AI Translation in 2026 Isn't the MTL You Remember
Why readers who swore off machine translation should look again, what fan translators still do better, and how to read untranslated novels fairly.
If you read web novels for any length of time, you probably have an MTL story. A hero who was Jin Wook in chapter one and Gold Feather in chapter nine. A "she" who turns out to be a he three sentences later. A line about "the cold wind of the nine heavens eating the rice" that you read four times before giving up. Plenty of readers tried AI translation once, got that, and wrote the whole thing off. That was a fair verdict on what they saw. It just isn't what's on offer in 2026.
What old MTL got wrong
Old machine translation worked a sentence at a time and forgot each sentence the moment it was done. That explains most of the damage. Korean, Chinese and Japanese all drop subjects constantly, so a sentence like "went to the market and bought it" has no he, she or they in it. A system with no memory has to guess, and it guessed differently every time. Names got the same treatment: a name that happens to be an ordinary word was translated as the word on Monday and spelled as a name on Tuesday.
Idioms were the other casualty. Matching words to words could produce a grammatical English sentence that meant nothing at all, and your job as a reader became decoding it. People who read MTL for years got good at that, and it's a real skill.
What changed
Current large language models read the whole passage before they write a word of it. They can tell who is speaking from three paragraphs back. They can see that a phrase is an idiom and reach for what an English speaker would actually say. The result reads like a novel with a voice, with jokes that land as jokes and dialogue that sounds like people talking.
The practical difference is the one you feel in the first page: you stop decoding and start reading. A story you have to translate in your head is homework, and a story that flows is one you can fall into and come up from three chapters later.
It isn't perfect. A model can still flatten a joke, miss a bit of cultural weight, or smooth over something a human translator would have fought to keep. Honorifics and forms of address are the usual casualties. But the failures now are the failures of a decent translator having an off day, which is a different category from the failures of a phrasebook.
Fan translators did the real work
None of this is an argument against fan translators, and it shouldn't be read as one. Fan translation is skilled, unpaid labor, often done late at night for a community, and it is the reason a lot of readers found their favorite books at all. A good fan translator makes choices a model doesn't: a consistent voice across a million words, a footnote that explains the pun, a decision about how to carry an honorific. Plenty of them beat any automated output.
AI translation isn't a way to publish something under a translator's banner or to compete with a team that spent a year on a book. If a book has a translation you like, read that one and thank the people who made it.
Most books will never be translated
The uncomfortable arithmetic is that the supply of web novels is enormous and the supply of translators is not. Korean, Chinese and Japanese platforms publish new serials every day, many running to hundreds or thousands of chapters. Only a thin slice of that ever gets an English version, whether from fans or from a licensed publisher, and the slice that does tends to be the biggest hits. Plenty of excellent books sit in the long tail with no English edition and no sign of one coming.
For those books, the choice isn't between a fan translation and an AI translation. It's between reading it with AI and not reading it. That's the honest reality for most of the shelf, and it's a good reason to take modern AI translation seriously even if you have decades of MTL scars.
Reading untranslated novels the right way
The way to do this without harming anyone is short.
- Buy the book on its official platform, whether that's Ridibooks, Kakao Page, Qidian, Syosetu or wherever it's published. The author gets paid, and that is what keeps the next chapter coming.
- Translate it for your own reading, the way you'd use a dictionary or ask a friend who speaks the language.
- Keep the translation to yourself. Don't post it, share it, or upload it anywhere, and don't hand it to a group chat. A private translation of a book you paid for is a reading aid. A shared one is someone else's work given away.
That keeps the money with the author and leaves fan translators to the books people care enough about to translate in public.
What a good reading setup needs
If you're going to read a long book this way, a few things matter more than they sound. Names have to stay spelled the same from the first chapter to the two-hundredth, or the book falls apart the way old MTL did. Long chapters should be translated in pieces, with a check that nothing was skipped, because a missing paragraph in the middle of a fight scene is its own kind of betrayal. And you should be able to read where you actually read: a phone on the train, a laptop at night, a tablet, or an e-reader with the chapters exported to EPUB.
That's the setup our app is built around. It doesn't host or share translated chapters. Each reader translates chapters they already have access to, bought or free, for their own reading, and the book builds its own glossary as it goes so names stay put. If you've been burned by MTL before, read it in Crossleaf with a chapter you know well and see whether it still feels like decoding.