EDPB Guidelines 03/2026 · Web Scraping & Generative AI
New GDPR Guidance Just Made Your
AI Training Data an Audit Target
What the guidelines demand, the risks they create for CTOs, CISOs and legal teams — and an honest look at which parts SBOMator DataBOM solves.
What Happened
Guidelines 03/2026 on Web Scraping in the Context of Generative AI
- Adopted 7 July 2026 by the European Data Protection Board; public consultation runs until 30 October 2026.
- Not a new law — interpretive guidance on how the GDPR already applies to scraping training data for generative AI. Supervisory authorities will enforce along these lines.
- Covers organisations that scrape data themselves, contract scrapers, or acquire already-scraped datasets from brokers.
- Core message: scraping "publicly available" data is fully in scope of the GDPR — with concrete, checkable expectations.
The Demands
Five Concrete Expectations Hidden in the Guidelines
Published source list
Even under the "disproportionate effort" exemption: publish a precise indication of sources — good practice is domain names, URLs, searchable format, collection dates.
Syntax-based filtering
Apply "syntax-based filtering mechanisms (e.g. regular expressions)" during and after collection to strip unnecessary personal data.
Special-category safeguards
Health, political, religious, ethnic data: incidental collection tolerated only with demonstrable detection and lifecycle safeguards.
Timestamps & reliable sources
Timestamp the data, validate it, and prefer maintained, trustworthy sources.
Accountability records
You must be able to demonstrate all of the above — for every dataset, including ones you bought.
Risk · CTO
Your Datasets Are Opaque — and Now That's a Liability
- Training corpora are terabytes of crawled text with no manifest. Nobody on the team can say which domains are in there, from when, or what personal data they carry.
- Datasets acquired from brokers or open repositories inherit the compliance burden — the guidelines put acquirers in scope, not just scrapers.
- Building internal provenance tooling means diverting engineers from the product, for a moving regulatory target.
- Retrofitting provenance after a model ships is far more expensive than evidencing it at ingestion.
Risk · CISO
Auditing the Data Without Creating a New Exposure
- Every cloud "data compliance" service means uploading your training corpus to a third party — a new processor, a new transfer, a new breach surface.
- Datasets often contain exactly what you don't want leaving the building: residual PII, customer content, proprietary crawls.
- Regulators and enterprise customers increasingly ask for evidence artifacts — you need reports you can hand over without handing over the data itself.
- Air-gapped and classified environments can't use SaaS scanners at all.
Risk · Legal / DPO
Legitimate Interest Now Needs Evidence, Not Assertions
- Art. 6(1)(f) balancing requires knowing what data, from which sources, under what expectations — you can't balance what you can't see.
- The Art. 14(5)(b) exemption is conditional: no source list, no exemption.
- Art. 9 incidental-collection tolerance (GC & Others, C-136/17) depends on demonstrable detection and remediation of special-category data.
- In an investigation, "we believe the dataset is clean" is not a defence. A dated, itemised provenance report is.
The Solution
SBOMator DataBOM — Dataset Provenance, Scanned Like Firmware
The scanner architecture your security team already uses for firmware SBOMs, pointed at AI training data. Runs 100% locally — your dataset never leaves your PC.
- Source inventory: every domain, record counts, first/last collection dates
- Regex PII screening: emails, phones, IPs, card numbers (Luhn-validated)
- Art. 9 keyword indicators: health, political, religious content flags
- Provenance completeness score with per-component breakdown
Traceability
Each Report Section Maps to a Guideline Demand
| DataBOM report section | Guideline expectation it evidences |
|---|---|
| Source inventory — domains, counts, dates, searchable | Art. 14(5)(b) published source list (paras 30–31) |
| Regex PII scan — emails, phones, IPs, cards | "Syntax-based filtering mechanisms (e.g. regular expressions)" (para 38) |
| Art. 9 keyword indicators | Post-collection detection of special-category data (para 68) |
| Timestamp coverage metric | Accuracy — "timestamp the data" (para 42) |
| Full report + CycloneDX ML-BOM | Art. 5(2) accountability record for the dataset (para 16) |
DataBOM output is evidence for your compliance file — it is not legal advice and not a compliance certification.
Honest Scope
What We Solve — and What We Don't
Evidence layer
- Source list generation (Art. 14(5)(b))
- Regex PII screening (para 38)
- Dataset accountability record (Art. 5(2))
Detection, not remediation
- Art. 9 indicators are keyword screens, not legal classification
- Timestamp coverage measured; source reliability is your judgment
- Removal / anonymisation stays your workflow
Lawyer-shaped problems
- Legitimate-interest balancing (Art. 6(1)(f))
- robots.txt / ai.txt opt-out checks (needs network — on the roadmap as an optional online step)
- Collection-time exclusions, privacy notices, DPIAs, data-subject rights
Our positioning is deliberate: DataBOM makes your dataset provable, your lawyers make it lawful. Tools that promise both are overselling.
Why Local Matters
Air-Gapped by Design = GDPR by Design
No new processor
No cloud vendor touches your data — no Art. 28 processor agreement, no sub-processor chains to audit.
No transfer risk
Nothing crosses a border. Chapter V transfer assessments simply don't arise for the scan itself.
Art. 32 inherited
Security of processing inherits the controls you already certified — disk encryption, access control, physical security.
Works fully offline
Same site-license model as SBOMator firmware scanning. Suitable for air-gapped and classified environments.
Know Your Data
Like You Know Your Code
Talk to us about dataset provenance scanning on your own hardware — and tell us which capability your compliance file needs next.
Contact ESL