Large catalogues in this field usually run on crowd-sourced tags — anyone can add descriptive labels to an entry. It's the only approach that scales to hundreds of thousands of works, and it comes with a specific set of failure modes worth understanding.

Why crowd-sourced tagging wins on coverage

A centralised team can tag maybe a few hundred works a week. A community can tag everything, including obscure material nobody employed would ever get to.

Coverage is the whole advantage, and it's a big one. Catalogues that rely on uploader-supplied tags alone end up with thin, inconsistent labelling, because uploaders tag for discovery rather than for accuracy.

Community tagging also self-corrects over time on popular entries, since more eyes means more chances someone fixes a wrong label.

There's a second benefit that gets overlooked. Because taggers are also readers, the vocabulary that emerges describes what people actually care about rather than what a cataloguer imagined they would. Categories nobody uses quietly die out, which is a kind of maintenance no committee performs as well.

Where it breaks

Consistency. The same tag means different things to different taggers, and without definitions nobody notices the divergence until filtering stops working.

Three specific problems recur. Synonyms multiply, so the same concept exists under four labels. Scope creeps, so a narrow tag gradually covers everything adjacent to it. And unpopular entries stay under-tagged forever, because nobody visits them.

Namespaces are the biggest single improvement

Grouping tags into families — artist, group, language, format, character, element — solves more problems than any other change. It stops unrelated concepts sitting in one flat list, and it lets a filter target a category rather than a string.

It also makes the metadata usable outside the site, because a namespaced tag is close to a proper field. That connects directly to the six fields we argue for in the piece on metadata.

Flat tag lists are the older design and they degrade badly at scale, which is why almost every large system eventually adds namespaces.

Retrofitting them is painful, though. Thousands of existing tags have to be sorted into families by hand or by guesswork, and the guesswork produces exactly the scope-creep problems namespaces were meant to solve. Better to start with them, which almost nobody does.

Exclusion is the main use case

This is the part system designers underestimate. Most tag use is people avoiding things, not finding them, and exclusion is far less tolerant of inconsistency than search is.

If a concept exists under four labels and you've blocked one, you'll still see it. A search that misses something is an inconvenience; a filter that misses something is a failure. That asymmetry is why definitions and alias tables matter more than tag count — the argument in our piece on how genre tags work.

Moderation and the safety floor

Tagging is not moderation and can't substitute for it. A correctly tagged work that breaches the floor is still a breach: characters must be fictional and unmistakably adult, with no ambiguity about age, no real-person sexual depiction, and no coercion presented as romance.

Crowd tagging can help moderators find problems, which is a genuine benefit. It also creates a record of who flagged what and when, which is useful evidence. But the decision itself has to sit with people who are accountable for it, because a vote is not a review. Our standards are in the control room, and connected creation with real account and moderation controls stays on models.

What keeps a community tag system usable

Four things, in order of impact. Namespaces, so tags have families. Written definitions, so divergence is visible. An alias table, so synonyms collapse. And moderation with weighted trust, so vandalism doesn't propagate.

Systems with all four stay usable at enormous scale. Systems with none become a flat list of thousands of near-duplicate labels, which is functionally the same as having no tags at all — and worse, because people trust the filters.