Social classification, or a folksonomy, a term coined by Thomas Vander Wal, is what emerges when ordinary users tag content themselves, rather than trained indexers applying a controlled vocabulary. Folksonomies are flexible and reflect genuine diversity of perspective, since different people naturally use different words for the same thing, or the same word for different things. Clay Shirky's assessment is a fair summary of where they stand: a folksonomy isn't necessarily better than a controlled vocabulary, but it's better than having no classification at all.
Their core weaknesses are structural: folksonomies can't express equivalence, hierarchical, or associative relationships between terms the way a thesaurus can, and the resulting lack of vocabulary control creates real findability problems. But folksonomies aren't necessarily a dead end, the tags people actually choose can directly inform the design of a more formal controlled vocabulary, revealing real user language a team might not have anticipated.
A folksonomy won't organize your content for you, but the words people freely choose to describe it are some of the most honest vocabulary research you'll ever get for free.
The inability to express equivalence relationships is the correct answer. A folksonomy has no built-in mechanism to declare that "JS," "Javascript," and "JavaScript Programming" all mean the same skill, each member tags independently, with no shared preferred term to converge on. This is exactly the structural weakness the source material identifies: folksonomies lack the vocabulary control that a thesaurus's variant-term mapping provides.
This doesn't mean three years of tagging data should be discarded, quite the opposite. Those tags are a genuine, large-scale record of the actual language members use to describe their own skills, which is precisely the kind of user-warrant evidence the controlled vocabulary development process calls for gathering early on. Rather than a team guessing at preferred terms from scratch, they can treat the folksonomy as a rich terminology-gathering source: analyzing which variant tags cluster around the same underlying skill, using that clustering to define preferred terms with strong user warrant behind them, and mapping the existing variants as the controlled vocabulary's variant terms.This is a good illustration of the source material's broader point about folksonomies: even though social tagging alone can't solve the platform's findability problem, it doesn't need to be thrown away once its limits are reached, it can become the raw material for the more structured controlled vocabulary that actually can.