Belgian DPA orders chatbot app to obtain consent for AI training data
The app faces two conditional €1,000 daily penalties, each capped at €100,000, if it has not complied four months after notification.
Belgium’s Data Protection Authority (APD) has ordered an AI chatbot application to obtain explicit consent before using users’ special-category conversational data to optimise its model, and to improve its disclosures. The Litigation Chamber’s Decision 320/2026, dated 16 November 2026, gives the company four months from notification to comply. The app is not named in the decision, which identifies the respondent as Y.
The order matters for chatbot providers that use conversations for model training. The APD found that pseudonymisation did not remove special-category data from the GDPR, and that consent to share sensitive information while using a chatbot did not extend to using it to improve the underlying model. The company must therefore build a consent mechanism, obtain consent before a conversation in which special-category data may be shared, and give clear information about both uses.
What the APD found
The application lets users create and converse with personalised chatbots, including in erotic role-play conversations. The complainants had shared information from which their sex life or sexual orientation could be inferred. The APD found that the application remained responsible for processing those conversations because it decided the purposes and means of collecting and using them to refine its AI model.
Removing user identifiers and keeping them separately reduced the risk of re-identification, but amounted to pseudonymisation rather than anonymisation. The training database therefore remained within the GDPR, including where its conversational content contained special-category data.
The Chamber accepted two different legal bases for ordinary conversational data. Article 6(1)(b) GDPR covered processing necessary to provide relevant chatbot responses, while Article 6(1)(f) covered model optimisation based on the company’s legitimate interest. The Chamber considered optimisation lawful, necessary and reasonably expected by users, taking account of voluntary sharing, pseudonymisation, content filtering and user notifications.
Special-category data had a different result. The Chamber accepted that voluntarily sharing such information during chatbot use could constitute explicit consent for that use under Articles 6(1)(a) and 9(2)(a). But that consent did not cover model optimisation. The APD found no Article 6 and Article 9 basis for that separate processing.
The company’s earlier privacy notice, dated 14 March 2021, also failed to explain the purposes and legal bases for processing conversational data. The company corrected that failure by updating the notice on 21 January 2025, so the APD issued a reprimand rather than a further compliance order for that historical breach.
What the order requires
The application must:
- implement a mechanism for explicit consent to special-category data processing for model optimisation;
- collect that consent before a potentially sensitive conversation begins;
- tell users clearly that special-category data is processed when they use the chatbot;
- explain that the processing relies on explicit consent and how consent is collected;
- add a concise warning to the existing AI interaction notification; and
- remind users through a technical or organisational measure that sensitive conversational data may be used to optimise the models.
It must provide the APD with the updated privacy notice, the revised notification and its triggers, and evidence or an explanation of the consent mechanism.
The order is backed by two conditional periodic penalty payments, one for each injunction. Each is €1,000 per day of delay, capped at €100,000. The payments do not begin automatically after four months: the APD will first review the evidence and notify the company if it has complied only partially or not at all. The penalty then starts on the date of that notification. The company may seek a reasoned extension within 45 days of notification if compliance is impossible despite reasonable efforts.
The APD dismissed four broader alleged breaches, including purpose limitation, data minimisation, international transfers and accountability, because the available evidence did not sufficiently establish or delimit them. The company and complainants have 30 days from notification to appeal to Brussels’s Market Court.
Permanent link to this Atlas analysis
Related Atlas analysis
Continue with coverage connected by market, topic and operating context.
- IcelandIceland orders Bus4u to halt Idrive AI coach monitoringThe authority found failures involving legal basis, fairness, transparency, security, information duties and DPIA requirements.
- CyprusCyprus NSIU tells persons to freeze assets after 145 EU Russia listingsThe EU listed 113 individuals and entities over war financing, and 32 over foreign interference and malicious cyber activity.
- European UnionECB keeps 2027 capital requirements stable and flags AI risks through 2029The average SREP score stayed at 2.5, while 40% of DORA incidents reported to the ECB originated with third-party providers.
Continue with Atlas
Move from this development into the relevant research, comparison and workflow.