How Munwyn uses AI
Which feature uses which model, what leaves your device, what never does — and how to turn it off.
The short version
Munwyn writes back to you. Insights, the Sunday letter, counter-arguments and summaries are written by a language model — not by a person, and not by us reading your notes. Everything a model wrote carries a label in the app and a machine-readable marking in the data.
Search, related notes and voice transcription run on our own server and never leave it. Only the features marked “cloud” below send text to the model provider, and only when you have cloud processing switched on.
Feature by feature
| Feature | Model | Runs | What is sent | AI Act |
|---|---|---|---|---|
| Sunday & monthly letter | Claude (Anthropic) | cloud | Text of notes from the period, excluding those marked “Don’t send to AI” | 50(1)+50(2) |
| Insights | Claude (Anthropic) | cloud | Titles, excerpts and topics of recent notes | 50(1)+50(2) |
| Counter-argument | Claude (Anthropic) | cloud | The one note you asked about, plus older notes that contradict it | 50(1)+50(2) |
| Note summary, topics, tone, language | Claude (Anthropic) | cloud | The text of that one note | 50(2) |
| Ask your archive | Claude (Anthropic) | cloud | Your question and excerpts of the notes found for it | 50(1)+50(2) |
| Open loops, blind spots, compost patterns | Claude (Anthropic) | cloud | Titles and excerpts of the notes concerned | 50(2) |
| Thinker profile, mood patterns, Wrapped narrative | Claude (Anthropic) | cloud | Aggregates and topic names only — never the text of a note | 50(2) |
| Connections between notes | Claude (Anthropic) | cloud | The pair of notes proposed by local similarity search | 50(2) |
| Photo description | Claude (Anthropic) | cloud | The image you attached | 50(2) |
| Semantic search, related notes, area suggestions | MiniLM (multilingual, 384 dim) | our server | Nothing leaves the server | — |
| Voice transcription | Whisper (small) | our server | Nothing leaves the server | — |
The exact Claude model version in use is recorded with every generated text and is visible in your sending log (Settings → Privacy and AI). We do not fine-tune the model, so Munwyn is a provider of an AI system, not a provider of a general-purpose AI model.
What we never send
● Notes you marked private
“Don’t send to AI” on a single note excludes it from every cloud feature above — including letters, insights and patterns that would otherwise quote it.
◉ Everything, if you switch cloud off
Settings → Privacy and AI turns off cloud processing entirely. Capture, search, related notes and voice transcription keep working — they never needed the cloud.
◆ Your credentials and attachments
Passwords, two-factor secrets and API key fingerprints never go to a model. Attachments go only when you ask for a photo to be described.
⌛ Nothing for training
What Munwyn sends is not used to train the provider’s models. Transfers outside the EU are covered by Standard Contractual Clauses.
How you can tell a text was written by AI
In the app and in email. Every text a model wrote carries a visible label, and emails that carry AI text say so above the footer. You are told at first launch, before you write anything.
In the data. API responses and the GDPR export carry a machine-readable marking:
ai_generated, the model that wrote it (ai_model), the provider and
the timestamp. Where only part of an object is synthetic — a note you wrote yourself
that has an AI summary — ai_fields names the fields the model filled in.
Public share pages carry the same information as <meta name="ai-generated">.
What is not marked. The text of your own notes, letters you sealed to your future self, and the cleaned-up transcript of a voice note — cleaning a transcript is an assistive editing function that does not substantially alter your input or its meaning.
What Munwyn deliberately does not do
- No emotion recognition from biometrics. Munwyn reads tone from the text of a note. It never analyses your voice, face or any other biometric signal — which is why it falls outside Article 50(3). In a workplace this is not merely restricted but prohibited under Article 5(1)(f), and that boundary is a hard rule for us.
- No generated images, audio or video. Recognising what is in a photo you took is not generating one, so the deepfake rules in Article 50(4) do not apply.
- No decisions about you. Munwyn scores nothing, ranks nobody and feeds no decision about employment, credit, education or access to services. It is not a high-risk system under Annex III.
- No profiling for advertising. There are no ads, no ad networks and no third-party trackers in the app.
For a security questionnaire
⚖ Our role
Munwyn is a provider of an AI system under Regulation (EU) 2024/1689. We build on a third-party model and sell under our own name; we do not fine-tune it, so we are not a provider of a general-purpose AI model.
▤ Transparency obligations
Article 50(1) and 50(5): disclosure at first interaction and a persistent label on AI output. Article 50(2): machine-readable marking of generated text. Article 50(3) and 50(4) do not apply — see above.
▣ AI literacy (Article 4)
A dated record of AI literacy for the people operating the system is kept in our documentation and reviewed at least annually.
⚑ Supervision
In the Czech Republic the general supervisory authority for the AI Act is the Czech Telecommunication Office (ČTÚ). Data protection questions go to the Office for Personal Data Protection (ÚOOÚ).
Last updated: 21 August 2026 · Questions about anything on this page: · See also the Privacy & GDPR page and the Terms of Service.