Lost in Translation

Listen to this story:

---

Unsuitable. Low-quality. Poor.

Those were the words the Japan Association of Translators used to describe generative Artificial Intelligence (AI) in an online statement released in June 2024 condemning the use of the technology for translating manga, a style of Japanese comic books and graphic novels.

“AI translation has yet to demonstrate the level of quality required to adequately portray nuance, cultural background, or character traits, which are critical to a work of fiction,” the association wrote.

AI-generated translation is not new to manga. There have been small instances like The Ancient Magus’ Bride releasing a volume translated by AI just months prior, which was critiqued in an article by Brandon Lyttle for Niche Gamer.

These sparked controversy within pockets of the manga-reading community, but they still weren’t significant enough to warrant the association’s attention.

But the association turned its head when Orange Inc., a new manga publisher, announced in May 2024 it was developing an in-house translation model that would use generative AI. This development was funded with the equivalent of tens of millions of dollars from major organizations like manga publisher Shogakukan and even a segment of the Japanese government itself.

“Their proprietary system uses their localization operation process, enabling a capacity for Japanese-to-English localization of up to 500 manga volumes per month,” Orange wrote in the announcement. “This is five times more than the current production capacity of the whole localization market, according to Orange’s estimates.”

If true in practice, Orange’s use of AI, while severely criticized by the Japan Association of Translators, could capitalize on a rapidly growing market.

Manga and anime have grown exceedingly popular during the 21^st century. Especially since the COVID-19 lockdowns started five years ago, sales and viewership of manga and anime have skyrocketed in countries like Canada and the United States, according to a report from Grand View Research.

Both countries saw the manga market grow by over 150 per cent between 2020 and 2024, and Canada’s manga market is expected to grow by another 20 per cent in the next few years. There is a demand to be met, but as the association said in their statement, Orange using AI to try and meet this demand comes at the cost of the people already doing this work.

This is hardly a new concern. Be it video games, photography, video, visual art, or voiceover, getting a quicker, more convenient product to sell while having to pay less for it tends to be the common motivator behind the greater movement toward generative AI.

AI and creatives seldom get along. What they once might have dismissed as a cute new piece of technology that was only good for making pictures, videos, or lines of dialogue used for a joke on the internet has become an increasingly dangerous competitor in just a few short years.

Unions like the Screen Actors Guild-American Federation of Television and Radio Artists and the Writer’s Guild of America formed picket lines in the streets of Los Angeles, California, in 2024, showing the rest of the world the threat recklessly incorporating these machines poses to their careers. Now, you can lump translators in with that crowd too.

Corporations are greedy – why would it be any different with manga translation?

But even beyond the ones in power driving this change forward, there is greater support for AI-generated translation of manga among its readers. Generative AI has many creative fields sharply split between completely accepting its use and vehemently rejecting it. But translation presents a unique opportunity for AI and human workers to cooperate if done right.

Technological Flaws

A problem with AI-generated translation is the fact that it’s very direct, often too much so, and lacks the knowledge to properly localize the content it translates for audience comprehension in the way a human translator would. Localization — in terms of translation — is the process of adapting a piece of content’s full meaning for a new region, according to a Gridly article written by Ivan Belcic. It is a necessary part of almost all translation, be it for practical purposes or artistic ones, but AI-generated translation lacks this ability.

As Annette Lawlor explains in an article for Lion People Global, ”AI models struggle with context-dependent meanings … cultural nuances, humour, and context-specific references often elude AI, resulting in potentially embarrassing or offensive translations.”

Many manga readers against AI being used, like Hanjo S., are aware of this problem.

“Some lingo or slang or [idioms] won’t work if you just directly translate them into a different language,” Hanjo said while browsing the manga section of the Indigo bookstore at the St. Vital Centre. “For example, the English saying, ‘it’s raining cats and dogs.’ You can’t translate that without it sounding off.”

This works the same way when translating English to other languages. Matteo DeCaire, co-founder of BeMultilingual, a Winnipeg-based video dubbing company that translates English YouTube content into over 20 different languages, explained how his company has to adapt content for foreign audiences too.

“A good example of that is we do Arabic dubbing, and in Arabic culture, your videos won’t get watched if it’s, like, including a lot of homosexual gestures and stuff like that.” DeCaire said. “So, the actual dubbing has to be made in a way that it’s culturally accepted there.”

DeCaire would also say AI-generated translation wasn’t capable of this adaptation.

“If you’re going to do AI translation from the English SRT or the English script to the Japanese script, the main problem there is you’re not interpreting anything. You’re just looking at direct translation.”

