AI training data and consent

Admin -> AI -> Training data has two parts: an export of your own workspace as a training set (a download for you, nothing leaves the instance), and the correction capture, which records what people do with AI proposals and shares it for training AI models. This page is about the capture and the consent it needs.

What is shared

While the capture is on, every AI proposal (a suggested title, labels, priority, estimate, summary, split, polished or translated text, an answer from Ask) is recorded with what the person did with it: accepted, edited or rejected, and the final value they set.

• User ids are pseudonymised and email addresses are removed.
• Private tickets, tickets in private milestones and attachments are never included.
• Resolved examples are relayed to the orboto control plane at most once an hour and used there to train models. Locally they are kept for 30 days.

Giving the consent

Added in v0.221.2 Under clause 3.5 of the terms of use, data of your workspace is used for training AI models only with your consent. The capture is off until an admin gives it, and the consent is an active opt-in: the switch does not flip by itself.

1. Open Admin -> AI -> Training data and choose Give consent.
2. The dialog states what is shared, the pseudonymisation and the relay, and quotes clause 3.5 word for word. Nothing in it is ticked in advance.
3. Tick the acknowledgement and choose Give consent and turn on.

orboto stores who gave the consent, when, and which version of the statement they confirmed, and writes an audit entry. When the statement changes, an earlier consent no longer counts: the page says the consent needs to be confirmed again, and nothing is captured until an admin does.

From the CLI, orboto admin ai-training consent prints the same statement and the version to accept; orboto admin ai-training consent --accept <version> gives the consent. orboto admin ai-training status shows the current state.

Withdrawing the consent

The consent can be withdrawn in total or for single projects. It can be withdrawn only for the future: data already used in training is not removed from the model.

• In total: Admin -> AI -> Training data -> Withdraw consent, or orboto admin ai-training withdraw. Nothing new is captured or relayed from then on. Examples already relayed are deleted on the control plane only if they have not been used in training yet; models already trained are not changed.
• Per project: a project admin ticks Exclude this project from AI training in the project settings, or runs orboto project set <key> --ai-training-excluded=true. From then on no AI proposal on that project is captured or relayed, and its examples not yet relayed are deleted. Unticking includes the project again.

Every withdrawal is recorded with the person and the time. The Training data page lists the excluded projects and the consent history.

"Purge local examples" on the same page deletes the copies kept in your instance. It does not touch examples already relayed to the control plane.