The European Data Protection Board has published a four-page "Guidelines at a Glance" summary, dated October 2026, that condenses its Guidelines 03/2026 on web scraping in the context of generative AI into a short set of obligations for AI developers - including a recommendation to stay away from websites that block scraping with robots.txt files, ai.txt files or CAPTCHA.

In Short

Europe's privacy regulators put out a short, plain-language version of their rules on collecting personal data from websites to train AI. It matters to anyone who builds AI models, sells scraped datasets or runs a website, because it says who is responsible and which sites should be left alone. The fuller guidelines are still open for public comment until October 30, 2026, so the details could still change.

A summary of a document still under consultation

The summary carries no day of publication. Its cover gives only "October 2026", and the file's metadata shows it was created on October 12, 2026. It runs to four A4 pages and links twice to the EDPB's public consultation page for Guidelines 03/2026. That page lists the feedback period as July 8 to October 30, 2026, at 23:59 CET.

The summary is explicit about its own limits. "This document provides a simplified overview of the guidelines. For more comprehensive legal explanations and examples, please consult the full text of the guidelines," it states on its final page. What follows here covers what the four pages contain, not the full guidelines.

The underlying text is not new. The EDPB adopted Guidelines 03/2026 on July 7, 2026, a 22-page version 1.0 document, at the same plenary that finalised guidance on blockchain and opened consultation on anonymisation. The blockchain guidelines were finalised as version 2.0 that day, while the anonymisation text, like the scraping text, stays open for comment until October 30. The new summary arrives with less than three weeks of that window left.

Why does a regulator write a four-page version of a 22-page document? The format, which the EDPB labels "Guidelines at a Glance", is written in the second person and addressed to "you (AI developer)". It strips out the legal reasoning and keeps the checklists. For companies that have not read the full text, it is the clearest signal yet of which parts the board considers operational.

Why scraping draws attention

The summary opens with a framing that sets the tone for everything that follows. "Web scraping is a commonly used technique for extracting large amounts of data from publicly available web services. This technique can be a powerful tool for generative AI development, but it poses significant risks as it typically operates 'invisibly'," according to the EDPB.

It defines generative AI as "a technology designed to create new content by learning patterns from existing data," producing "diverse outputs, including text, images, and audio." The guidelines, the summary says, "clarify how to comply with the GDPR, covering both legal and technical aspects."

Web scraping is treated as relevant both for training new models and for fine-tuning existing ones "for specific tasks", according to the summary. That second point matters. Fine-tuning is often done by companies that never build a foundation model themselves, including agencies, ad tech vendors and publishers adapting a model to their own content.

Two scenarios are covered:

  • An AI developer scraping data from the internet itself, or contracting another party to do it.
  • An AI developer obtaining and re-using a dataset that another organisation, described as a data broker, has already scraped.

Who carries the responsibility

The GDPR applies to scraping when it involves personal data processing operations, "such as collection, storage, organisation and retrieval," according to the summary. From there it allocates responsibility across three arrangements.

Hiring a scraper

When an AI developer hires another party to scrape on its behalf, the developer "typically decide[s] how and why scraping happens," which makes it the controller. If the scraper builds the training dataset "according to your documented instructions," the summary says the scraper "may be regarded as a processor."

Buying a pre-scraped dataset

Here the summary separates the two parties. The scraper and the buyer are, "in principle, responsible for your own, separate, processing activities, since each of you determines the objectives and the essential means of your own processing," according to the EDPB. The scraper "is not, in principle, responsible for the re-use of the data." For data brokers and the companies that buy from them, that sentence draws a line: the buyer cannot point back to the seller for its own use of the dataset.

Deciding together

When the scraper and the developer "determine the purposes and means of the processing together," they are jointly responsible. The summary reproduces one example from the full guidelines (section 3.2.1, paragraph 20, page 9, example 1). Two organisations decide jointly to develop an AI model. One scrapes the web and builds datasets; the other develops the model using them.

The EDPB's assessment: "The collection criteria for the scraping of personal data are determined jointly by both parties. Even though they perform separate processing activities, they determine the purposes and means of the processing (the web scraping and the development of the model) together and therefore are jointly responsible for the processing of personal data."

