The European Data Protection Board (EDPB) released draft guidelines on web scraping for generative AI on July 7, 2026. If your team scrapes data for AI training or buys pre-scraped datasets, you're now operating under clearer compliance requirements. These guidelines apply whether you're running the scrapers yourself or purchasing datasets from brokers.
The feedback window closes on October 30, 2026, so the final framework is coming soon. Here's how to implement the EDPB's requirements before that deadline.
The Problem You're Solving
You're collecting personal data at scale from public internet sources. Under GDPR, "publicly available" doesn't mean "freely usable." Every scraped record containing personal data triggers controller obligations: legal basis, data minimization, transparency, accuracy, and special category protections.
The EDPB's position is clear: compliance starts before you scrape the first byte. If you're waiting until model training to think about GDPR, you've already failed the test.
What You Need Before Starting
Role documentation. Determine whether you're acting as a controller, processor, or joint controller. If you're instructing a scraper to collect specific data categories from defined sources with no independent decision-making on their part, they're your processor. If you're both agreeing on collection criteria, you're joint controllers. If you're acquiring a pre-scraped dataset, you and the original scraper are separate controllers with independent obligations.
Legitimate interest assessment template. Consent won't work at scale when you can't identify data subjects before collection. Document all three parts of the legitimate interest test: identify your legitimate interest, demonstrate necessity, and complete a balancing exercise showing your interests don't override data subjects' rights.
Data minimization criteria. Define in advance what you will and won't collect. This includes category filters, source exclusions, and post-collection cleansing rules.
Privacy notice infrastructure. Even if individual notification is impracticable, you need a publicly accessible privacy notice ready to publish.
Step-by-Step Implementation
1. Configure Collection Filters
Before your scraper touches any website, implement technical filters:
Exclude high-risk sources. Block sites directed at minors, healthcare forums, and any domain that explicitly prohibits scraping in robots.txt or terms of service. Maintain an exclusion list you can audit.
Filter sensitive data categories. Use pattern matching to exclude data revealing racial or ethnic origin, political opinions, religious beliefs, trade union membership, genetic data, biometric data, health data, or sexual orientation. The EDPB acknowledges this is challenging but doesn't exempt you from trying.
Timestamp everything. Tag each record with the collection date. You'll need this for accuracy validation and transparency documentation.
2. Document Your Legal Basis
Draft your legitimate interest balancing test now, not retroactively. Address:
- Your legitimate interest: Specify what you're building and why scraping is necessary (e.g., "training a code completion model requires real-world code examples").
- Necessity: Explain why you can't achieve this through synthetic data, licensed datasets, or narrower collection.
- Balancing: Acknowledge growing public awareness of data scraping. Document technical measures that reduce privacy impact (pseudonymization, filtering, access controls). Show you've considered data subjects' reasonable expectations.
This document must be specific to your processing activity. Generic templates won't survive scrutiny.
3. Build Post-Collection Cleansing
After collection but before training:
Run validation passes. Check for directly identifiable data you can remove. Validate timestamps against source material where possible.
Implement deletion workflows. If your filters caught special category data, delete it. Don't just flag it for review.
Consider pseudonymization. Where you've collected names, email addresses, or other direct identifiers, evaluate whether pseudonymization reduces risk without undermining your training objective.
4. Publish Transparency Documentation
Create a publicly accessible privacy notice containing:
- List of scraped sources "to the greatest extent possible" (the EDPB suggests domain names and URLs in searchable format)
- Whether sources are publicly accessible
- Date range of collection
- Crawler characteristics
- If you purchased data, contact details of the originating controller
If you're relying on the disproportionate effort exemption for individual notification, document why. The EDPB's example: large-scale collection spanning 20 years, no direct identifiers, thousands or millions of individuals, with published privacy notice and excluded directly identifiable data qualifies. Targeted collection from 5,000 identifiable individuals spanning two years doesn't.
5. Handle Special Category Data
If you can't technically prevent collecting special category data, document how you meet the GC & Others framework:
- Your processing is analogous to a search engine
- Special category data collection is incidental, not deliberate
- It's genuinely difficult to assess whether the data is present at collection time
- You've implemented robust technical and organizational measures across the full lifecycle
Those measures must span collection (filtering), post-collection (deletion), development (resistance to privacy attacks), and deployment (output filtering with ongoing monitoring).
Validation: How to Verify It Works
Audit your filters weekly. Sample 1,000 collected records. Check for special category data, directly identifiable information, and sources on your exclusion list. If your filters are catching less than 95% of prohibited content, recalibrate.
Test your privacy notice. Can an external party find it via search engine? Does it list specific domains, not just "various internet sources"?
Review your legal basis documentation. Have a colleague unfamiliar with the project read your legitimate interest assessment. Can they understand your necessity argument without additional context?
Check output filtering. Run test queries designed to surface special category data. If your model reproduces scraped content verbatim or generates outputs containing personal data from training, your output filters aren't working.
Maintenance and Ongoing Tasks
Update your source list quarterly. New domains enter your scraping scope; others should be excluded. Keep your privacy notice current.
Re-run balancing tests when processing changes. If you expand to new data categories or sources, your original legitimate interest assessment may no longer apply.
Monitor for data subject requests. When someone exercises their right to object or requests deletion, you need processes to locate their data in your training set and, where technically feasible, retrain or remove affected records.
Track regulatory developments. The consultation period ends October 30, 2026. When final guidelines publish, compare them against your implementation and adjust.
The EDPB's position is clear: publicly available data remains subject to GDPR, and controller obligations arise before you scrape. If you're building AI systems on internet-sourced data, this playbook gives you a defensible starting point. The compliance burden is real, but so is the regulatory risk of getting it wrong.
