AI provenance

In data products, people can't tell which values come from the source and which an AI produced. So they either trust everything or nothing.

Mark every AI-generated value with an AI label. Opening it shows why the value was produced and where it came from, with accept, reject and undo. See the risk alert table for a working example.

The AI label is built inside the risk alert table. It is not a design system component yet. It uses:

• Button: Accept, Reject and Undo
• IconButton: closes the popup
• The Base UI Popover for the popup, used directly
• The color.ai tokens for the label: color.ai.label-background, color.ai.label-content and color.ai.border

• The AI label is a button. Its name starts with "AI", then says what it marks and whether it has been reviewed, for example "AI sentiment, not reviewed". After a decision it says "accepted" or "rejected".
• The popup is marked as a dialog and named by its title, for example "AI-rated sentiment".
• After Accept or Reject, focus moves to Undo. After Undo, focus moves back to Accept.
• Escape closes the popup and focus goes back to the label.
• Rejecting a value changes the cell: a summary shows the source's own title, and a sentiment shows "Not rated".
• Automated tests check all of the above, and run axe with the popup open in both themes and both densities.