The conclusion follows from agreeing on collection criteria, not from sharing the work. Each party can carry out a different task and still be a joint controller, which brings the shared liability that status carries under the GDPR.

Legitimate interest as the default basis

On legal basis, the summary is brief. "Organisations often use 'legitimate interest' (Art. 6(1)(f) GDPR) as the legal basis for scraping data to train AI models," it states, and links to a separate EDPB summary on that legal basis rather than restating the conditions.

That brevity hides a contested area. The EDPB's Opinion on AI models in December 2024 set out a three-step test of interest, necessity and balancing for AI development. A review of 19 regulatory guidelines published this year found that regulators converge on legitimate interest as the main basis for AI training but differ sharply in how they apply it. A separate EDPB case digest reviewing 62 One-Stop-Shop decisions documented how the basis has failed in practice.

The legislative picture is also moving. The Irish Presidency's Digital Omnibus compromise text renumbered the Commission's AI clause as Article 88 bis and struck the unconditional right to object that the Commission had proposed as a safeguard. If that clause survives trilogue, the legal footing for AI training could shift beneath guidance that is still in draft.

One notable absence: the summary contains no section on consent. PPC Land's July report on the full guidelines described consent as one of its central points, saying the EDPB considered it most probably unworkable for scraping at scale. The four pages do not repeat that.

Purpose limitation and the bias exception

The summary lists lawfulness and purpose limitation as principles requiring "particular attention." Lawfulness means "carefully checking any datasets you use or websites you scrape, and making sure your scraping purpose is lawful."

Purpose limitation carries a specific allowance. Processing personal data "to avoid the risks of potential biases and errors" can fall within the purpose "when this is clearly and specifically identified within the purpose, and the personal data is necessary for that purpose." The summary qualifies necessity in a way that recurs throughout the document: the purpose must be one that "cannot be effectively achieved by processing other data, including synthetic or anonymised data."

Transparency, and what 5,000 people changes

The default is that individuals should be told how their data is processed, which rights they have and how to exercise them, in information "easy for users to find and understand."

Individual notice may not be required where it "proves to be impossible or require excessive effort," according to the summary. The EDPB's example is collecting "a very large dataset from public social media pages or forums, including very old content," where individuals cannot be contacted directly. Even then, information should be provided "in another effective way," such as "a clear, plain and easy-to-find notice" on a website, along with protective measures "such as excluding data from web pages where individuals are directly identifiable."

The worked example the summary chooses runs the other way. Drawn from section 3.2.2, paragraph 34, pages 11-12 (example 3) of the full guidelines, it describes an AI provider further training a language model on data from "a closed group of 5,000 individuals who have discussed a specific topic over the last two years using a particular business or social jargon." The individuals can be directly identified by "full names, email addresses and unique participant IDs."

The verdict: "Since the data is recent and the provider can contact the individuals concerned, providing the information required under Art. 14 GDPR would not involve disproportionate effort. The provider should therefore inform each person individually about the scraping."

Three facts drive that outcome - the data is recent, the group is closed and bounded, and contact details are in hand. Scale alone does not exempt a scraper. A dataset of 5,000 identifiable people with known email addresses is, in the EDPB's view, small enough to notify one by one.

Six steps for data minimisation

The minimisation section is the most operational part of the summary, and the one most directly relevant to publishers and site operators. "Before scraping begins, you should assess what data is truly needed for the development of the model," according to the EDPB. Six measures follow:

  1. Consider using synthetic data instead of personal data.
  2. Define precise collection criteria.
  3. Undertake a data mapping and inventory process.
  4. Apply filters to exclude certain categories where they are not necessary, with bank transaction data and location or geolocation data given as examples.
  5. Exclude certain categories of websites, such as "sites or social networks used mainly by minors," and websites containing sensitive personal data like financial or location information.
  6. Avoid websites that block scraping with technical measures.

The sixth item names specific mechanisms. AI developers should "Avoid websites that block scraping with technical measures (such as robots authentication to view content, robots.txt or ai.txt files, or CAPTCHA) aimed at ensuring that an action can only be carried out by a human being," the summary states.

