EPIC Comments to EDPB on Guidelines on Web Scraping for AI: Go Further

EPIC has submitted comments to the European Data Protection Board (EDPB) Call for Comments on their draft Guidelines on Web Scraping to Train Generative AI. The Guidance is intended to address the many conflicts that have arisen between the AI developers’ practice of using scraped data to train generative AI systems and the data protection and rights elements of the GDPR and privacy principles, making clear what specific efforts and circumstances must be in place for the practice to be legal.

EPIC’s comments identify some gaps within the Guidelines and urge the EDPB to declare more firmly when web scraping cannot be legally used for AI training – which is nearly always. First, the current scope of the Guidance excludes web scrapers if they are creating data sets not explicitly intended for AI training at the outset. This completely ignores data brokers who often create massive data sets using scraped data that they will sell to other parties for multiple purposes. Without including data brokers in the scope of the Guidelines, many of the pre-scraping recommendations (such as determining a data scope, ensuring a legal basis for collection, performing a risk assessment, etc.) will not be applied, leading to increased risk and harm to individuals over their data.

Second, EPIC analyzes how web scraping for AI training can be made compliant with the GDPR. We conclude that in most cases, it cannot. The only possible legal basis under which this data may be scraped is legitimate interest. However, legitimate interest requires both a balancing test where risk to the individual’s rights and freedoms is weighed against the controller’s interest. Due to how generative AI systems incorporate their training data into the algorithm itself, AI’s increased ability to create inferences and to re-identify de-identified individuals, and the unlikelihood that anyone made their data public contemplating its eventual use by a third party for AI training, AI developers will almost certainly lose this balancing test. Even if legitimate interest could be used as a legal basis, individuals will have a nearly impossible time exercising their rights to correct or delete personal data once an AI system has been trained on that data.

Third, EPIC highlights the many ways in which the practice of data scraping is incompatible with privacy principles. In particular, mass data scraping is by its nature antithetical to data minimization and the notice portion of the transparency principle is wholly unaddressed. The Guidelines acknowledge that notifying individuals about the scraping and use of their personal data may be impossible or unduly burdensome to AI companies and suggest the possibility of simply posting a notice on the company’s website that they use scraped data. This is not a solution, does not meet the transparency requirement, and is practically absurd. The suggestion posits that an individual will somehow sense that their data may have been scraped for this purpose (how), will magically know what company or companies may have used their scraped data for AI training (again, how), and will then be able to find the notice and contact the company to exercise their data rights. That is not how any of this works.

Finally, while EPIC maintains that web scraping for generative AI training as it currently stands is illegal under the GDPR and a violation of privacy principles, we listed some suggestions as to how individual rights may be better respected if web scraping continues. We recommend requiring all companies using scraped data for AI training purposes to register in a public registry that includes information about what data was scraped, from what sources, what they are using it for, how to exercise individual rights, and other pertinent data. We further recommend that the default assumption be that web pages cannot be scraped unless they have an indicator, like a robot.txt, that affirmatively permits scraping for AI training purposes. We also recommend more concrete requirements for cleaning scraped data sets prior to using them for training.

The AI industry is counting on authorities ignoring its illegal and exploitative practices in the name of supporting innovation. There is nothing innovative in violating people’s privacy to make more money. EPIC will continue urging enforcement bodies to uphold their laws and protect their constituents by holding companies that violate those laws to account.