A new framework for a practice that has become widespread
On 7 July 2026, the EDPB adopted Guidelines 03/2026 on web scraping in the context of generative artificial intelligence. This is the first time a European authority has published a GDPR framework devoted specifically to the automated collection of publicly available data to train AI models. The consultation is open until 30 October 2026 at 23:59.
This deadline is not incidental. It means the final text will not be available before late 2026 at the earliest. Nevertheless, the principles set out in version 1.0 of these guidelines are already precise enough to guide your data governance choices now.
Who is concerned?
The guidelines apply in two directions. The first: your organisation extracts data from the web to build training datasets, whether for its own models or for third-party services. The second, often overlooked: your organisation publishes content online, and its data is, or may be, scraped by third parties to train generative AI models.
The two situations create different obligations. In the first case, you process potentially personal data and must justify that processing. In the second, you are the controller of the data you publish, and the users whose data is concerned hold rights that scraping may undermine. The EDPB addresses both cases.
What the guidelines actually say
On the legal basis, the EDPB confirms that legitimate interest (Article 6(1)(f) of the GDPR) is the most likely legal basis for web scraping for AI training purposes. This does not mean it is automatically applicable: the organisation must demonstrate that its interest is real, necessary and proportionate, and that the rights of the data subjects do not override it. A documented balancing test is expected.
On protection signals, the EDPB treats robots.txt files, ai.txt files, CAPTCHAs and login walls as indicators of the reasonable expectations of data subjects. A site that publishes an ai.txt file prohibiting the use of its content for AI training sends a signal that the scraping organisation must respect and document in its legitimate interest assessment.
On the full lifecycle, the guidelines cover not only the initial collection but also the use of the dataset, the training of the model and its deployment. Data collected in compliance may become non-compliant if it is used in a context the data subject would not reasonably have anticipated.
The political tension of September 2026
Guidelines 03/2026 arrive in a context of lively debate at European level. In September 2026, several Member States are examining the possibility of relaxing consent rules for training AI on personal data, citing European competitiveness against American and Chinese models. Mistral's EUR 3 billion fundraising at a valuation exceeding EUR 20 billion, announced in September 2026, illustrates the economic stakes involved.
This tension creates regulatory uncertainty that your GRC programme must anticipate. The final text of the guidelines could be shaped by the positions of Member States before the end of 2026. Taking part in the public consultation, or at the very least documenting your analysis of the current requirements, is a way to prepare the organisation for several scenarios.
A concrete example: the SaaS publisher using public data to refine its model
A compliance software publisher scrapes sanction decisions published by European data protection authorities to refine a GDPR risk classification model. These decisions contain the names of companies, legal representatives and, at times, natural persons. The processing involves personal data. The publisher must: document its legitimate interest, verify the absence of a restrictive robots.txt file on the source sites, minimise the data collected, and define a retention period for the training dataset. If these steps are not formalised, the publisher risks a formal notice from the competent authority once the final guidelines are adopted.
What your organisation must do before 30 October
Three actions should be started now. The first: map your web scraping activities, whether active or planned, whether carried out in-house or subcontracted. This mapping must appear in your record of processing activities. The second: assess your exposure as an online content publisher. Does your terms of use permit or prohibit scraping for AI training purposes? Is an ai.txt file in place, or is one relevant? The third: decide whether your organisation will submit comments to the EDPB consultation before 30 October. Organisations whose business models depend on large-scale scraping or publication of data have good reason to make their position known.
The documentary governance of AI, of which Guidelines 03/2026 form a new part, is a project that goes beyond the question of scraping alone. The article on the structural gaps that EU AI Act documentary compliance conceals describes the common blind spots in AI governance programmes, which apply here too.
Why not wait for the final text
Version 1.0 of Guidelines 03/2026 is already public and precise enough to guide your data architecture choices. Waiting for the final version means postponing a risk analysis you will have to carry out anyway, with less time to adapt your practices. Organisations that document their analysis today will have a structural advantage: they will be able to adjust their position in light of the final text, rather than starting from scratch against a constrained deadline.
The GDPR framework does not change. What changes is the precision with which the EDPB explains how it applies to the training of generative AI. That precision is a resource, not an additional constraint.