How to use AI labeling for Icons & Images
Good to know - Technical constraints & behaviors:
- Individual Item Analysis: The AI labeling system evaluates items individually instead of as a group.
When you upload multiple items at once, the AI analyzes each one separately. For example, if you upload two items quickly, the AI might not notice that one is a broken link while the other is valid; it may just identify them both as "links.
- Abstract Logo Interpretation: For logos, the AI tries to extract only the company names, as other details might not be relevant. However, abstract or less-known logos may not provide enough context for accurate identification.
For instance, Stripe's abstract tilted square logo might be seen as a "shape," leading to suggestions related to squares or tilted shapes.
- Label Linking Approaches: We only suggest links to existing categories and subcategories, while we can also generate tags.
- Subcategory Linking: If the AI is confident about a subcategory but lacks understanding of the overarching category, it will attempt to make the best guess.
- Redundancy Threshold (Stop Logic): We have set up a prompt to tell the AI when to stop generating new labels.
If the AI decides that no more labels are needed, an admin will receive a notification stating that the AI does not see the need for additional labels. If existing labels are sufficient, the AI will issue a warning but will not change the current view.
- Guidance for Category and Subcategory Descriptions: A mild prompt has been added to help the AI with linking decisions. Clear prompts are essential, as vague wording may not yield the desired results.
For example, if you have a category "Size" with subcategories "Portrait" and "Widescreen," describing "Widescreen" as “Exclude images taller than they are wide” will likely be followed by the AI (not 100% guaranteed tho). However, a vague prompt like “Include all vertically long images” may not work as intended.
Updated on: 12/18/2025
Thank you!