- Clear indication of whether something was clickable or not, vs. everything flat and featureless
- Good contrast
- Normal-sized widgets which left enough room for content vs. the huge widgets we use today for some reason (it's weird for me to say this, but Linux desktops are the worst offenders here). I get why they're important on touch-enabled systems but that's no reason to use them anywhere else.
- Useable scrollbars, no hamburger menus
- And -- although Windows-specific: the Start menu was something you could actually use.
The state of testing, examination and debates about user interfaces was also light years ahead of what we see today. I was genuinely fascinated about what my colleagues who did UI design were doing, and about the countless models they developed and metrics they used. If it was the same bikeshedding we see today, they sure as hell knew how to make it look like they were having a real debate...
I suspect the reason behind this drop in quality is largely economical. Fifteen years ago, you needed a great deal of understanding about perception, semiotics, about computer graphics, and a remarkable degree of mastery of your tools in order to produce an icon set. This made icons costly to develop, to a point where it was pretty hard to explain it to managers why you need to pay a real designer a heap of money for a real icon because, dude, just look at every other successful app on an OS X desktop!