Placing robots.txt and ai.txt in a data minimisation checklist gives those files a role under data protection law, not only under copyright. Until now, the most concrete European legal hook for crawler opt-outs has been the text and data mining exception in Directive 2019/790. Czech publishers built a robots.txt standard grounded in Article 4(3) of that directive in March 2026. Infrastructure providers have moved too: Cloudflare in August began generating robots.txt from dashboard settings and blocking crawlers that do not separate search from training.

The summary does not say how a scraper should treat a site with no such signals, or how to read a robots.txt file that disallows some crawlers but not others. Those questions, if they are addressed at all, sit in the full text.

The exclusion of sites used mainly by minors also has a recent history. In August 2025, a German court confirmed that Meta's AI training included children's data despite protections, and the Hamburg data protection authority raised concerns about unregistered users' data in public posts.

Accuracy: record when data was taken

Personal data must be "accurate and up to date for its intended purpose," according to the summary. The EDPB recommends scraping only from reliable sources - "official, trustworthy, and maintained" - and avoiding "secondary and outdated aggregators."

Two practical measures follow. The date and time of collection should be recorded, so the age of the data is clear. And data should be checked before training, "for example by making sure email addresses and phone numbers are in the correct format and by reviewing random samples to confirm the information is accurate."

The guidance against aggregators has a commercial edge. Many scraped datasets sold for AI training are themselves compilations of other compilations. A preference for maintained primary sources points developers toward licensed or official data, and away from the cheapest bulk files.

Special categories and four conditions

Processing data revealing racial or ethnic origin, political opinion, health or sexual orientation "is in principle prohibited," the summary states. If scraping involves such data, "both a lawful basis under Art. 6 GDPR and an exception under Art. 9(2) GDPR are required."

The EDPB acknowledges a practical problem: "it is not always possible to know for sure whether collection involves this kind of data until after scraping has occurred." For accidental collection, organisations "should assess these cases individually and be able to show" four things:

  • the processing is similar to that of a search engine;
  • the sensitive data is collected only incidentally and not intentionally;
  • it is difficult or impossible to know in advance whether such data will be collected;
  • appropriate safeguards are in place to prevent its collection and dissemination.

All four must be shown. The search engine comparison is a narrow route, and the burden sits with the organisation doing the scraping. PPC Land's July report on the full text noted that the guidelines draw on the CJEU's 2019 search engine ruling for this analogy.

Why this matters for marketing and media

The audience for the summary is AI developers, but the people affected are wider. Publishers whose pages are scraped now have a European data protection document naming robots.txt, ai.txt and CAPTCHA as signals that a scraper should respect. Data brokers selling scraped datasets learn that buyers are, in principle, separately responsible for what they do with the data. Agencies and ad tech companies fine-tuning models on scraped material are inside the scope, since fine-tuning is named alongside training.

The tension with commercial practice is visible elsewhere. Outside Europe, Japan's data protection commission secretariat proposed in October that companies publish statistics-exception data deals on a crawler-readable web page - a disclosure approach rather than an opt-out one.

What the four-page summary cannot settle is how much of this survives. The consultation closes on October 30, 2026. The Council's Digital Omnibus mandate is still being negotiated, and a version that presumes a legitimate interest in AI training would change the balancing exercise the guidelines rely on. For now, the summary is the EDPB's shortest statement of what it expects from anyone scraping the European web to build AI.

Timeline

Summary

Who: The European Data Protection Board, addressing AI developers that scrape personal data themselves, hire scrapers, or buy datasets from data brokers.

What: A four-page "Guidelines at a Glance" summary of Guidelines 03/2026 on web scraping in the context of generative AI. It sets out controller, processor and joint controller roles, names legitimate interest as the commonly used legal basis, gives a 5,000-person example where individual notice is required, lists six data minimisation steps including avoiding sites protected by robots.txt, ai.txt or CAPTCHA, and sets four conditions for incidental collection of sensitive data.

When: The summary is dated October 2026, with file metadata showing creation on October 12, 2026. The full guidelines were adopted on July 7, 2026, and their public consultation closes on October 30, 2026.

Where: The European Union, under the GDPR, published by the EDPB at edpb.europa.eu.

Why: The EDPB describes web scraping as a technique that "poses significant risks as it typically operates 'invisibly'." The summary condenses the operational parts of the guidelines while the full text is still open to comment and while the Council negotiates changes to how the GDPR treats AI training.