They are saying that there is more detail at low ISO with the 50D compared to the 40D, but that higher ISO image quality is roughly on par. All things considered they are saying that the 50D is a bit better, but not enough to justify an upgrade for existing 40D owners.
That sounds about right to me, Bernard. It's in accordance with my own tests. I would agree that image quality alone is not sufficient reason to upgrade. Or to be more precise, image quality directly related to the increased pixel count is not sufficient reason to upgrade from a 40D. But there are at least a couple of sweeteners that Canon have thrown into the mix, namely, a higher resolution LiveView LCD screen which is just amazing at 10x magnification with a 400mm lens attached (on tripod of course and preferably with IS enabled), and an autofocus micro-adjustment feature. Both of these additional features will potentially contribute to better image quality over and above what can be achieved merely with the increase in pixel count.
Edit: As BJL mentioned, Bob Atkins in his review of the 50D demonstrated that, even at ISO 3200, the 50D image lends itself to a degree of noise reduction which can lower noise to the 40D level, and even better, whilst still retaining the resolution edge over the 40D.
I haven't tested this myself and it looks like Chasseur D'Image haven't either, but there's a certain logic to the process. We know that noise reduction programs tend to destroy fine detail. We know also that sacrificing resolution by downsampling can reduce apparent noise, resulting in high-ISO 21mp 1Ds3 images having similarly low noise to a 12mp D3 image, for example, when downsampled.
It looks as though a similar process applies to high-ISO 50D images when some degree of resolution is sacrificed, not by downsampling, but by the application of a noise reduction routine. It looks as though this might be a more successful or more efficient way of equalising noise than downsampling.
This probably deserves more investigation because, if the effect is consistent, it means that the 1Ds3 and 5D2, with appropriate noise reduction, have lower noise than the D3 .
Of course, I'm aware of the counter argument. Any noise reduction that is applied to the 5D2 or 1Ds3 can also be applied to the D3 image to reduce its noise even further, but alas! not without also reducing D3 resolution.
The resolution gap remains and the bottom line might well be, after all noise reduction options have been explored, that a 5D3 image at high ISO can have either lower noise than the D3 but equal resolution, or equal noise but marginally higher resolution.
Something to consider