Know what is on your site before a plaintiff’s firm does.Free website scanScan Your Site
Log in Sign up Book a demo
Home› Editorial› Belgium’s APD Clears Chatbot Training on Legitimate Interest,…
GDPR

Belgium’s APD Clears Chatbot Training on Legitimate Interest, but Not on Users’ Sensitive Conversations

Belgium’s data protection authority has drawn a line every company training a chatbot on user conversations should study.

Nov 17, 2026 9 min read

Know what’s on your site before a plaintiff’s firm does

A free privacy audit shows which cookies, pixels and trackers are running — and which ones commonly appear in privacy claims.

Get My Free Audit

Belgium’s data protection authority has drawn a line every company training a chatbot on user conversations should study. In a decision dated November 16, 2026 and published the next day, the Litigation Chamber of the Autorité de protection des données (APD, known in Dutch as the GBA) ruled that an unnamed AI companion app could rely on legitimate interest to fine-tune its model on ordinary chat data, but had no legal basis at all for training on the sensitive material users reveal in those chats, including erotic role-play. Early summaries said the company lacked a basis for AI training generally and faces a fine. Neither is right: the problem is limited to special-category data, and the sanction is a reprimand plus orders backed by daily penalty payments, not a fine.

TL;DR

  • The APD’s Decision 320/2026 found that training a chatbot model on ordinary, pseudonymised conversation data was lawful under legitimate interest, but training on sensitive conversation data had no valid legal basis under Articles 6(1) and 9(2) GDPR.
  • The company received a reprimand for an outdated privacy notice and was ordered to collect explicit consent before sensitive conversations and to explain clearly how that data is used, within four months.
  • Missing the deadline triggers penalty payments of €1,000 per day for each of the two orders, capped at €100,000 each; the APD chose not to impose a fine, and the decision can be appealed within 30 days.

Watercolor of Brussels-style stepped-gable rooftops with a speech-bubble kite tethered to an open blank ledger book on a rooftop ledge

The case behind Decision 320/2026

Two users filed the complaint on April 15, 2023. The app lets people build personalised chatbots from preset parameters (behaviour, language, goal) and then talk to them, including in erotic role-play, which the decision abbreviates as “ERP”. The complainants raised misuse of their data, minors and vulnerable users’ mental health. Neither the decision nor the APD’s press release names the company.

Several procedural points shaped the outcome and are worth knowing before reading too much into it:

  • No EU establishment. The company has no establishment in the European Economic Area, and a 2023 mutual assistance request under Article 61 found no other authority investigating it, so the one-stop-shop did not apply.
  • A narrowed scope. Because the file reached the APD’s inspectors outside the statutory 30-day window, the Chamber set their report aside and dismissed four of five grievances (purpose limitation, minimisation, transfers, accountability). Only lawfulness and transparency under Articles 5(1)(a), 6 and 9 was decided.
  • Adults only. The complainants were adults who did not represent minors, so the Chamber did not assess processing of children’s or at-risk users’ data.

How the Chamber sorted the data and the purposes

The Chamber separated ordinary conversation data from special-category conversation data, and two purposes: running the chatbot, and “optimising” the model, which in practice means fine-tuning.

The company fine-tunes two open-source pre-trained models on user conversations, after stripping user identifiers and pseudonymising or masking personal details in the text. Moderation filters screen inputs and outputs, and a notice warns users they are talking to an AI that may give inaccurate or fictional answers. It argued that this meant it processed no sensitive data, that training data was anonymised, and even that users were controllers of what they shared. The Chamber rejected all three positions:

  • Still personal data. Users had shared names, occupations and dates of birth, and conversations were tied to account identifiers. Because the company deliberately kept the ability to re-link pseudonymised data, for example to answer access requests, the data was pseudonymised rather than anonymous, following the Court of Justice’s September 2025 judgment in EDPS v SRB (C-413/23) and the EDPB’s Guidelines 02/2026 on anonymisation.
  • Special categories by inference. Relying on the CJEU’s December 2025 Russmedia judgment (C-492/23), which also features in the CJEU’s reshaping of platform liability under the GDPR, the Chamber held that ERP conversations reveal information about a person’s sex life or sexual orientation and therefore fall under Article 9(1).
  • The company is the controller. It decides why and how conversations are used to build and refine its model, which makes it the controller under Article 4(7), whatever users volunteer.

Legitimate interest passed for ordinary chats

For ordinary conversation data, the company won on lawfulness. Using that data to run the chatbot was necessary to perform the contract under Article 6(1)(b). Using it to fine-tune the model was justified by legitimate interest under Article 6(1)(f), after the Chamber applied the familiar three-part test, citing the EDPB’s Guidelines 1/2024 on legitimate interest. The EDPB set out the same framework for AI models specifically in its Opinion 28/2024, adopted in December 2024, although the Belgian decision itself relies on the guidelines rather than the opinion.

  1. Purpose test. Improving a model so its replies are coherent and relevant is a lawful interest, clearly defined, and real rather than hypothetical.
  2. Necessity test. Conversation data is needed to spot trends and positive or negative user feedback. Pseudonymisation limits what is processed, and the Chamber accepted the company’s explanation that synthetic data alone is not yet good enough to replace real conversations without degrading results.
  3. Balancing test. The Chamber identified real risks: loss of control over intimate conversations, profiling, and bias feedback loops in which a model trained on user chats reproduces gender, racial or sexual stereotypes. It still found the balance tipped toward the company, because users share the data voluntarily, processing is minimised, outputs are filtered, an AI notice is shown, and users of a chatbot can reasonably expect it to learn from conversations.

