then add this sheet. Related: AI vendor questions for deployers · Legitimate interest assessment sheet · Article 28 DPA clause check · Pseudonymisation check
Sources
- EDPB Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models, adopted 17 December 2024
- Regulation (EU) 2016/679 (GDPR), official text: EUR-Lex
- Regulation (EU) 2024/1689 (AI Act), official text: EUR-Lex
Unadhi Guides are free, independent explainers. We correct errors as soon as we find them: write to mininglife7@gmail.com.