I'm a software engineer and although I currently don't work on anything related to computer graphics, I understand how downscaling works, what I don't understand is why most modern games default to over 100% on my computer (GTX 1080, 1080p@144Hz monitor). For instance, Overwatch defaults to 140% render scale, which translates into a huge performance hit (stable 144 vs 60~80 fps). Why is the game doing this? I understand that downscaling makes the quality of the image a bit better, but this doesn't justify that performance hit. I read in some places that downscaling also helps to reduce some graphics artifacts, can someone please elaborate on this? I would like to know what are the benefits of downscaling and why modern games are defaulting to settings above 100%.
What you are describing is supersampling, a technique for combating aliasing (a.k.a. jaggies), which is a term from signal theory that also applies to graphics. It's a spatial domain approach. Basically it means that the sampling resolution (render target pixels in this case) makes signals with different frequencies be reconstructed as the same frequency. Or, in other words, the scene is sampled too infrequently for sub-pixel detail to be properly represented.
It is especially prevalent in modern games since the recent advances in GPU computing power have made that quality/performance tradeoff (that you deem unjustified, but unless we're talking VR, everything above 60 FPS is just buttery-smooth and undistinguishable) even possible, and because the (fairly) recently introduced physically-based shading models are much more susceptible to aliasing from sources other than just undersampled geometry: a group of artifacts collectively called shading aliasing, with specular aliasing (a.k.a. fireflies) being especially visible. Epic Games' Brian Karis explains some of that in the opening of his presentation on temporal AA in UE4 (temporal AA is just another approach to anti-aliasing, one that distributes samples in time instead of space).
To sum up: it's the brute force approach to fixing aliasing. An anti-aliasing technique that simply captures more samples of the scene, and thus captures more sub-pixel detail that would otherwise be lost.
It is also worth mentioning that while it appears superficially similar to Multi-Sampled Anti-Aliasing, it differs in the sampling pattern (SSAA simply increases the pixel grid density, whereas MSAA has a sort of sheared grid pattern) and the sampled signal resolution (SSAA is a full colour capture, MSAA cheats slightly by sampling triangle coverage more finely than colour).