The EDPB Says Scraping the Web for AI Still Counts as Data Processing

“Public” is not the same as “free of privacy rules.” The European Data Protection Board’s new guidance on web scraping for generative AI makes that point in the context where it matters most: automated collection at internet scale. At its July 8, 2026 plenary, the EDPB adopted draft guidelines on web scraping in the context of generative AI. The board says web scraping that includes personal data processing—collection, storage, organisation, or retrieval—falls within the GDPR framework. The guidance is open for consultation until October 30, 2026. It does not end the policy debate, but it gives AI developers and data controllers a clearer list of questions they need to answer. Public pages can still contain personal data A public profile, forum post, photograph, review, or code repository can contain information about an identifiable person. Automated collection changes the scale and context of that information. A detail that was hard to find on one page can become searchable, correlated, and reusable after it is copied into a training or retrieval pipeline. The EDPB’s framing rejects a shortcut: public access does not by itself settle whether collection is lawful. Organizations must still examine the purpose, legal basis, transparency, minimisation, security, and downstream use of the processing. That is especially important when a scraper collects more than a model needs. A system optimized to gather “everything” may capture names, contact details, health information, political views, children’s data, and private context embedded in pages that were technically reachable. What responsible scraping requires The EDPB highlights several practical controls: Scrape only from reliable sources and record when information was collected. Validate information before it is used for training or model outputs. Minimise collection instead of treating the open web as an unlimited warehouse. Explain the processing to people where required, while documenting any reason an exemption is claimed. Test whether special categories of personal data were collected and address both the GDPR legal basis and the additional Article 9 requirements. Put safeguards around retention, access, deletion requests, and vendor transfers. Accuracy matters too. A scraped page can be outdated, satirical, incorrect, or describing someone else with a similar name. Training or retrieval systems can preserve the error while presenting it in a more authoritative voice. Legitimate interest is not a blank cheque AI developers may argue that model training serves a legitimate interest. The EDPB’s approach treats that as an analysis, not a permission slip. An organization must identify its interest, show that the processing is necessary, and balance it against the rights and expectations of the people whose information is collected. A person who publishes a post for a small audience may not reasonably expect it to be copied into a global dataset, combined with records from elsewhere, and used to generate inferences. Context does not decide every case, but it is part of the balance. The existence of an opt-out mechanism also does not answer every question. An opt-out may be difficult to discover, may come after collection, and may not address data already copied into derivative datasets or model weights. Questions to ask an AI provider Before treating an AI service as a neutral search or writing tool, ask: What sources were collected and what categories of personal data were included? How are scraped records filtered, retained, corrected, and deleted? Can a person access or object to the processing? Are special categories or children’s data removed before training? Which vendors and jurisdictions receive the data? What does the system do when a source is wrong or a person requests correction? The EDPB’s draft guidance makes a simple principle harder to avoid: the web may be public, but the people represented on it are not raw material without rights.