MorseSign inStart freeOpen Morse

Belgium's chatbot decision: training on sensitive chats needs explicit consent

The Morse teamPublished 9 min read

In Decision 320/2026 of November 16, 2026, the Litigation Chamber of Belgium's data protection authority (the Belgian DPA, called APD in French and GBA in Dutch) found that a chatbot app trained its AI model on users' sensitive conversations without the explicit consent the GDPR requires, and ordered it to fix that and tell users clearly within four months. The same decision found the app could train on ordinary conversation data under its "legitimate interest." In this decision, the line runs between ordinary chats and chats that reveal sex life or sexual orientation, or other special-category data, and only the second kind needs a yes from the user before it goes into training.

We read the 35-page decision, published in French on November 17, 2026, the authority's press releases in French and Dutch the same day, and the GDPR's text on EUR-Lex. Quotes from the decision are our translation from the French. This isn't legal advice. The Chamber deserves credit for showing its work. It walks through how the app builds its model step by step, and it explains why it set aside its own inspection report.

What did the Belgian DPA decide about the chatbot app?

That the app had no legal basis to train on sensitive chats, and wasn't open enough with users about it. The decision describes an app that lets users create chatbots and talk to them, including conversations "de type erotic role play," which it shortens to ERP. Two users complained on April 15, 2023 that the company used their personal data to train and run its chatbots without a valid basis.

The Chamber split the data in two. Ordinary conversation data is the personal data in the chats. The complainants had shared their names, jobs, and birth dates in chats, and each conversation was tied to their user ID. Special-category data, the kinds Article 9 of the GDPR prohibits processing except in listed cases, includes "data concerning a natural person's sex life or sexual orientation." Citing the Court of Justice's December 2, 2025 Russmedia judgment, the Chamber held that data revealing those things indirectly, by deduction, counts too, and that ERP conversations reveal them about the person who starts them.

Which chatbot company was it?

The decision doesn't name it. It calls the company "Y," and the Chamber says it isn't necessary to mention "the identification data of the parties" directly when publishing. It gives three facts about the company: it has no establishment in the European Economic Area, no other EEA authority had opened an investigation into it as of October 21, 2023, and its turnover is under 50 million euros.

Why was training on ordinary chats allowed?

Because the Chamber found the three-part legitimate interest test passed. The company said it uses conversations to "identify trends, adjust tone, and ensure the coherence of responses." The Chamber accepted that improving the model is a lawful, real, and present interest, that conversation data is necessary for it, and that users of a chatbot "could reasonably expect" their conversation data to be used to improve the model.

In the Chamber's words, users share the data "voluntarily," and the company pseudonymizes it before training. The Chamber added that the balance "depends heavily" on those minimization measures. It also named risks on the users' side: loss of control over intimate conversations, and a feedback loop in which models trained on users' chats "reproduce or amplify" gender or racial bias.

If you'd rather a company not use your chats this way at all, the GDPR's right to object applies to processing based on legitimate interest. We cover how to use it in how to object to AI training under the GDPR.

Why do sensitive chats need explicit consent for training?

Because consent covers only the purpose it was given for. The Chamber held that when you share something sensitive in a conversation, that counts as explicit consent under Article 9(2)(a) for the chatbot to use it in replying to you. Article 9(2)(a) allows processing where the person "has given explicit consent to the processing of those personal data for one or more specified purposes." Training is a different purpose. The consent "is granted for processing carried out in the context of using chatbots and not for optimizing the AI model," and the company showed no other basis, so the Chamber found a breach of Articles 6(1) and 9(2).

The company's privacy notice didn't help. Its March 14, 2021 version said nothing about special-category data. Its January 21, 2025 version said, in English: "During chats, you might voluntarily share personal data, including sensitive information." The Chamber held that this "brief mention" doesn't meet the duty to inform. Nor does the app's in-app notice warning that replies may be "inaccurate, false or fictional," which tells users they're talking to an AI but not what happens to their data.

Does removing your user ID make chat data anonymous?

No, according to the Chamber. The company said it strips user identifiers from conversations, masks or replaces personal details, and so "fully anonymizes" the data before training. The Chamber called the result pseudonymized, for two reasons from the company's own impact assessment: the company keeps identifiers separately so it can re-link conversations to a user, for example to answer an access request, and sensitive details stay tied to a person "on the basis of the data contained in the conversation" itself. The GDPR's recital 26 says pseudonymized data that could be attributed to a person "should be considered to be information on an identifiable natural person."

The company also said it uses some synthetic data. The Chamber noted the file had no information on how that data is made, and said synthetic data is only part of the training set, so the set as a whole isn't anonymous. We look at what other AI companies strip before training in are AI chats anonymized.

What does the company have to change, and by when?

Two things, within four months of being notified. The first order requires a way for users to give explicit consent to training on their special-category data, collected "prior to the start of a conversation with a chatbot during which the user is likely to share special-category data (e.g., an ERP conversation)." The second requires the company to:

  • tell users clearly that their special-category conversation data is processed to run the chatbots,
  • say that this rests on explicit consent, and how that consent is collected for each processing operation,
  • add a short prevention message to the app's existing in-app notice, and
  • add a reminder that sensitive data a user may share in a conversation "will be used to optimize" the company's AI models.

