From Ritu Raj | Product & Market Analysis
EDPB Web Scraping Guidelines Are Still a Draft: 12 Records to Keep Now
On this page
The EDPB's web scraping guidelines closed for comment on 30 October 2026, and they are still a draft. Guidelines 03/2026 carry version 1.0, which the Board's own version history labels "adoption of the guidelines for public consultation". For GDPR AI training, the only final EDPB text is Opinion 28/2024. You can document against both today, and you should.
Key takeaways
- Guidelines 03/2026 are still a version 1.0 draft. The EDPB adopted them on 7 July 2026 for public consultation, feedback closed on 30 October 2026, and the consultation page listed no final version on 10 November 2026.
- Opinion 28/2024 is the settled EDPB text on GDPR AI training. Adopted on 17 December 2024, it sets the anonymity bar at an insignificant likelihood of extraction and lists six kinds of documentation supervisory authorities should look for.
- The draft treats robots.txt as a one-way signal. A missing robots.txt file is not consent, but a site that refuses scraping through robots.txt, ai.txt or CAPTCHA should be excluded, and scraping it anyway weakens a legitimate interest case.
- B2B AI vendors can build 12 records now without waiting for the final text. Every record in the list below traces to a numbered paragraph in Opinion 28/2024 or Guidelines 03/2026, so the work survives most changes in the final version.
The short answer
Guidelines 03/2026 on web scraping in the context of generative AI are not final. The EDPB adopted version 1.0 on 7 July 2026 for public consultation, and feedback closed on 30 October 2026. No final version was listed on 10 November 2026. Opinion 28/2024, adopted on 17 December 2024, is the settled EDPB text.
What the EDPB has actually published on web scraping
The document is called Guidelines 03/2026 on web scraping in the context of generative AI. It runs to 77 numbered paragraphs over 22 pages, with seven worked examples. The cover says version 1.0, adopted on 7 July 2026.
The EDPB consultation page gives the feedback period as 8 July to 30 October 2026, closing at 23:59 CET. When this post was checked on 10 November 2026, the page listed the guidelines PDF and a short summary called "Guidelines at a glance". Nothing on it was marked final.
What version 1.0 means
The version history has a single row. It reads "adoption of the guidelines for public consultation". That phrase is the whole status check. The Board adopted a text so it could ask for comments, not so it could close the question.
The EDPB marks the difference when a text does become final. Its page for Guidelines 3/2025 on the interplay between the DSA and the GDPR now shows a final version, adopted on 17 September 2026. Until the scraping page shows something similar, the text you can read is a draft, however official its formatting looks.
Commentary has used three verbs for the same event: adopted, published and issued. All three are accurate. None of them tells you the text is a draft unless the writer adds the qualifier, and several headlines did not.
The summary document adds a second layer of confusion. The at a glance PDF is dated October 2026 on its cover and was uploaded in November. A reader who finds it after the deadline could reasonably assume it summarises a final text. Its own footer calls it a simplified overview of the guidelines, and the guidelines are still the draft.
The anonymisation draft runs on the same clock. The EDPB page for Guidelines 02/2026 on anonymisation, dated 8 July 2026, shows the same 30 October closing date and a closed consultation. Expect the two to be read together, because the scraping text points to anonymisation as a safeguard more than once.
| Document | Adopted | Status on the EDPB site | How to treat it |
|---|---|---|---|
| Opinion 28/2024 on AI models | 17 December 2024 | Final Article 64(2) GDPR opinion | Addressed to supervisory authorities, who take utmost account of it |
| Guidelines 03/2026 on web scraping | 7 July 2026, version 1.0 | Draft, feedback closed 30 October 2026 | The Board's current direction, open to change |
| Guidelines 02/2026 on anonymisation | July 2026, page dated 8 July | Draft, feedback closed 30 October 2026 | Will move the anonymity reference point once final |
| Guidelines at a glance: web scraping | Dated October 2026 | Summary document | Simplified overview of the draft, not a separate position |
Opinion 28/2024 is the final EDPB word on GDPR AI training
The Irish supervisory authority asked the EDPB four questions on 4 September 2024. The Board answered them in Opinion 28/2024, adopted on 17 December 2024 under Article 64(2) GDPR. The draft scraping guidelines cite it for purpose limitation, for examples of legitimate interests and for reasonable expectations.
The opinion is addressed to supervisory authorities. Paragraph 15 says authorities should take utmost account of it, and that it is without prejudice to the obligations of controllers. In practice it tells you how your regulator will read your file.
