Unique Research · Compliance · December 19, 2024
On Wednesday the European Data Protection Board (EDPB) issued an Opinion on how AI developers can use personal data to build and deploy models such as LLMs without breaking EU privacy law. The Board guides GDPR implementation and supports enforcement — so its view matters.
Topics include: whether AI models can be treated as anonymous (if so, privacy law may not apply); whether “legitimate interests” can be a lawful basis for developing and deploying models (if so, no individual consent); and whether models trained on unlawfully processed data can later be lawfully deployed.
Legal basis for GDPR-compliant AI remains hot and unsettled — OpenAI’s ChatGPT has already hit trouble. Non-compliance risks fines up to 4% of global annual turnover and/or orders to change how tools work.
About a year ago Italy’s DPA initially found OpenAI’s chatbot breached GDPR. Complaints followed in Poland, Austria, and elsewhere — legal basis for processing, tendency to invent facts, inability to correct wrong statements about people. GDPR also grants access, erasure, and rectification rights — hard for chatbots that “hallucinate.”
Despite rapid GenAI complaints, enforcement so far has been relatively light; authorities are still working out how long-standing rules apply to data-hungry training. The Opinion aims to help supervisors decide. Ireland’s DPC — now overseeing OpenAI’s GDPR compliance after recent legal changes — said the Opinion will “promote proactive, effective and consistent regulation of AI models across the region” and support engagement with new-model developers and AI complaints, Commissioner Dale Sunderland noted.
Main takeaway for developers: no one-size-fits-all fix for legal uncertainty.
Model anonymity
EDPB defines anonymity as “highly unlikely” that individuals used to create the model can be identified directly or indirectly, and “highly unlikely” users can extract such data via prompts. Assessment must be case by case. Non-prescriptive methods to evidence anonymity include: training-data source choices that avoid/limit personal data; minimization and filtering before training; robust method choices that cut identifiability (e.g., regularization against overfitting, differential privacy); measures that reduce extraction via queries.
Design choices affect how far GDPR applies. Only truly anonymous data (no re-identification risk) falls outside the Regulation; for AI models the bar is “highly unlikely” identification. Anonymity is not assumed — case-by-case is required. Pre-Opinion, DPAs disagreed; some argued models can never be personal data — the Board rejects that default.
Legitimate interests
Few GDPR bases fit AI; OpenAI learned that via Italian enforcement. Legitimate interests is often preferred — no consent from every person whose data trains the tech (consent is neither attractive nor scalable for LLM volumes).
DPAs must assess fit via the standard three-step test: purpose and necessity (lawful and specific; less intrusive alternatives?); balancing impact on individuals’ rights. EDPB leaves room for models to meet all criteria — e.g., powering conversational agents that assist users, or better threat detection in information systems (lawful purpose). Necessity must weigh proportionality and data minimization. Balancing must weigh case specifics, risks to fundamental rights, and data subjects’ “reasonable expectations” (public data? source/context? relationship with controller? further uses?).
If balancing fails, mitigations may limit impact — technical measures (as under anonymity), pseudonymization, masking/replacing personal data in training sets, opt-outs and other rights-exercise measures, transparency. Web-scraping risks get special attention.
Unlawfully trained models
Again: consider each case’s specifics. The Opinion appears to leave a path for models built on shaky legal foundations (e.g., scrape-everything) if developers anonymize personal data before deployment — and can show subsequent operation involves no personal-data processing: “unlawfulness of the initial processing should not affect subsequent operation of the model.”
Independent adviser and King’s College London AI Institute member Lukasz Olejnik (whose ChatGPT GDPR complaint has sat with Poland’s DPA for over a year) warns against “systematic abuse schemes”: focusing only on the end state (anonymization) could unintentionally legitimize scraping without a proper legal basis and weaken GDPR’s principle that personal data must be lawful from collection onward. He adds the Opinion does not bind national DPAs; Poland’s will likely consider it — and his complaint goes beyond training to accountability and privacy by design.
— END —