The Comics Journal (TCJ) recently published "How 'I Love You' Becomes 'The Moon Is Beautiful, Isn't It?': The Craft of Translation and the Limits of AI" by Gina Gagliano. The article quotes me in four places, as "Peter Anderson, the founder of the AI Manga Translator". I build AI Manga Translator, a manga translator that runs in the browser, and I run it on my own. TCJ used a good part of what I sent. Here are my full answers, including the parts that did not make it in. Thanks to Gina Gagliano for the thoughtful questions and for representing my view fairly.
What TCJ quoted
TCJ quoted me in four places.
On CAT tools: "They are different things that get lumped together. A CAT tool is a workbench for a human translator. AI translation produces the translation."
On how the tool works and why I built it: "I built it for myself first. The series I like tend to be the less popular ones, and those are exactly the ones nobody translates."
On climate: "Translating a page is a small amount of compute, but I have not measured it carefully and I would rather not claim more than I know."
On scanlation: "Scanlation groups have been using machine translation for years, quietly, and cleaning it up by hand. What is changing is the floor."
The rest of this post is everything I sent, unedited.
How it works
From the user's side it is simple. You upload a manga page, or a whole chapter, and a few seconds later you get the same page back with the dialogue in your language. Under the hood it does roughly what a scanlation group does by hand: it reads the text in the bubbles, whatever language the original is in, translates it with the rest of the page and the previous pages in mind so the dialogue holds together, cleans the original text out of the art, and letters the translation back into the bubbles. The result looks like the original page, just readable.
Why I built it
Honestly, I built it for myself first. The series I like tend to be the less popular ones, and those are exactly the ones nobody translates. Scanlation groups go where the readers are, and hiring a translator is not something one reader can afford. So for years it was wait, learn Japanese, or give up. Once language models could handle natural dialogue, I put the rest together around that, and at some point it made sense to open it up to other people. I still run it on my own.
CAT tools versus AI translation
They are different things that get lumped together. A CAT tool is a workbench for a human translator. It splits the text into segments, remembers how you translated a phrase before, keeps your glossary consistent, and manages the files. The translation itself still comes from the person. AI translation produces the translation. If a human is involved, they are editing rather than writing. For manga specifically, CAT tools have never been much use because the text lives inside images, and the hard, tedious part is getting it out and putting it back. That part is where automation has actually changed the workflow.
On cultural context, slang, and meaning that depends on the art
Those translators are right, and I would not argue with them. The tool translates from the text, not from the drawings. If a line only makes sense because of a character's expression, or a visual gag, or something that happened forty pages ago, the model may miss it. It does get the whole page and the preceding pages as context, so it is not translating bubbles in isolation, and it handles ordinary dialogue and register better than people expect. Where it falls down is puns, wordplay, honorifics, sound effects, and anything that a good translator would stop and think about for ten minutes. My position is that this is a tool for readers. For an individual reader it is already good enough. You can follow the story, the characters sound like themselves, and a series you could not read at all becomes one you can sit down with in the evening. That is what it is for.
Scanlations
Scanlation groups have been using machine translation for years, quietly, and cleaning it up by hand. What is changing is the floor. The gap between a quick machine translation and a mediocre scanlation has mostly closed, and an individual reader can now do for themselves what used to require a group. I expect the low-effort speed scan groups to lose their reason to exist, and the good groups, the ones with strong translators and typesetters who care about a series, to keep their audience, because what they offer was never just the translation. On my side, the tool is for a person reading their own files, and each user only sees their own results. And some of the most useful feedback since I opened it up has come from people in scanlation groups themselves, which tells me at least some of them see it as a tool rather than a threat.
On the concern online
I take it seriously, and I do not think it is misplaced. From what I see, most of what goes through the tool is material that was never going to be professionally translated, so readers are getting something rather than nothing, not something instead of a translator. But I understand why the direction of travel worries people who do this for a living, and I would not tell them their concern is wrong. On quality, I have said what I think above. It is good enough for a reader, and I do not claim more than that. On climate, translating a page is a small amount of compute, but I have not measured it carefully and I would rather not claim more than I know. I am building something readers clearly want, and I try to be straightforward about what it is and is not.
What has changed since I wrote these answers
I wrote these a while ago, and a few things I listed as weak spots have moved since then. I have left the answers as I sent them and added the changes here.
Honorifics are now a choice, not a guess. There are three translation styles: a natural localized one, one that keeps honorifics like -san and -senpai, and a faithful one that also keeps original terms and does not condense. If you are not sure, you can run one page in all three side by side for the cost of one page.
Names stay the same across a volume. You can set character names and terms once in a glossary and they apply to every page.
Bulk translation. When I wrote these answers I described uploading "a page, or a whole chapter". It now takes whole volumes: up to 1,000 pages in one upload, or CBZ, CBR, ZIP, RAR, PDF and EPUB files up to 1 GB each. You can drop in several files at once and each one runs as its own job, with pages processed in parallel. Before it starts you can untick covers, credits and blank pages so they don't use up your balance. I wrote up the full workflow in how to translate a whole manga volume at once.
Reading on the site itself. The Chrome and Firefox extension translates pages in place on the reading site, and can do it automatically as you read.
Puns, wordplay and sound effects are still the hardest part. Translating with the whole page and the previous pages in mind has made them better than when I wrote these answers, and the simpler ones now come through. The ones a good translator would stop and think about for ten minutes are still where it falls down.
A shout-out to the scanlation groups who wrote in after I opened the tool up. Your feedback has been some of the most useful I've had, and the tool is better for it.
If you translate or scanlate and think I got something wrong here, I'd like to hear it. You can try the manga translator on your own files first: the first 5 pages are free, no sign-up needed.