Prefer to listen? Click play for AI narration

Most generative AI models learn from text, images and audio collected from the open web: blogs, forums, news sites, social media profiles and public registers. Much of that material is about real, identifiable people. For years, the industry's informal assumption was simple: if it is publicly accessible, it is fair game.

European regulators have now spelled out, in detail, why that assumption does not hold. In July 2026 the European Data Protection Board (EDPB) published draft Guidelines 03/2026 on web scraping in the context of generative AI, and the public consultation closes on 30 October 2026. On 8 October 2026 the UK Information Commissioner's Office (ICO) published the closing report of its two-year supervision of the biggest foundation model developers. Together they give the clearest picture yet of what “lawful scraping” looks like, and of what companies will be expected to prove.

This matters beyond AI labs. If you commission a dataset, buy one from a broker, fine-tune a model on public posts, or build a lead-generation tool that collects profiles, these rules reach you too.

30 Oct
Deadline to comment on the EDPB draft guidelines
10
AI developers that changed practices after ICO scrutiny
3%
Maximum AI Act fine for training-content summary failures (or €15M)
0
Digital Omnibus changes in force today. The GDPR applies in full

Why scraping is under the spotlight now

Three developments have landed within a few weeks of each other:

Meanwhile, the EU's proposed GDPR amendments for AI are stuck in the Council (more on that below). For now, the existing rules are what count.

“Public” doesn't mean “free to use”

The EDPB's starting point is blunt. The GDPR applies to web scraping whenever it involves personal data, from extraction through cleaning, structuring and storage, whether the data identifies someone directly or indirectly. Two further points are worth remembering:

The ICO reaches the same place from a different angle. In its words, vague statements that a model was trained on “publicly accessible information” are “unlikely to be sufficient” for transparency purposes.

Who is responsible: the scraper, the developer or both?

Scraping supply chains often involve several companies, and the EDPB says roles must be assessed case by case, on the facts. Its general guidance:

The ICO adds that developers should take responsibility for ensuring everyone in their training-data supply chain, including data brokers, meets their transparency obligations.

Legitimate interest: the three-step test, applied to scraping

The ICO calls legitimate interest (Art. 6(1)(f) GDPR) realistically the only lawful basis for training generative AI on scraped personal data. The EDPB says it is the basis private companies usually rely on, and that the others are generally less likely to apply. Both insist it has to be earned through a documented three-step test.

  1. A real, specific interest. It must be lawful, clearly articulated, and real and present. The ICO adds that broad interests such as “developing and improving our products and services”, “training our models” or “benefitting humanity” are unlikely to be enough without specification or justification.
  2. Necessity. Narrowing collection criteria, rather than “scraping a wide part of the internet”, may be crucial, says the EDPB, and pseudonymised or synthetic data may be less intrusive alternatives. The ICO expects developers to explain why other sources, such as licensed datasets, are not suitable.
  3. Balancing. The EDPB lists risks such as loss of control, a “chilling effect” on free expression, de-anonymisation of people who posted anonymously, and deepfakes. It also notes that, with today's technology, personal data “cannot be easily deleted” from a trained model.

Reasonable expectations carry a lot of weight, and the EDPB ties them closely to what websites signal. Someone posting on an open platform that allows scraping, and says so, can reasonably expect AI developers to collect their content. Someone posting on a platform that blocks scrapers with robots.txt and CAPTCHA, and says it does not allow AI training on users' data, cannot.

The draft also shows where the line falls. An organisation that collects public voice recordings to sell a voice generation tool, with no safeguards against malicious reuse, “cannot rely” on legitimate interest. A text model trained only on freely accessible content that people manifestly made public, respecting copyright opt-outs, with safeguards against memorisation, easy rights requests and a public list of sources, may generally pass.

Transparency when you can't tell everyone

Article 14 GDPR requires you to inform people whose data you obtained indirectly. When you have scraped data about millions of people, individual notices are often impossible. Article 14(5)(b) allows an exception for impossibility or disproportionate effort, but the EDPB says it “should not be routinely relied upon”, and the ICO warns against using it as a “blanket” exemption.

If you rely on the exception, the EDPB says you must at least make the information public, for example in a dedicated notice on your website. It should cover the categories of data, purposes and legal basis, people's rights, and a precise indication of sources, including the crawler's characteristics. As good practice, list the domains and URLs scraped, in a searchable format, with the collection period.

Scale matters. The EDPB contrasts scraping 20 years of forum posts with no direct identifiers, where individual notice would be disproportionate, with fine-tuning on posts from a closed group of 5,000 identifiable people. In the second case, each person should be informed individually.

This overlaps with the AI Act. The Commission's template asks general-purpose AI providers to name their crawlers, give the collection period and list the top 10% of scraped domains (for SMEs, the top 5% or 1,000 domains, whichever is lower). One good source inventory can serve both laws.

🧭
GDPRGard Tool
Not sure which GDPR duties apply to your data collection?

