Belgian DPA Orders Chatbot App to Get Explicit Consent to Train AI on Sensitive Data
In decision 320/2026 the Belgian DPA found a chatbot app could use ordinary chat data to train its AI on legitimate interest, but needed explicit consent for sensitive data. It has four months to comply, with penalty payments of up to EUR 2,000 a day.
What happened
The Belgian Data Protection Authority’s Litigation Chamber adopted decision 320/2026 on 16 November 2026 and announced it on 17 November. A complaint said an app for building personalised chatbots used users’ chat data to train its AI without a valid legal basis.
What it changes
The authority split the data in two. The company could rely on legitimate interest to train its model on ordinary chat data, given the users’ reasonable expectations and the safeguards in place. For special category data, such as health, sexual orientation or religious beliefs, it needed explicit consent and did not have it. It was also not transparent enough about either use.
The company was reprimanded and ordered to bring the processing of sensitive chat data and its user information into line within four months. Penalty payments are EUR 1,000 per breach per day of delay, EUR 2,000 a day in total. The parties have 30 days to appeal.
Who is affected
Providers of chatbots and other AI services that train models on what users type, and organisations that deploy them in Belgium.
What to do now
Find out whether your training data can include special category data. Separate it from ordinary data, get explicit consent or exclude it from training, and update your privacy notice to say what you train on and why.
Data Protection Impact Assessment (DPIA)
Our Data Protection Impact Assessment maps which chat data reaches model training, sets the legal basis for each kind of data and lists the consent and notice changes. Our Data Protection Compliance Audit then checks them against the GDPR.
