For all the people saying there is no difference so its not an issue, if nvidia did this everyone would go on and on about cheating...
bilinear filtering does 5 texel lookups per pixel, trilinear does 10 and then does a linear interpolation between 2 colors, anisotropic uses the view angle to vary the sampling range of the texels, each anisotropic level enlarges the sampling range(non-uniformly) and so the texels sampled increase, it then applies trilinear filtering with the new ranges. This allows the textures to look smooth even when moving through a scene and as a result heavily anti-aliases textures and reduces pixel swimming but comes at a heavy cost.
For example in a standard 1680x1050 scene, the texel lookups counts are as follows:
bilinear: ~8.82 million texel lookups
trilinear: ~17.64 million texel lookups
Anisotropic: (depending on AF level) potentially more than 4x the texels of tri-linear depending on the AF level.
In a still image there is no problem, the major filtering artifacts are movement based artifacts such moire patterns and pixel swimming and mip-level pops, static images will never really show the full picture when it comes to filtering. I cant believe ATI would do this!
I'd would love someone used the 10.3s and the previous versions and see if there is a difference, that nice 10% increase across the board the new drivers bring might be