I wonder whether language models could be used to solve this problem. Convert every tag to a latent space vector and show the user all of the closest existing tags in a dropdown with the option to create a new one at the bottom.

That would be a cool tool to have. In the end, it still comes down to humans being conscientious every time they add something to the library. The equivalent would be having really nice software development lifecycle processes: test suites, merge request gating, etc, and once in a while (ha ha) people still manage to push out half-baked changes when they're in a hurry. In the end it's a human issue, tools only do so much. And in domains outside of coding, it's harder to find regressions automatically, so they tend to go uncaught and accumulate.

Taking this a bit further, do we actually need “tags” that are identical? Rather than filter by a tag, perhaps you could enter a label or set of labels and sort by which results minimize a distance function in latent space. No matching or conscientiousness necessary.