The anonymity test
Paragraph 34 settles the headline question in one sentence. "AI models trained on personal data cannot, in all cases, be considered anonymous." Anonymity is assessed case by case, against specific criteria.
Paragraph 43 sets the bar. The likelihood of extracting personal data from the model, and of obtaining it through queries, should be insignificant for any data subject. Authorities should assume by default that a thorough evaluation is needed. Paragraph 134 adds that "a mere assertion of anonymity of the model is not enough".
Paragraph 55 names the tests an authority may look for. They are attribute and membership inference, exfiltration, regurgitation of training data, model inversion and reconstruction attacks. The same paragraph warns that passing known attacks only proves resistance to those attacks.
Legitimate interest in three steps
Paragraph 66 asks authorities to verify that controllers "carefully assessed and documented" three cumulative conditions. These are a legitimate interest, necessity, and a balancing test the controller passes. Paragraph 69 gives three examples that may qualify, including a conversational agent built to assist users.
Necessity has teeth. Paragraph 73 says that if the purpose can be met by a model that does not process personal data, processing personal data is not necessary. That is the sentence most vendor assessments skip.
I think paragraph 73 is the most underrated line in the opinion. It turns data minimisation from a principle you cite into a test you either pass or fail, and it puts the burden of proof on you.
What the draft EDPB web scraping guidelines add
The draft does not overturn the opinion. It narrows the analysis to one activity, scraping by private entities for generative AI training, and makes the expectations concrete. Paragraph 43 says legitimate interest is the basis private bodies often rely on. The other Article 6(1) grounds will generally be less likely to apply.
Consent is effectively off the table. Paragraph 44 says consent "would most probably not be an applicable legal basis", because scrapers have no direct relationship with the people in the data.
Transparency gets the most detail. Paragraph 27 says the Article 14(5)(b) disproportionate effort exception "should not be routinely relied upon" outside archiving, research and statistics. Where you do rely on it, paragraph 30 says you must still publish the Article 14 information, including sources and the crawler's characteristics.
Example 3 is the one B2B vendors should read twice. A provider fine-tunes on a closed group of 5,000 people from the last two years, identifiable by name and email. The draft says informing each person would not be disproportionate, so each must be told individually.
robots.txt cuts one way
Paragraph 45 is blunt. "The absence or non-applicability of a robots.txt file on a web site does not amount to consent within the meaning of the GDPR." Silence from a site gives you nothing.
A refusal does take something away. Paragraph 37 tells controllers to exclude websites that clearly oppose scraping through robots authentication, robots.txt, ai.txt or CAPTCHA. Paragraph 64 adds that scraping such a site anyway makes it less likely people expect the processing, and Example 5 says those users cannot reasonably expect it. The commercial side of that choice for publishers is covered in the analysis of whether to block AI crawlers in robots.txt.
Special category data and the search engine test
Paragraphs 72 to 75 borrow the CJEU's search engine ruling in GC and Others, C-136/17. The draft says it can be relevant to incidental and residual collection of special category data. It also says the ruling "should not be seen as a general exemption".
To use it, you must show four conditions. The processing resembles a search engine's, collection is only incidental, you cannot know in advance what you will collect, and you take measures within your responsibilities, powers and capabilities. Paragraph 74 then sets four minimum measures across collection, cleaning, training and deployment, including output monitoring once the system is live.
The 12 records a B2B AI vendor can keep before the final text
None of the items below depends on the draft surviving unchanged. Each maps to a paragraph in the final opinion, the draft guidelines, or both. If the final version moves, you edit a record rather than build one from nothing.
If I ran compliance at a B2B AI vendor, I would not wait for version 2.0. The cost of these records is mostly engineering time spent logging things you already know at the moment you know them.
