AI Transparency Policy
In short: minumeeting works on your device by default. Cloud AI runs only if you choose it and agree. Every AI result shows where it was made, and every AI result has a "Report" button.
1. Three ways to run AI
You choose the mode in Settings → AI engine. A new install starts with On this device only.
| Mode | Where it runs | What leaves your device | Cost |
|---|---|---|---|
| On this device only | On your phone, tablet or computer | Nothing (apart from the one-time engine downloads described in the Privacy Policy) | Free |
| Smart — Cloudflare | Our AI service on Cloudflare Workers AI | The meeting text needed for the feature you run and, if you choose cloud transcription, the meeting audio while you record — only after you agree on the consent screen, which names Cloudflare, Inc. and asks separately for text and for audio | Free within a daily and weekly allowance |
| My own key | Directly at the provider you choose: OpenAI, Anthropic, Google Gemini, Groq or OpenRouter | The meeting text needed for the feature you run, sent from your device to that provider with your key — only after you agree for that provider | Billed by the provider to your account |
If the chosen mode cannot be used at that moment — no internet, allowance used up, the service paused, a key rejected — the app makes the result on your device instead and tells you why. It never switches to another cloud provider on its own.
2. Which feature runs where
| Feature | On this device | Smart — Cloudflare | My own key |
|---|---|---|---|
| Live transcription while recording | Whisper (tiny, base, small or large-v3-turbo — chosen to fit your device, changeable in Settings), on your device. In the web app you can choose your browser's Web Speech service instead | Whisper large-v3 turbo, if you choose "Smart — Cloudflare" as the transcription method and allow meeting audio; any piece it cannot send is transcribed on your device | Not available |
| Analysis when a recording ends (summary, key points, topics, action items, decisions) | Always on your device: rules that extract what was said, or — if you choose it in Settings — a small language model on your device | — | — |
| Summaries (short, detailed, executive) | Sentences extracted from what was said | Written by the model, citing transcript lines | Written by your model, citing transcript lines |
| Action items (owner, deadline), decisions, risks, open questions | Found with rules in what was said | Written by the model, citing lines | Written by your model, citing lines |
| Chapters and title suggestions | Simple rules | Written by the model | Written by your model |
| Formal Arabic minutes and follow-up email | A template filled with the items found on your device | Written by the model in formal language | Written by your model |
| Next meeting agenda | Assignments and anything said to be postponed or not settled, found with rules in what was said, without lengths | Proposed by the model from what the meeting left open, citing lines, with suggested lengths | Proposed by your model, citing lines |
| Ask your meetings | Keyword search, with the meeting and line of each result | Search by meaning: embeddings are computed in the cloud, then stored and searched on your device; the answer shows the closest passages as they were said | Search on your device, then an answer written by your model that cites the passages |
| Translation | OPUS-MT models on your device for Arabic↔English and English↔French, Spanish, German and Chinese | The language model (or M2M-100) | Your model |
| Read aloud | Your device's voice | MeloTTS for English, Spanish, French, Chinese, Japanese and Korean; Arabic always uses your device's voice | With an OpenAI key, OpenAI text-to-speech; otherwise your device's voice |
| Speaker suggestions from voiceprints (optional) | On your device only | — | — |
"On this device" results that come from rules are extracts of what was said, not newly written text; the app labels them "on-device rules".
3. Providers and models
On your device (downloaded once, then used offline): Whisper by OpenAI for transcription (tiny, base, small or large-v3-turbo, chosen to fit your device; you can change it in Settings); OPUS-MT translation models by the University of Helsinki; and, only if you choose the on-device language model, Qwen3 0.6B or Qwen2.5 0.5B Instruct (Alibaba Cloud). They run through Transformers.js and ONNX Runtime Web, on your graphics processor (WebGPU) when available.
In Smart mode — run by Cloudflare, Inc. on Workers AI for our service:
| Task | Current model |
|---|---|
| Summaries, action items, decisions, chapters, titles, minutes, emails, agendas, translation | Gemma 4 26B A4B (Google); if it fails, Qwen3 30B A3B (Alibaba Cloud) |
| Embeddings for "Ask your meetings" | BGE-M3 (BAAI) |
| Alternative translation engine | M2M-100 1.2B (Meta) |
| Read aloud | MeloTTS (MyShell.ai) |
| Live transcription (if you choose it) | Whisper large-v3 turbo (OpenAI) |
We may change Smart-mode models on our server without an app update, for example when a provider retires a model. We update this page when the default models change. Licences of all models are on the Licences page.
With My own key, the model is the one you choose from your provider's list ("Test key" shows the models available to your account).
4. How we keep results tied to your meeting
- The instructions we give the models tell them to use only the transcript and to cite the transcript lines each item comes from. Items without a valid citation are removed.
- Names of people who do not appear in the transcript or in the participant list are removed from owners and attendees.
- The model may give a calendar date only when a date was said explicitly. Relative deadlines such as "next Thursday" are calculated on your device from the meeting date and marked as calculated.
- The transcript is passed to the model as data, and the model is told to ignore any instructions spoken or written inside a meeting.
- Every result is checked against the expected structure before it is shown or saved.
- In My own key mode, the app uses the same instructions and the same checks.
5. Limits
AI can be wrong. Transcription can miss words or attribute them to the wrong speaker, especially with noise, overlapping speech, dialects, names and numbers. Summaries and minutes can leave out or misstate what was said; translations can change meaning. Results are not legal, medical, financial or other professional advice. Check every result against the recording before you rely on it or share it, and treat formal minutes as a draft to review and approve.
6. Labels on every result
Each AI result shows its source: On your device, Cloudflare (with the model), or Your key · provider. When a result was made on your device instead of the mode you chose, the app says why — for example "you're offline" or "today's free allowance on this device is used up".
7. Your consent and control
- Smart mode asks for your consent before the first use, naming Cloudflare, Inc. and asking separately for meeting text and meeting audio.
- My own key asks for your consent for each provider, naming the provider's company.
- You can withdraw consent at any time in Settings → AI engine; the app then works on your device.
- "Delete AI data on this device" (Settings → AI engine) deletes the install ID and session, the "Ask your meetings" index and your saved keys. See Delete your data.
8. Training
We do not use your meetings to train AI models. Cloudflare states that it does not use content processed by Workers AI to train models and does not keep it after the request. In My own key mode, your provider's own terms decide how it uses and keeps what you send — check them in your account with the provider.
9. Reporting a problem with an AI result
Every AI result has a Report button. Choose a reason — offensive, hateful, sexual, violent, harmful or dangerous, misleading or wrong, reveals private information, or something else — add a comment if you want, and decide whether to include the generated text. The report is sent only after you tick the consent box. It goes to our team through our service on Cloudflare, without an install ID or IP address, and is deleted automatically after 90 days. You receive a report number; quote it if you write to us about it.
You can also report by email to support@tibrcode.com. This includes results made with your own key or on your device.
We review reports and act on them — for example by changing the instructions given to the models, changing a model, or switching a feature off for everyone.
10. Uses we do not allow
Do not use minumeeting's AI features to create false records (such as minutes of a meeting that did not take place), to harass or deceive people, or to try to produce illegal, hateful, sexual, violent or otherwise harmful content. The full rules are in the Terms of Use.
11. Changes and contact
We update this page when features, providers or default models change, with a new date. Questions: support@tibrcode.com.