Everyone knows what an MTL is. Then there was some genius thinking that using deepl is worth mentioning and called his "translations" "dmtl" now we got a new genius who calls his MTLs "AI Translation"
What's the fucking point? It's all an MTL, can we please stick to the term? Did you edit it? "Edited MTL" it is.
Quality wise, they're all more or less the same shit.
They really aren't. There are currently three main categories of translation software available.
An edited AI based translation that starts with a Multipass from a specialized LLM model and an adversarial GAN is damned close to a human hand translation. These haven't landed in this space yet because not too many people know how to run them, the tools people use for rapid mtl's don't support them yet, and they currently need 20gb-30gb of vram. But they almost certainly will eventually. There's someone thats done a couple games this way on the translators hub discord, and they are damned good- but he'll only distribute translation patches, not full games.
Then we have the local nn model stuff like sugoi and deepls cloud based stuff. Inconsistent for sure and REQUIRES some editing to avoid parsing problems- but still blazing fast compared to most alternatives doesn't need special hardware, and generally an intelligible result,especially with a bit of prep work. This is what I think DLSite is calling an 'AI Translation'; that's technically correct since everyone is calling neural networks 'AI' even if I think the correct nomenclature is 'expert systems'.
A step down from that is 'MTL'. These are rules based translation tools, like Fujitsu ATLAS, powertrans, or google translate on the remote end. You can get good results out of these, and they are FAR more consistent than the other solutions-but it takes a lot of work, generally more work than people are willing to put in. Most of what we see on this site are in this lowest tier category.