The Chamber added a warning: the balance “depends heavily” on minimisation measures, especially pseudonymisation. Remove them and the same training could fail.

Where the company failed: sensitive data and its privacy notice

The sensitive-data analysis produced a split that will interest anyone designing consent flows. For running the chatbot, the Chamber applied Russmedia by analogy and held that a user who voluntarily shares sensitive information with a chatbot gives explicit consent to its processing for that purpose, under Articles 6(1)(a) and 9(2)(a). That consent, however, covers only the conversation itself. It does not extend to using the same material to train the model, and the company had no other Article 9(2) condition to point to. Training on sensitive conversation data therefore breached Articles 6(1) and 9(2).

Transparency was the second problem. The company’s March 14, 2021 privacy notice listed legal bases without naming the contract or the legitimate interest concerned. An updated notice dated January 21, 2025 fixed that for ordinary data, so the Chamber found a past breach of Article 5(1)(a), read with recital 39, and issued a reprimand for it. For sensitive data, the 2025 notice only said that users “might voluntarily share personal data, including sensitive information” during chats, which the Chamber said was not enough.

The decision also makes a point product teams often miss. The app’s “you are talking to an AI” notice is the kind of labelling the EU AI Act encourages, but the Chamber stressed that the AI Act applies without prejudice to the GDPR. Telling users a reply may be fictional says nothing about how their sensitive data is processed, so it does not satisfy GDPR transparency. Teams drafting chatbot disclosure notices should treat the two as separate obligations.

What the company must do, and what happens if it does not

The Chamber chose a reprimand and two compliance orders over a fine, citing the structural changes the orders force, steps already taken, and turnover below €50 million. The decision (in French; quotations here are our translation) requires the company to:

  • build a mechanism for users to give explicit consent to the use of their sensitive conversation data for model training;
  • collect that consent before a conversation in which users are likely to share sensitive data begins, with ERP given as the example;
  • tell users clearly that sensitive conversation data is processed, that the basis is explicit consent, and how consent is collected for each processing operation;
  • add a concise prevention message to its existing in-chat notification, plus a reminder that sensitive details shared in a chat will be used to optimise its models;
  • send the Chamber proof, including the updated privacy notice, the notification text and its triggers, and a description of the consent mechanism.

The company has four months from notification. After that, if the Chamber finds partial or no compliance, penalty payments of €1,000 per day apply to each order, so up to €2,000 a day in total, capped at €100,000 per order. The company can request an extension within 45 days, and either party can appeal to the Market Court (part of the Brussels Court of Appeal) within 30 days of notification. That court has overturned APD measures before, as the APD’s long fight with IAB Europe over the TCF showed, so this decision may not be the last word.

What chatbot operators should take from it

The ruling joins European enforcement against companion apps such as Italy’s €5 million Replika decision, but its message is more practical. Legitimate interest can support training on user conversations, if the safeguards are real. Sensitive content needs its own track.

  • Map data by category and purpose. One legal basis for “the chatbot” will not survive scrutiny; split serving from training and ordinary from sensitive data.
  • De-identification is not anonymisation. If you can re-link training records to accounts, you are processing personal data.
  • Ask first for sensitive features. Where a feature invites intimate disclosure, collect explicit, separate consent to training before the conversation starts.
  • Document your safeguards. The legitimate-interest result rested on pseudonymisation, filtering and notices; record them in your assessment.
  • Write notices for people. Name the interest you rely on and remind users in context that what they share may train the model.

Elsewhere, some regulators loosen rules on pseudonymised data for AI development while others write rules aimed at emotional companion chatbots. For services reaching EU users, the Belgian decision is now the concrete reference point.

FAQ

Did the APD ban training chatbots on user conversations?

No. It held that training on ordinary, pseudonymised conversation data was lawful under legitimate interest in this case. Only training on sensitive conversation data lacked a legal basis.

Was the company fined?

No administrative fine was imposed. The company received a reprimand and two compliance orders. Penalty payments of €1,000 per day per order apply only if it fails to comply after four months, capped at €100,000 per order.

The Chamber applied the CJEU’s Russmedia reasoning by analogy: a person who chooses to disclose sensitive information about themselves consents to its processing for that disclosure. That consent was limited to running the conversation and did not cover model training.

Is pseudonymised training data outside the GDPR?

Not when the company keeps the means to re-identify it. The Chamber found the training database was pseudonymised, not anonymous, partly because the company could re-link records to answer access requests.

Which company is it?

The APD did not identify it, and the published decision redacts the parties’ names. We have not named it either.

The model can learn, but not from everything

The Belgian decision gives AI companies something they rarely get from a regulator: a worked example of legitimate interest succeeding for model training. The price is discipline. Ordinary chat logs can feed the model if they are pseudonymised, filtered and honestly disclosed, but the moment a product invites people to share their health, beliefs or sex lives, training on those confidences needs a clear, separate yes from the user, asked before the conversation starts.

Put this into practice

Turn privacy guidance into working controls.

Captain Compliance helps teams discover tracking technologies, enforce consent, manage privacy requests and document the evidence behind every decision.

Book a demo Run a free scan