That's because MSAA is temporally stable whereas DLSS is much less so, so in motion MSAA is going to look much better.
Sorry, can't let it go, this is factually incorrect.
(I hate long posts too, but here comes one, apologies)
MSAA, which works by supersampling pre-designated "edges" in a scene, takes no motion vector data into account whatsoever. It has no temporal component. It is expensive because it samples down specific edges per frame, with no sense of what has happened or what is coming.
The biggest problem with MSAA specifically is that it is
not temporally stable, particularly with deferred rendering elements (although this deferred issue isn't applicable in AC, I believe). Thin edges, like the back of hypercar rear wings, shimmer and pixelate when in motion, even at 4K resolution and high MSAA values. An easy to spot, identifiable example is AMS2's Madness Engine's MSAA - it has a ton of shimmer, even on very high MSAA values at 4K. (TAA finally came to AMS2 in early fall, to mixed results).
TXAA was a semi-evolution of this concept, which added a temporal filtering algorithm to MSAA. TXAA was
literally designed by NVIDIA in the early 2010s as an attempt to fix MSAA's lack of temporal stability. It ultimately failed, due to its high performance cost, its need for a difficult bespoke per-engine (sometimes even per-game) implementation process, and its mixed visual results.
Temporal Anti-Aliasing (TAA) then unfortunately took over, as a console-centric deferred-rendering system in the mid-2010s (generally unified per proprietary engine, i.e. UE4, Unity, Frostbite, etc.) - vastly cheaper performance-wise than MSAA and considerably less "jaggies" in motion, but at the cost of an unacceptable amount of blur and ghosting to the image, due to its high speed and lack of data.
DLSS/DLAA is a temporal solution like TAA - but exponentially accelerated. It takes motion vector data from both past and incoming frames, and extrapolates them through a vast, frequently updated neural network database of game rendering data (apparently called the "
convolutional auto-encoder neural network"), using the tensor cores on an NVIDIA GPU to do so.
DLAA and DLSS are supercharged forms of TAA (and TAAU) - and at proper resolutions and frame rates, completely nullify the many sins of TAA. They are, by their very nature, designed to be temporally stable to an AI-accelerated extreme. And while this certainly was not often the case in its early iterations from ~2018-2021, its current versions (4.0 and 4.5) are mature - and the most temporally stable solutions available by a very wide margin. DLAA in particular is crystal clear at 4K 90+ FPS using Transformer, with essentially zero ghosting, blur or aliasing to speak of. And (in my opinion) 4K DLSS Quality is visually identical, even on a very large 4K OLED - with a very nice performance uplift from native 4K.
The only case in which MSAA might possibly be more temporally stable than DLAA/DLSS is with a very old model (DLSS 1/2/early 3), at a very low base resolution like 720p, at a ~60 FPS frame rate or less. In that case, MSAA would likely be superior, and even then it's a toss-up. Every RTX-compatible card does not fit that profile - so if you have an RTX card, DLAA/DLSS is the best AA solution by far. I've been driving this week with DLAA and DLSSQ and AC has never been clearer, especially when it comes to camera movement and tracking other cars in front.
Of course, there often needs to be minor tailoring of DLAA/DLSS itself per-game - and considering the lengths Ilja has gone to fold proper motion data into ExtraFX, there will likely be a few growing pains. For example, a few fence meshes don't play nice, heat haze is completely incompatible, and there's some wobble when going under billboards. But it is worth the effort to create a temporally stable AA platform - with best-in-class upscaling to boot. Now, I'd wager even a 2060 Super with the latest DLSS3 (CNN) model, with tweaked CSP and AC video settings, can achieve a near-locked, high-fidelity 100 FPS at 1080p DLSS Quality. And that is an excellent upgrade - at zero cost to an (NVIDIA) end user.
Try to target as high a frame rate and resolution as possible - due to the amount of data fed to the DLSS algorithm. More frames + high number of pixels = more data = more temporal accuracy. There is a considerable clarity difference between 4K60 DLSS Quality and 4K 120 DLSS Quality. I personally wouldn't go lower than DLSS Balanced (or maybe Performance) on 4K screens, and DLSS Quality on 1440p or 1080p screens. At lower than those base resolutions, visual artifacts will probably appear - the upscaling algorithm is black magic, but even black magic can only do so much.
Not mentioned for brevity: Presets (use Preset K for 4K, M for lower, pass on L for now), DLSS version differences, base DLSS resolutions, MSAA and its effectiveness on forward vs. deferred rendering, off the top of my head. Tried to make it as readable as possible. lol.
Kool Links:
https://en.wikipedia.org/wiki/Temporal_anti-aliasing - which lead to:
Valve Dev Anti-Aliasing Wiki
ADDENDUM: The problem of this, and perhaps the source of (justifiable) ire towards DLSS, is that it is simply too good at its job. So, naturally, many detestable modern publishers and developers now see it as a crutch to lean on - an avenue to just skip any and all GPU optimization whatsoever. A "Flex-Seal" solution for a critical water line burst - cheap and automated, literally plug and play - perfect for modern-day big-budget slop. And the nerve by executives to extol it as "native" performance when it very much is not.
Randy Pitchford and his coding abomination Borderlands 4, and "HighGuard" AKA Concord II, are the most obvious recent examples of this principle - the "runs horribly, **** it, slap DLSS on it, that'll do" approach. Sadly, this will only become more prevalent as time passes. But as long as it's not relied on for performance, DLSS is the premier anti-aliasing solution as well as an excellent means of adding performance and fidelity to legacy hardware - a great thing in a vacuum, but an opportunity to cheap out for greedy publishers.
None of this has anything to do with DLSS 5 - which is not even remotely DeepLearningSuperSampling at all, bewilderingly. DLSS 5 is in essence an outrageously computationally expensive, completely needless, artistically destructive form of AI filtering, far more akin to something as worthless as Path Tracing than "Super Sampling" or anti-aliasing of any kind.
