Your scenario on tagging documents about atomic energy reminds me of my frustration about serious problems with assigning keywords (the problem is the same for tags): subjectivity and consistency, when I was director of a computer information center many, many years ago. Back then, searches of our computer tapes could only be done on keywords.
Subjectivity: Different persons given the task of assigning keywords (or tags) to the same document will invariably differ in their interpretation of the important features or content of the document. The more complex the document, the worse this issue becomes. Worse, the same individual reading similar documents will display the same issue, depending on the individual’s reaction to the content of each document.
Consistency: Should the keyword (or tag) be ‘atomic energy’ or ‘atomic power’ or ‘nuclear energy’ or ‘nuclear power’? This is an example of a keyword or tag that should contain at least two terms, as we wouldn’t be happy with a search in a large database that simply queried for ‘atomic’ and ‘power’, as (at least in my main database) the search results would include hundreds of irrelevant results. Because the literature often varies in the use of terms, the person doing the tagging is likely to be inconsistent in applying the same tag always to similar documents. If in the process of tagging we’ve jumped about among those 4 tags noted above, a search for any one of them, such as “atomic energy” won’t pick up the similar documents that were not tagged that way.
Indeed, tags can be very useful to allow searching of a collection of documents that are in different languages - provided that the tags are applied correctly and consistently.
Frankly, I almost never bother to tag the hundreds of new items that I add to my databases every week, as I enter them. Most of those new items are interesting to me and I do categorize them early on into an appropriate topical database and a group within that database.
Here’s why: to do a good job of tagging a priori, as the item is entered into a database, could take a good deal of time and effort. But I concluded long ago that a priori tagging is a poor investment of my time and energy. Nor (except in an exceptional case) do I want to import tags created by someone else, when I import documents. Why should I trust their tags?
That doesn’t mean that I think tags are not useful and important. When I’m actually working on a research project is the stage when I’ll spend a good deal of effort searching for relevant material and often using See Also to see suggestions for other possibly interesting items. That’s when I use tags, to help pull together resources that are useful in various ways for that project. In the old days, I used to create new groups holding resources for a project and replicate items into those groups. Tagging is less work to accomplish the same objective. Just as I used to delete those project groups containing replicants of resources when the project was finished, I erase the tags created for a similar purpose, when the project is finished. That’s because I found that the resource groups (or tags) created for the purposes of one project are not likely to be useful for the next project, and might even be impediments to approaching my resources from a different perspective.
Tags are important to me for various other uses than in projects. For example, I’ve got a tag set for the writings of a colleague, which also includes a critiques of his work that I plan to respond to, and some of my preliminary notes on a response. As these documents are differently grouped in my database, I created a hierarchical tag group to pull them together for future reference.
I use tags more extensively in my financial database than any other, although that may change with the added capabilities of date range searches (in a forthcoming update of DEVONthink), as smart groups based on date ranges will replace some of my existing uses of tags in that database.
I’ve been lovingly building my main database for years, even before the initial appearance of DEVONthink. I prune it occasionally, removing obsolete or otherwise unsatisfactory references, and use tags to mark items for possible removal.
I usually exclude groups from tagging (File > Database Properties), so I see only tags that I’ve created (or, in one instance, imported). As you can guess, I don’t use a very large number of tags (which helps me apply them consistently).
Disclaimer: If you don’t agree with my perhaps eccentric views about tagging (although the literature of information science doesn’t disagree with me), feel free. Carefully done tagging can distill the information content of documents, and can while away many an hour in hopefully productive ways. 
