System tags and user tags need a visible source

What it is

Words on an object come from two places: people and the system (type detection, spam, language, invoice, AI keywords). Source has to be readable on the chip—shape, region, or an explicit Auto—because trust and disposal differ completely. Human words can be edited, deleted, used as a retrieval cue. System words may return on refresh, or should never be a personal class. This is not whether the vocabulary is open, and not AND versus OR.

Why it happens

People treat every chip they see as their own classification promise. Mixed in one row, system words get used as retrieval cues and deleted as errors; they come back, and the product looks broken. The other way: human words that look like system badges are left untouched, and the vocabulary freezes. A visible source matches disposal to origin: system words offer “wrong type” correction, not chip deletion; human words offer delete and rename. A flat filter list mixing both makes “auto: invoice” and a human “invoice” look AND-able, while one drifts with detection and the other with the person, so the result set is unstable. Source also splits blame: a wrong system tag is a model issue; a wrong human tag is a vocabulary issue. Mixed, neither can be fixed.

Studying it

Put human words and an automatic type on one object. Ask people to “remove marks that should not be here” and “find this class with invoice.” Compare mixed with no source, partitioned, Auto written on the chip.

Independent variables: whether source is visible, whether removing a system word is a true delete or a correction, whether the filter list is grouped. Dependent variables: confusion after deleting a system word that returns, failures from treating system words as a stable cue, reluctance to delete one’s own words.

A lab that says “some of these are automatic” spoils the read. Source should come from the chip. When detection updates a system word, watch whether people notice the source. Do not count permission role names as tags.

Where it stops holding

A product with only hand-applied tags has no split to show. System words used only for internal ranking and never shown need not become chips. A legally required provenance mark (“this line was generated”) may be heavier than a tag chip; a weak color difference is not a substitute. Words from a user rule (“if title contains invoice, mark it”) sit between; label them Rule, neither hand nor model.

Applying it

• Put human and system words in two rows or two chip styles; mark system as Auto or place them in Detected.
• The negative act on a system word is correct type, with a note that detection will learn or will not mark this again. The negative act on a human word is delete.
• Group filters: my tags / system detection. Do not AND the two classes in one unexplained operation.
• Verify: ask someone to remove an automatic type chip; watch the act’s name and whether it returns. Then retrieve with “invoice” and ask whether they relied on detection or their own word. If they cannot name the source, the split failed.

Related

• Within the group: H8.10.1 Tag systems need user-defined labels, not only presets · H8.10.2 Multi-tag retrieval must state AND versus OR · H8.10.3 Unbounded tagging destroys retrieval value
• Adjacent: L1 Intelligent System Interaction · H8.13 Content Search and Location · H8.05 Save for Later
• Search terms: tag provenance · machine tags · auto classify