# Your data and the AI. Dimmy. > Where your words go, who can see them, who could train on them. Local when you choose local models, telemetry only with your yes, and an honest per-provider table for when you choose the cloud. Plain language, English and Italian. Source: https://dimmy.app/data-and-ai/ your data & AI # Your words. Your choice. Dimmy turns your voice into text with AI. This page explains, in plain language, where your words go, who can see them, and who could train on them. Short version: with local models, nobody, because nothing you say or type leaves your computer. Verified: November 2026 ## The three states of your data Offline ### 1\. Local. Nothing leaves. When you choose local models (on a Mac Parakeet is the default; on Windows you pick local or cloud during setup), Dimmy transcribes on your machine (Whisper, Parakeet and more) and can rewrite with an optional local LLM (about 2.5 GB). Dictation audio is not kept unless you turn on audio in history; meeting recordings are saved on your computer. No account, no API key, works offline. Your words never reach us, or anyone: there is no Dimmy server involved. For your dictated content, we are not in the loop at all. Only if you say yes ### 2\. Telemetry, only with your yes. Numbers, never words. At first launch Dimmy asks whether it may send usage statistics and crash reports; nothing is sent until you say yes. If you do, we send event names and numbers (counts, durations, error categories) to PostHog EU and Sentry EU, linked to a random installation ID: pseudonymous, not anonymous. Never audio, never text, never prompts, never API keys. IP addresses are not sent. Two separate switches in Settings → Privacy. The full event-by-event list, with payloads, is on the privacy page. Your call ### 3\. Cloud, if you choose it. With your key. You can plug in a cloud provider for faster or fancier models, with your own API key or your own subscription. Then, and only then, your text or audio travels: directly from your computer to that provider, never through us. Which provider sees your words is your choice, so we tell you exactly what each one can do with them. That's the table below. The same applies to any integration you switch on: enabling it sends the relevant content (for example a recap to Notion) to that service, and to no one else. ## If you use your own key: who can train on your words? Checked on the providers' official documentation, July 2026\. Policies change: re-check before sending anything sensitive. "No" means the provider states it does not use API data to train its models by default. | Provider | Trains on your data? | Worth knowing | | ----------------------------- | -------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | OpenAI API | No | No training on API data by default. Different product from consumer ChatGPT. | | Anthropic API | No | No training by default. | | Groq | No | No training; zero-data-retention available in the console. | | Deepgram | No | Trains only via its Model Improvement Program. Standard API rates can include it: send mip\_opt\_out=true to stay out. | | Google Gemini (paid) | No | Paid API: no training on your data. | | Google Gemini (free tier) | Yes | The free API tier may use your content to improve products, including human review. Don't send anything sensitive here. | | Together AI | No | No training; zero-retention toggle in settings. | | Fireworks AI | No | No training by default. | | OpenRouter | Depends | Depends on the downstream provider. ":free" models often allow training: turn off trainable providers in your OpenRouter settings. | | Claude subscription (via CLI) | If enabled | Consumer accounts train by default (opt-out): the "Help improve Claude" toggle is on unless you turn it off in your claude.ai settings. Some safety-flagged chats may still be used regardless. | Two honest traps to remember: the Gemini free tier, and a consumer Claude subscription with "Help improve Claude" left on (it is on by default). Everything else, with your own paid key, does not train on you by default. ## Meetings: consent first Recording a meeting involves other people's words, not just yours. When you start a meeting yourself, a dialog asks you to confirm you have everyone's consent, offers a spoken and copyable announcement for remote participants, and keeps a local, append-only log of what was confirmed and when. If you turn on automatic recording of calls, you confirm once in Settings; after that each call starts recording without the dialog, and Dimmy plays the announcement and copies it so you can paste it in the call chat. People on the call may not hear it. The audio file, the transcript and the recap are saved on your computer; with a cloud provider, audio and text also go to that provider. Dimmy gives you the tools to ask for consent, but obtaining valid consent, and complying with the recording laws that apply to you, remains your responsibility. ## Where Dimmy stands on the EU AI Act Dimmy is an AI system, and we are its provider under the EU AI Act (Regulation 2024/1689). Based on our analysis it sits outside the heavily regulated tiers: voice dictation, transcription and meeting summaries are not among the prohibited practices, nor among the high-risk uses listed in Annex III. At that level no conformity assessment or certification applies, and none exists to be obtained, so you will never see us claim one. What does apply is transparency, and that deserves a precise answer rather than a slogan. The transparency rules (Article 50) apply from 2 August 2026\. Following the Commission guidelines adopted on 20 July 2026, here is how they land on a tool like this. Dictation, transcription and grammar or style rewriting count as standard editing: they don't change the substance of what you said, so no marking duty attaches to them. Summaries are a different matter. Condensing an hour of meeting into a recap does restructure meaning, so an AI-generated recap has to carry a mark, readable by machines, saying it came from AI. Dimmy already adds that marking: every recap saved to recap.md carries a machine-readable tag (), and the app shows "AI-generated summary. Review it before sharing." above it. Dimmy never passes for a human, and it is clear that transcriptions, rewrites and recaps come from AI. Not every output waits for you: dictated text is pasted where you are typing, history is saved, a recap runs when a meeting stops, and a recap can go to Notion automatically if you turn that on. Review AI output before you share it. We don't train any model on your data, and Dimmy does not identify who is speaking from their voiceprint, does not infer emotions, and does no biometric categorisation. If our reading turns out to be wrong, or the guidance moves, we will say so on this page and fix it. ## Don't trust us. Check. The code is public under AGPL-3.0, every telemetry event is listed with its exact payload on the privacy page, and the app logs locally everything it sends. Claims you can verify beat promises you have to believe. [Technical privacy page](https://dimmy.app/privacy) [Terms](https://dimmy.app/terms) [Source code](https://github.com/KonradDallaOrg/dimmy) This page is a plain-language explanation, not legal advice, and provider policies are theirs, not ours: always check the provider's current terms before sending sensitive data. Everything here reflects how Dimmy works to the best of our knowledge on the date shown above.