0003 AI training, by anyone, only with the depositor's consent
The Korean version is authoritative. This page was translated by AI (Claude) and has not yet been checked by a person. 한국어 원문
This settles the second item under “To look at separately” in 0001 (arkive protocol article 10 and principle 8).
- Principle 8 changes from “Not sold, not used for training” to “Not sold, not used for training without the depositor’s consent”.
- AI training happens only with the depositor’s consent, whether by Nodeone, an institution or a company. We do not draw the line by who does it.
- A fourth use consent field, “use for AI training”, is added. Off by default. At the time of consent, we show who is training and what is being taught, and the data is used only within that.
- What stays the same:
- There is still no field for selling, or for handing over records wholesale for training (arkive article 12).
- Connected model APIs (transcription, assistant conversations) are used only with settings that exclude training.
- Decided together with this:
- “Sealed” records, and passages of other people’s voices without that person’s consent, are not used even when there is consent.
- When consent is withdrawn, the data is not used from then on. At the time of consent we tell the depositor first that it cannot be removed from a model that has already learned from it.
- An AI that answers questions about content does not train on it; it reads and answers each time (arkive article 10, second part, as it is).
- manifest: a
trainfield inconsent, and"train"insent_to.purpose.
Documents updated: 00-Principles sections 1 and 4 (v0.2), 03-manifest sections 3 and 8 (v0.2).
- Decision of the head of Nodeone (2026-10-12): “Training, whether by an institution or by Nodeone, is possible only with the depositor’s consent.”
- It fits principle 1, “The original belongs to its owner”. The owner decides whether it is used. If Nodeone blocks it on the owner’s behalf with “never”, that also takes the choice away from the owner.
- arkive protocol article 8 (separate consent for each use, with “AI training” as one field) and article 10 (learning how to read, within consent) already have this shape. Now layer 0 and layer 1 say the same thing.
- If we split it as “not Nodeone, only institutions”, the same consent from the same depositor would mean different things depending on who does the training. One rule is simpler.
Other options
Section titled “Other options”| Option | Why not chosen |
|---|---|
| Nobody trains (the original principle 8) | It blocks institutions such as the Korean Studies Advancement Center from improving their transcription tools (arkive article 10). Even a depositor who wants it could not allow it |
| Not Nodeone, only institutions with consent | With the same consent, there is little reason to treat it differently depending on who does it. The head of Nodeone did not choose it |
| Allow only “learning how to read” | The consent field records “what is being taught”, so a depositor who wants to narrow it can narrow it there |
Where it has effect
Section titled “Where it has effect”- arkive: almost nothing to change.
ai_trainingin arkive’s policy code is already one use consent field. At the time of consent, also record who is training and what is being taught. - secretary: four consent fields on the onboarding screen. manifest
consent.train. (This is a data structure change, so it needs separate approval on the secretary side. There are no plans to train now, so there is no hurry.) Whether to change the wording of principle 8 in secretary’s working rules to match layer 0 is for the secretary session to decide.
0 won.