| # | Record | What it contains | EDPB source |
|---|---|---|---|
| 1 | Role map | Controller, joint controller or processor for scraping, cleaning, training and deployment, decided before processing starts | Opinion 62; Guidelines 16 to 20 |
| 2 | Purpose statement | Model type, expected functions, internal or external deployment, commercial or research objective | Opinion 64; Guidelines 50 |
| 3 | Legitimate interest assessment | All three steps, dated at the outset, with DPO input and any DPIA or decision that one was not needed | Opinion 58(a), 66 |
| 4 | Collection criteria and source list | Domains, languages, topics, excluded site categories, and when each item was scraped | Guidelines 37, 42; Opinion 50 |
| 5 | Opt-out signal log | How robots.txt, ai.txt, CAPTCHA and login walls were detected and honoured | Guidelines 37, 66(c); Opinion 105(c) |
| 6 | Filtering record | Regular expression filters, pseudonymisation, synthetic substitution, and why any were not used | Guidelines 38; Opinion 51 |
| 7 | Public Article 14 notice | Data categories, purposes, legal basis, sources, crawler characteristics, unlisted source types with reasons | Guidelines 30, 31 |
| 8 | Disproportionate effort file | Dataset-level weighing of the number of people, the age of the data and the safeguards | Guidelines 28 |
| 9 | Special category file | Analysis of the four conditions, plus a deletion log for sensitive items found later | Guidelines 73 to 75 |
| 10 | Rights handling log | Objections, opt-out requests, erasure and regurgitation claims, received and resolved | Guidelines 66(f); Opinion 102 |
| 11 | Model test reports | Who tested, when, how, against which attacks, and the results | Opinion 55, 58(e) |
| 12 | Upstream model diligence | For models you did not train: data sources and any regulator or court finding | Opinion 129, 130 |
The numbering and grouping are Zan Digital's. The paragraph references are the EDPB's: Opinion means Opinion 28/2024, Guidelines means Guidelines 03/2026 version 1.0. Paragraph numbers in the draft may change in the final version.
Records about collection
Records 4 to 8 are the cheapest to produce and the hardest to recreate later. A timestamp you did not capture at scrape time cannot be reconstructed honestly. Paragraph 42 asks for exactly that: record when the data was scraped, to show how current it is.
Record 7 doubles as a sales asset. Paragraph 31 calls it good practice to publish domain names and URLs in a searchable format, with the collection period. Enterprise buyers increasingly ask for something close to this in security reviews, and it sits naturally beside an AI bill of materials disclosure.
Records about the model
Record 11 is where most vendors will find the largest gap. Paragraph 58(e) asks for reports on how the model was tested, by whom, when, how and to what extent, plus the results. A slide saying the model was red-teamed is not that record.
Paragraph 58(c) also asks for measures per source of training data, including relevant source URLs. That is a much finer grain than a typical model card carries. If your training data came through a third-party dataset provider, the same paragraph expects a description of the measures that provider took.
Fine-tuning on customer data sits outside the scraping draft
This is the gap B2B vendors miss. Paragraph 5 of the draft says it does not address processing of an organisation's own personal data. Paragraph 4 excludes data brokers who scrape and hold datasets without training models themselves.
So if your product fine-tunes on customer tickets, CRM records or call transcripts, Guidelines 03/2026 are not your text. Opinion 28/2024 is. The Irish request asked explicitly about first-party data, and paragraph 74 of the opinion says less intrusive means may differ depending on whether you have a direct relationship with the people concerned.
The direct relationship cuts both ways. It makes individual notice more feasible, which is the logic Example 3 applies to scraped data. It also means your processor terms decide who is the controller for training. The contract language for that question is covered in the clause-by-clause read of AI data processing agreements.
My view is that many B2B vendors should start their evidence work here rather than with scraping. For a lot of B2B products, customer data sits closer to the model than any crawled page does.
If you build on someone else's model, Scenario 2 of the opinion applies
Many B2B AI vendors do not train a foundation model. They deploy one. Opinion 28/2024 covers that position in Scenario 2, where personal data retained in a model is processed by another controller at deployment.
Paragraph 129 says authorities should check whether the deploying controller "conducted an appropriate assessment" that the model was not developed by unlawfully processing personal data. Paragraph 130 names what to check. Did the data come from a breach, and did a supervisory authority or court find an infringement?
The depth expected scales with risk. A support assistant that summarises tickets carries different deployment risk from a model that pre-screens job applicants. Paragraph 131 adds that an AI Act declaration of conformity "may not constitute a conclusive finding of compliance under the GDPR".
Record 12 is the output of that assessment. Keep the provider's training data statement, its published notice, any regulator findings you can locate, and the date you checked. The AI Act side of the same supplier relationship is set out in the note on GPAI enforcement and downstream obligations.
Where this argument is weakest
Three things could make this list less useful than it looks, and you should weigh them before you staff the work.
Guidance is not statute
Neither text is law. The opinion is addressed to supervisory authorities, and the draft is a consultation document. A national authority or a court can read Article 6(1)(f) differently, and the CJEU, not the EDPB, has the final word on interpretation.
The examples are also generous in places. Example 7 says the balancing test "may generally be considered as met" for a text model trained only on manifestly public, copyright-cleared content with memorisation safeguards. That bar is lower than many readers expect, and the final text may tighten it.
The draft can move, and so can the anonymity reference
The 114-day consultation exists to change the text. Submissions will push on paragraph 27 and on the robots.txt language. I would not bet on either softening much, but I cannot rule it out.
