I think it's the result of several factors.
As pcchen pointed out, the hardware is just a vehicle for the artists and designers.
In general, games have very short shelf life before they hit the bargain bin. That means that you want to reach as large an audience as possible at release. Thus, the game has to be playable on hardware that is at least a couple of years old. Of course they might give the user the opportunity to raise the level a bit, but you are typically just adding to the baseline which isn't the same as if you had designed for the higher level to start with. (There are counterarguments to this - if it was better looking/scaled higher, the game might have a longer shelf life. But the bleeding edge will probably never be able to sustain game development. After all, games cost roughly the same regardless of how many they sell to.)
More of the same, take 1. If you give game designers more hardware capabilities, they might want to do what they couldn't pull off on the previous project. Example: Instead of modelling trees with 8 polygons, they model them with 800. There. Much better. And that's it - their polygon budget went to model the trees better. In general, the game is stuck as similar to the last, only modeled a bit better. Another designer was frustrated because he could only have a few trees, hell, he wanted a forest! So next time around he uses 80 polygons per tree but makes a whole bunch of them. Er. Hmm. Turns out he could only affort 50 or so. Still, it's better than 5. So the game gets more trees and slightly better as well. Both these examples use 100 times the polygon complexity. But is it 100 times better? Which brings us to...
Visual impression of quality is not a linear function of the resources put in. The detail you add gets smaller and more insignificant for each step. Say you model a car. You use 1000 polygons. It looks so-so. If you use 2000 polygons, that won't make it twice as good. Likewise, if you use 100.000 polygons, using 1.000.000 won't make it much better visually. I'll try to get the point across - you need a only very little to provide a sketch, LOTS more to provide a good model, LOTS more again to provide the impression of detail and LOTS more to create the impression of "come closer and you'll see more". A car is a car even at 100 polys. And textures supplies lots of the impression of detail. So even the factor of 20-40 improvement that we have seen since the time of Quake1 in terms of processing power and and fillrates can't do all that much. (I'll echo pascal here and recommend that you break out GLQuake and use good textures - you'll be astounded. Or depressed, depending on your perspective.)
More of the same, take 2. Game designers use new resources predictably. Realism isn't polygons alone, polygons are a case of strongly diminishing returns. One thing game developers should be credited for is that they are working with lightning. Another area that should be worked on is model movement - after all, we are not rendering static scenes. Not only should more things move - water, trees, grass - but they should move more realistically. We haven't really come a long way since Tomb Raider (another good point to refer back to, and then compare with new games). Oni brought smooth motion animation blending, but the takeup has been negligeable. To be fair it's difficult to model natural movement. That is, you can't get very far just adding more of the same.... Physics - it is interesting to note that as important as consistent physics is for realism, the improved physics engine in Unreal2003 wasn't part of the original design spec, but the result of an outside group coming along and pressing successfully to sell their work.
Related to 3D rendering, sound has actually regressed since the demise of Aureal, and the subsequent disuse of A3D. Nowadays we don't have any use of reflective positional sound in games, Creatives exclusive-for-Audigy API hasn't seen any use. The state of sound in games is deplorable. And sound is strongly tied to our perception of spaces and surroundings, just as smell is tied to our sense of taste.
So, many of the reasons we don't see much improvement in games graphics are intrinsic, some depend on limited resources/knowledge/time to incorporate better technology, some are related to market realities.
Unfortunately, I can't see anything to change this substantially.
Fortunately, that means that we will still move forward at a similar pace for at least 5-10 years.
Entropy