Answer a few questions in our GDPR decision tree to see which obligations apply to a specific use of personal data or AI.

Open the decision tree →

Minimisation and accuracy: decide before you crawl

Data minimisation does not prohibit training on large volumes of data, the EDPB stresses. It prohibits collecting personal data you do not need, and that assessment has to happen before scraping starts:

On accuracy, the EDPB recommends reliable sources such as official registries, timestamping what you collect and spot-checking samples before training.

Sensitive data: the “search engine” route and its limits

Special categories of data, such as health, political opinions or sexual orientation, need a legal basis under Article 6 and an exception under Article 9(2). Scraping almost inevitably picks some up.

The EDPB's draft proposes a narrow path for data collected incidentally and residually, despite efforts to avoid it. It draws on the Court of Justice's ruling in GC and Others (C-136/17, 24 September 2019), which held that search engine operators are bound by the sensitive data rules “within the framework of [their] responsibilities, powers and capabilities”. The EDPB says this reasoning may apply to AI scraping if you can show that:

Those measures are demanding: filters and site exclusions before collection, deletion as soon as sensitive data is identified (including on a plausible request), extraction testing and output filters during development, and continuous output monitoring after launch. The Board is clear that this is not a general exemption from Article 9.

The ICO is more sceptical. It considers the “manifestly made public” exception unlikely to cover sensitive data collected through scraping, because developers cannot verify that the person intended to make it public, and says the scientific research condition is unlikely to cover all model development. The ICO openly acknowledges that this “limits the ability for developers to lawfully train their models using special category data” and is discussing the issue with the UK government.

The model may be personal data too

Scraping risks do not end when training finishes. The ICO restates that models can contain personal data because training data can be extracted from them, a view compatible with the EDPB's Opinion 28/2024. It warns that web-sourced training data can include email signatures, API keys and passwords, and expects developers to test models for “singling out” and “linkability”. It also highlights good practice on rights: Apple, for example, lets anyone flag webpages that should not be used for training and tells site owners how to block its crawler for training while still allowing search indexing.

Where the law may be heading

The Commission's Digital Omnibus, presented in November 2025, proposes GDPR amendments relevant to AI training. According to a leaked Council compromise text of 3 September 2026 (document 12535/26), reported by GDPR Local and the privacy group noyb, a recital would say that processing personal data to develop and deploy AI “may be regarded as” a legitimate interest, unless the person's rights override it. Another recital would allow sensitive data that ends up in training sets “incidentally and residually” in limited circumstances. noyb strongly opposes the proposal, and its chair, Max Schrems, called it “a digital expropriation of Europeans.” We unpack the Council text in our article on the Digital Omnibus and legitimate interest for AI.

The file is far from settled. EU ambassadors failed to agree a Council negotiating position on 7 October 2026, mainly over trade secrets in the Data Act part of the package. According to MLex, the item was then removed from the agenda of their 11 October meeting, with no new date set. The European Parliament had tabled roughly 1,840 amendments in August. As of 12 October 2026, none of this is law, and the GDPR applies in full.

Beyond the EU: Japan and Brazil

Japan. Japan has amended its privacy law (APPI) to allow consent-free use of data, including publicly available sensitive data, for statistics, which can cover AI development organised as statistics creation. A draft paper from the regulator's secretariat, published on 1 October 2026, suggests companies relying on this route should publish the required details on a single web page that crawlers can read. It is a draft, not final rules.

Brazil. The LGPD (Law 13.709/2018) includes legitimate interest among its legal bases, but not in its Article 11 list of bases for sensitive data. Speaking at FGV Direito Rio on 29 April 2026, ANPD board member Iagê Miola said the authority has no definitive position yet on legitimate interest for AI training. He said consultation contributors did not see it as invalid, that views on scraping were divided, and that the lack of a legitimate-interest basis for sensitive data was the main concern. If used, it would need robust balancing tests and meaningful transparency. Brazilian companies offering goods or services to people in the EU are already covered by the GDPR through its Article 3(2), so the EDPB's guidance applies to them directly.

For global companies, “lawful scraping” will mean different things in different markets. Building to the stricter EU and UK standard is the safest default.

Practical checklist for companies that scrape or buy scraped data

Conclusion

The message from both Brussels and London is consistent: the open web is not a free-for-all, but it is not off-limits either. Scraping personal data to train AI can be lawful when it is targeted, minimised, transparent and backed by safeguards that you can show actually work. The companies that will struggle are those that cannot say where their data came from, why they needed it or what they did to protect the people in it. The EDPB's guidelines are still a draft and the Digital Omnibus may yet shift the debate, but mapping sources, filtering data, publishing notices and testing models will pay off under any version of the rules.

Book a free consultation with GDPRGard → and we will help you check your data sources and training notices.

Sources

Share this article

Also read:

⚠️ This article is for informational purposes only and does not constitute legal advice. For complex compliance situations, consult a qualified data protection professional. GDPR and EU AI Act requirements are subject to ongoing regulatory guidance — verify current obligations with your legal adviser.