AI Notice
Reference translation. The Korean original prevails in case of any difference.
Effective date: October 11, 2026 · Version 1
SAYNARY is a service that uses generative artificial intelligence. In line with the transparency aims of Korea's Framework Act on the Development of Artificial Intelligence and the Establishment of a Foundation for Trust (the AI Framework Act), the Company provides the following information so that users can recognize and check results created by AI.
1. Where AI Is Used
Generative AI (Anthropic's Claude, USA)
| Feature | What the AI does |
|---|---|
| Minutes, briefs, summary cards | Organizes a one-line summary, decisions, action items, issues, commitments and numbers from the transcript |
| AI titles, conversation types | Suggests the recording title and the kind of conversation (meeting, call, consultation, etc.) |
| AI context correction | Suggests corrections for words that appear to have been misrecognized |
| Ask AI, voice mode | Answers questions about one recording or several (adding general knowledge or web search results if you want) |
| Organization AI questions | Answers within the recordings an organization admin can see |
| Briefings, period summaries | Updates since the last meeting, summaries of missed meetings, suggested questions to check before a meeting |
| Speaker name suggestions | Suggests who a speaker is from the content of the conversation |
| Decision change and contradiction checks | Checks whether something differs from an earlier decision |
AI on the Company's servers (no external transfer)
Speech recognition (transcription), speaker separation, precise re-transcription, semantic search, voiceprint recognition and dictation are processed with AI models on the Company's servers (Republic of Korea). Minutes and summaries of recordings that do not use external AI are created with rules.
2. How AI Output Is Marked
- Below AI-created minutes, briefs, answers and summaries we show "Created by AI. Check important details against the original."
- Important items (decisions, action items, issues, numbers, commitments, answers) carry their source (recording time, original utterance, speaker). Tapping a source takes you straight to the transcript and to the audio of that moment.
- The server checks whether the original text quoted by the AI really exists in the transcript, and shows "needs source check" if it does not.
- When it is not clear who spoke, the speaker is not asserted but marked uncertain. Questions, proposals and opinions are not shown as decisions, and decisions are classified as confirmed, proposed, on hold or withdrawn.
- Parts answered with general knowledge that is not in the recording are set apart as "Good to know · not from the recording", with sources when web search was used.
- Names added automatically by voiceprint are marked (🎙) with their match score and can be confirmed with "That's right / That's not right".
- Words corrected by AI are shown together with the original word and can be reverted.
- Parts where the recording was interrupted are marked as a "recording gap", and the AI is kept from guessing what was said there.
- On each recording's 🔒 privacy screen you can see where each step was processed (on-premises server / Claude API).
3. Not Used for Training
- Your recordings, transcripts, minutes and questions are not used to train AI models. Anthropic also does not use data received through the API for training.
- The only exception is when you separately consent to training use in the settings and your organization allows it (off by default); see the Privacy Policy for details.
- Before anything is sent to Claude, resident registration numbers, card numbers, account numbers, phone numbers and email addresses are replaced by masked labels (default). Recordings caught by security keywords, or whose organization has turned external AI off, are not sent to Claude.
4. Limits and Possible Errors
AI can be wrong. Pay particular attention to the following:
- Mishearing: Noise, overlapping speech, dialects, technical terms, names and numbers can be transcribed incorrectly.
- Speaker confusion: People with similar voices, or who spoke only briefly, may be split into different speakers or merged.
- Summary errors: Things may be left out, the context misunderstood, or something not actually settled organized as if it were a decision.
- Answers that differ from the facts: The AI can plausibly make up content that is not in the recording (check the sources shown). General knowledge may be out of date.
- Legal, medical, tax, investment and similar matters: AI answers are general information only and do not replace a professional's judgment.
For important matters such as contracts, amounts, personnel and legal decisions, always have a person check the original (recording and transcript) before use. AI output is reference material that helps people decide; the Service does not make automated decisions about users or the other parties to a conversation.
5. How to Object or Correct
| What you want to do | How |
|---|---|
| Correct a result | Edit transcripts, speaker names, minutes, action items and decision status directly. Regenerate the minutes |
| Revert an AI correction | Tap the corrected word to restore the original |
| Wrong voice recognition | Tap "That's not right" to correct it (it learns not to make the same mistake again). Delete the voiceprint |
| Avoid AI use | Ask your organization admin to turn off external AI or set security keywords; choose to answer only from the recording |
| Report a problem / contact us | support@metavu.io (telling us the recording title, the screen and the time helps us check quickly). The Company will tell you what it found and what it did |
6. AI We Use
| Category | Provider | Where it is processed |
|---|---|---|
| Generative AI (summaries, answers, corrections, etc.) | Anthropic, PBC — Claude API | United States |
| Web search (general-knowledge answers, if the organization turns it on) | Anthropic web search tool | United States |
| Speech recognition, speaker separation, search, voiceprints | Open-source-based models operated by the Company | Company servers (Republic of Korea) |
The models in use may change to improve quality; if the place of processing or the provider changes, we will announce it in advance.
| Version | Effective date | Description |
|---|---|---|
| 1 | 2026-10-11 | First version |