Because the render engines anti-alias filter is built with a false assumption. Same problem applies to most commercially available vector rendering engines, or in fact many of them are way worse. This is the same reason why the conflation artifacts happen. For different errors caused by the same assumption. Some related posts here, here and here.
Quite simply, we are educating our students with erroneous/imprecisely worded information and nobody bothers to check the state of the art before they start doing things like this. This is not exactly a bug since its working as intended, just the intention is wrong.
The error assumption is that you can replace coverage calculation with alpha blending. This is not true! A simple example, would be a pixel half covered with a shape over a shape that covers the other half. In the real world this pixel would not show through at all but because of alpha blending rules the pixel is now according to rules of alpha blending
alpha = alphaFG+alphaBG(1-alphaFG)
Which results in
0.5+0.5*(1-0.5) = 0.5 + 0.25 = 0.75 while the analytical result is
1.0. Similarly your specific case if the surfeace overlaps but overlaps perfectly with a other shape you get
0.5+0.5*(1-0.5) = 0.5 + 0.25 = 0.75 with 50% coverage while the analytic solution is
So you see basically the mathematics behind the rendering engines anti-aliasation is just simply wrong. If you export that image with art optimized anti-alias it will not have this problem since sampling is not prone to such errors.
Adobe is currently working on fixing this problem after 10-15 years of inactivity so the issue will, probably, be gone some time in the future. In fact you can hardly ever notice it in the newest version. Something i can not say for chrome or firefox.