Deleting something certainly _can_ improve Wikipedia.
I agree with this, and I think you have to be dishonest to not accept it. I just don't think the benefits of deletion, in cases where it is not legally mandated, outweigh the potential for abuse. I think there's always a better solution, though in some cases this would require changes in the mediawiki software.
I don't think the fatal attraction of deletion is that people enjoy "being a cop." I think the attracton is that VfD has about the right amount of traffic, social interaction, conflict, drama, winning and losing to be addictive. I can write whole articles without getting a single "attaboy," but almost everything I do in VfD gets interesting positive and/or negative responses.
This is probably more the deletionist perspective, though. Personally, I am drawn to VfD more because it requires immediate attention. Writing a stub can always wait. Improving an article on cleanup can always wait. Why put off til tomorrow what you can put off til the day after tomorrow? But when something's listed on VfD, there's a good chance that it is going to be permanently erased (at least from my view) unless something is done about it. That's why I feel so strongly about having access to deleted articles. If I could only access the article after it were deleted, then it wouldn't matter if it were deleted. I could archive it somewhere and worry about it later.
But creating a new article has an instant-gratification factor. How else can you explain the fact that people would rather create eighteen substubs than a single article with eighteen short sections?
Creating eighteen substubs is much much easier than creating a single article with eighteen short sections. In order to do the former, you just search google 18 times. In order to do the latter, you have to find multiple sources on the same topic which contain 18 different facts distributed among the different primary source evenly enough that you aren't infringing the copyright of any single primary source.
And really, that's probably the single biggest reason to keep a poorly written stub. It allows a type of [[clean room design]] which can free information from restrictive copyrights. I'd rather see 18 people add 1 line to each of 18 articles than to see those same 18 people each add 18 lines to 1 article. With the latter, you get more fact-checking, you guess less risk of copyright infringement, and you get a more neutral point of view.
There are certainly arguments that we shouldn't have substubs and poorly written articles. But even these arguments I don't think necessitate deletion. If you've got a substub on a topic which could theoretically grow (but might take a long time to do so), why not move it to the talk page? Talk pages aren't included in random page, and talk pages don't reflect as poorly on Wikipedia. But if someone comes along and decides to improve on the article, the information will be sitting there in the talk page to help that person along.
This is part of what I'm talking about when I say that there's always a better solution. It doesn't make sense in every situation, such as an article which is patent nonsense, but in that case we could always mark the article as deleted (thus taking it out of random page, search engines, and the database dump) but still allow people to view it if they want to check and make sure that it really was patent nonsense.