The company must send the Chamber proof: the updated privacy notice, the text of the in-app notice and what triggers it, and the consent mechanism. The decision dates each step of the case:

DateWhat happened
April 15, 2023Two users file the complaint
April 27, 2023Declared admissible; the Litigation Chamber takes the case
July 19, 2023The company has no EEA establishment, so the authority asks other EEA regulators under Article 61 of the GDPR whether they have a case
October 21, 2023None has opened an investigation, and none confirms an establishment in its country
June 14, 2024The Chamber sends the case to the Inspection Service
January 21, 2025The company updates its March 14, 2021 privacy notice
April 10, 2025The inspection closes
July 25, 2025The Chamber opens the merits on five grievances
November 16, 2026Decision 320/2026
November 17, 2026Published, with press releases in French and Dutch
30 days after notificationDeadline to appeal to the Market Court (Brussels Court of Appeal)
45 days after notificationDeadline to ask the Chamber for more time
4 months after notificationCompliance deadline; penalty payments can start after it

How much could the company have to pay?

At most 200,000 euros, and only if it misses the deadline. The press releases give the penalty as 1,000 euros per breach per day, 2,000 euros a day in total. The decision adds the limits. Each penalty starts only on the day the Chamber, after reviewing the company's proof at the four-month mark, notifies it that it has complied partly or not at all, and each "cannot exceed 100,000 EUR."

There's no fine. The Chamber gave a reprimand for the 2021 privacy notice, which named its legal bases for conversation data without saying which contract or which legitimate interest they meant, until the January 2025 update. It said the orders already force the company "to structurally transform" how it works, credited the measures it had taken, and said the penalties would push compliance "without unduly weakening the financial stability of the company." For the fines other regulators have issued against chatbot makers, see our table of AI chatbot privacy fines and orders.

Did the decision look at minors or vulnerable users?

No. The complaint raised risks for minors and for people whose mental health could suffer from the app, but the Chamber found the two complainants are adults, don't represent minors, and nothing in the file shows they are at risk, so it didn't assess that processing. It dropped four other grievances too, on purpose limitation, data minimization and privacy by design, transfers outside the EU, and accountability.

Those went because of a missed deadline. The Chamber sent the case to its Inspection Service on June 14, 2024, more than a year after the complaint was declared admissible, past the 30 days Belgian law allows. So it set the inspectors' findings aside, and without them it found the four grievances no longer "sufficiently characterized." The Chamber also said the company isn't responsible for processing by the third parties that build the two open-source models it fine-tunes, and didn't rule out joint responsibility for other processing. We compare what 14 companion apps' own policies say about training in AI companion app privacy.

Why Morse has no training consent to ask for

Every finding in this decision follows from the app keeping users' conversations, intimate ones included, in a training database and feeding them to its model. Once they were there, the only questions left were legal ones about which basis, which consent, and which notice.

Morse doesn't train AI on your chats. It also doesn't sell your data or show ads. Your chat history, character chats included, along with the characters you make and your personas, is stored in your browser, encrypted, not on Morse's servers, and our servers don't keep your messages or log what you write. Morse is for adults only, and every user confirms they're 18 or older.

For conversations in Chat, you choose a privacy mode for each chat:

  • E2EE, for the subjects Article 9 protects. Your browser checks the enclave's hardware attestation, then encrypts the message so only that enclave can read it, and Morse's servers relay text they can't read. The enclaves are run by third-party providers, not by Morse, and the guarantee rests on the enclave hardware and attestation working as designed. E2EE runs open-weight models with no web search or image generation, and comes with Pro at $19 a month.
  • Private, for chats you don't want any host to keep. It's the default for new chats. Morse has zero data retention agreements with its inference, search, and other providers, so they've agreed to keep nothing, and where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
  • Anonymous, for GPT, Claude, or Gemini without the provider knowing it's you. Our proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider still sees the text without knowing it's you, and may keep it for a limited time to check for abuse, under its own policy, so leave out names and details you'd rather not share.

Because we don't store chats, we usually have very little to give authorities. Clearing your browser or losing your device deletes your chats, and Morse can't recover them, so save any story you'd hate to lose. The details are on our privacy page.

You can send a first message from the Morse home page without an account, or start free with unlimited character messages and no card.

Common questions

Which chatbot company did Belgium's data protection authority order?

The decision doesn't say. It calls the company 'Y', describes it as having no establishment in the European Economic Area, and puts its turnover under 50 million euros. The authority published the decision without the parties' identifying details.

Was the chatbot company fined?

No. The Litigation Chamber gave it a reprimand and two orders backed by penalty payments of 1,000 euros a day each, and said a fine wasn't needed because the orders already force structural changes and the company had started fixing things.

Can the company appeal the Belgian decision?

Yes, within 30 days of being notified, to the Market Court, part of the Brussels Court of Appeal. It can also ask the Chamber for more time to comply within 45 days of notification.

Does sharing something sensitive in a chat count as consent?

For running the chatbot, the Chamber said yes: choosing to share it in the conversation counts as explicit consent for that purpose. For training the model on it, the Chamber said no, and the company needs a separate explicit consent.

Ask what you actually want to ask.

80+ models, characters, and images, with every chat encrypted and kept on your device. Free to start.

Start a private chat

Read next