Opinion 28/2024 tied its anonymity analysis to WP29 Opinion 05/2014 "at the date of this Opinion", in paragraph 40. When Guidelines 02/2026 on anonymisation become final, that reference point changes, and records 6 and 11 should be reviewed against it.
The copyright layer is separate again. Example 7 links the balancing test to text and data mining opt-outs under Directive 2019/790, a question tracked in the review of AI training data copyright cases. A clean GDPR file does not settle it.
Frequently asked questions
Are the EDPB web scraping guidelines final?
No. Guidelines 03/2026 carry version 1.0, adopted on 7 July 2026, and the EDPB's own version history describes that as adoption for public consultation. Feedback closed on 30 October 2026. As of 10 November 2026 the consultation page lists no final version. The text can still change, so treat it as the EDPB's current direction rather than its settled position.
Can I use legitimate interest to train AI on scraped data under GDPR?
Possibly. Both Opinion 28/2024 and the draft scraping guidelines accept that legitimate interest under Article 6(1)(f) GDPR can work, subject to three cumulative conditions: a legitimate interest, necessity, and a balancing test you pass. The draft says consent will most probably not apply to scraping. Your assessment must be documented at the outset and supported by mitigating measures where the balance is close.
Does robots.txt matter under GDPR for AI training?
Yes, in two directions. The draft guidelines say a missing robots.txt file does not amount to consent. They also say controllers should exclude sites that clearly oppose scraping through robots.txt, ai.txt, CAPTCHA or login walls. Scraping a site that signals refusal makes it less likely people expect the processing, which weakens your balancing test under Article 6(1)(f).
Is an AI model trained on personal data anonymous under GDPR?
Not automatically. Opinion 28/2024 says models trained on personal data cannot, in all cases, be considered anonymous. Anonymity requires the likelihood of extracting personal data from the model, and of obtaining it through queries, to be insignificant for any data subject. Supervisory authorities will look at your documentation and testing, including membership inference, regurgitation and model inversion tests, before accepting the claim.
Do I have to notify people whose data was scraped for AI training?
Usually you must give Article 14 information, but individual notice can be dropped where it proves impossible or needs disproportionate effort under Article 14(5)(b). The draft says that test runs on the whole dataset, weighing the number of people, the age of the data and your safeguards. Even then you must publish a notice naming sources and crawler characteristics. A small, recent, contactable group must be told individually.
What documentation does the EDPB expect for AI training data?
Opinion 28/2024 lists what supervisory authorities should ideally find: DPIA records, including decisions that one was not needed; DPO advice; technical and organisational measures with source URLs; lifecycle measures; resistance testing with who, when, how and results; and what you gave deployers. The draft scraping guidelines add collection criteria, timestamps, a public source list and the analysis behind any Article 14(5)(b) reliance.
How to start this week
Pick one model in production and write records 1 and 3 for it before Friday. The role map and the dated legitimate interest assessment are the two documents every later record hangs from, and both are short if you write them while the decisions are fresh.
Then open your scraper or data pipeline configuration and check one thing. Does it log, per source, the date of collection and the robots.txt or ai.txt state at that moment? If not, add it now, because a day without that log is a day of data you cannot account for later. If you also ship into the EU AI Act's disclosure rules, the AI Act transparency checklist covers the obligations that sit beside these.
Last verified 10 November 2026 against the EDPB consultation pages, Guidelines 03/2026 version 1.0 and Opinion 28/2024. Re-check the consultation page for a final version before relying on any draft paragraph number.
Related on this topic
The same training data questions show up in contracts and in copyright. Read the AI data processing agreement clauses next, then the training data copyright cases.
References
- European Data Protection Board, Guidelines 03/2026 on web scraping in the context of generative AI, version 1.0, adopted 7 July 2026. Used for status, scope, paragraph references and examples.
- European Data Protection Board, Public consultation page for Guidelines 03/2026, feedback period 8 July to 30 October 2026, accessed 10 November 2026. Used for consultation dates and status.
- European Data Protection Board, Guidelines at a glance: web scraping for generative AI, how and when to use it, dated October 2026. Used for the summary document's framing.
- European Data Protection Board, 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, with its landing page. Used for the anonymity test, legitimate interest, documentation and Scenario 2.
- European Data Protection Board, Public consultation page for Guidelines 02/2026 on anonymisation, accessed 10 November 2026. Used for status and dates only.
Weakest point in the source base: this post relies only on EDPB texts. It does not draw on national supervisory authority decisions, and the full text of the anonymisation draft was not analysed, only its status.
Related reading