Generative Artificial Intelligence and Web Scraping
The report examines the legal limits on the automated collection of publicly available internet data for training, fine-tuning and evaluating generative models. Its central question is not whether a program can technically download a page, but under what conditions an organisation may extract, combine, store and repeatedly use information about specific individuals. The authors start from the premise that public availability does not replace a legal basis for processing under Article 6 GDPR, the restrictions of Articles 9 and 10, or the rights of data subjects.
The analysis covers the GDPR, Article 53 of the AI Act on the obligations of general-purpose AI model providers, the recommendations of the French CNIL and the UK ICO, and the EDPB’s draft Guidelines 03/2026 on web scraping in the context of generative AI, which at the date of the report remains a document for public consultation rather than a new law. It addresses the roles of controller, processor and data seller, the three-part legitimate interest test, transparency under Article 14 GDPR, data minimisation, data on children, health and criminal convictions, memorisation in model weights, due diligence on suppliers of ready-made corpora, and the relationship with copyright.
The key conclusions are that the lawfulness of collection requires an assessment of purpose, legal basis, data categories, people’s expectations and the actual flow of information across the entire technical chain; that independent verification must rely on real logs, sources and filter effectiveness rather than on declarations; and that removing a record from a dataset and eliminating its influence on an already trained model are distinct technical tasks. Hypothetical scenarios of erroneous sanctions attribution and of outdated information on international searches for wanted persons and extradition are examined separately.
The practical part comprises 50 annexes: legitimate interest assessment and DPIA forms, source intake sheets and provenance manifests, procedures for vetting data brokers and handling data subject requests, an independent audit programme and training cases. Observatoire ARGA sets out its own proposals for the EDPB draft — a verifiable minimum standard for data provenance, honest disclosure of the limits of opt-out and machine unlearning, specific safeguards for sensitive categories and reproducible audit indicators — while noting that their inclusion in the final text has not been determined.
The study is normative and methodological in nature: the authors had no access to developers’ internal training datasets and do not attribute violations to any organisation. External verification cannot guarantee a full reconstruction of a training corpus, and the follow-up research programme includes an empirical evaluation of filters in several languages and a comparative analysis of practice across EU Member States.
Citation Rules
- Mandatory source attributionWhen using ARGA Observatory materials, the full name must be cited.
- Date and version indicationFor analytical reports, the year of publication is mandatory.
- Link to the originalElectronic materials must be accompanied by an active link to the official website.
- Context preservationCitations must not be shortened or altered in a way that distorts the original meaning.
- Note on adaptationIf the text is abridged or translated, state: "adapted from ARGA Observatory report".
- No commercial use without written permission from the organization
- Data accuracy preservationCharts and tables must be reproduced without changes.
Ask the ARGA experts
Four short steps. We reply within one business day.