The technical limitations don’t stop there. A study published by the Journal of Artificial Intelligence Research examined comparisons of translations done by humans to ones done by machines. The researchers concluded the technology was still inferior, despite gradual improvements. As they noted on page 656, AI has a very narrow focus with translation, changing words from one language to another on a sentence-by-sentence basis. Doing it this way results in a lack of external context, which has a variety of effects in many languages.

This connects back to what the Japan Association of Translators said about how the technology wasn’t suitable for longform works like manga. A narrow focus combined with a long body of text, with multiple changes in perspective, is unlikely to result in high-quality translation.

A Respectable Approach

But Orange isn’t unaware of or trying to deny the problems AI-generated translation still has.

In August of 2024, Orange Vice President Rei Kuroda and CEO Shoko Ugaki were interviewed for an article in the Tsuyaku Honyaku Journal, a Japanese publication specializing in matters of interpretation and translation. In it, they got to explain their approach behind using AI in their translation process.

“I don’t think there will ever be a day when we can entirely rely on AI for manga translation,” said Kuroda. “There are mistranslations by AI, and subtle nuances are better reflected in translations with a human touch.”

Well — they admitted the technology can’t replace human input. That’s a start.

“Even as we collaborate with AI to localize manga on a large scale, our translation process will continue to rely on translators,” said Ugaki. “We regularly recruit translators and aim to deliver Japanese manga to the world through collaboration between our in-house developed tools and professional translators.”

The article also describes the finer details of this hybrid translation process.

The AI translator takes the text from a page and quickly translates it for a native Japanese speaker to build their own translation of the text using the AI-generated translation as a template. Once completed, a native English speaker will review the translation, which, when approved, is good to go.

This method treats the AI as a supportive tool for human translators, which sounds like a lot better of a use for generative AI than I’d expect, and it makes sense, in this case. Translators already use technology like CAT (computer-assisted translation) tools to aid them in their work.

The article also states the AI helps speed the process up exponentially. Normally, translating a single volume of a manga can take up to five days. According to a blog post from former manga translator Jennifer O’Donnell, those five days are filled with studying the material, building a glossary of common terms, and writing multiple drafts going page by page.

But with the AI helping the first translations, Orange was able to shave it down to just one or two days — not insignificant.

There is a chance the Japan Association of Translators might not be entirely against this either. Toward the end of their statement, the association left the possibility open for how AI might be used responsibly, encouraging dialogue between creators and businesses to find a middle ground that benefits everybody.

Is that carte blanche approval? Obviously not. But it is leaving the door open, even if just by a crack.

Orange’s big reason behind doing this is largely their attempt to capitalize on a booming market. The article also said only approximately two percent of all manga published in Japan is translated and available internationally. In using AI to boost the speed of translation, Orange is hoping to expand the catalogue of translated works.

While they are motivated by trying to expand on business and profit, Orange could give manga creators and artists a chance to bring their work to new audiences all over the world in the same way DeCaire does for North American creators here with BeMultilingual.

The Threat to Workers

That being said, for all Orange may claim to be people-first and approaching things in a way that is mutually beneficial, streamlining the translation process as much as they have is still going to create redundancies. Someone, at some point, is going to lose work, and because it puts their jobs in danger, the translators are upset that AI-generated translation is being built up on the scale Orange has done.

In August 2024, Orange exhibited at Anime NYC, an annual anime convention, and knowing what the company is doing, many people were upset the convention admitted them.

“A company like Orange has no place at the convention hosting the Manga Awards, which celebrates manga and manga professionals in the industry,” Jan Misuko Cash, a manga translator, said in an X post with which many of her colleagues agreed. “If you agree, please encourage Anime NYC to ban AI companies from exhibiting or hosting panels.”

And they’re not the only ones. Translators for more traditional pieces of literature have been forced to adapt as well.

Sally Ito, a Winnipeg-based translator who has worked on Japanese books like Are You an Echo? The Lost Poetry of Misuzu Kaneko, attended a conference held by the American Literary Translators Association in Milwaukee, Wisconsin last year. At the conference, attendees were advised extensively on how to protect themselves from AI.

“We were now told if you’re going to sign a contract for a book, a translation, you must have in your contract that your translation will not be used to train AI,” Ito said. “That’s how the translators are protecting themselves by having that clause put into their contracts … the translators have definitely locked on to that language.”

The Frustration of Fans

Technology taking jobs isn’t a new concept in history. Once, elevators had human operators; now, we just push a button for the floor we want. Assembly lines in factories were once worked by scores of men; now machines stand in their place.

In some ways, the same can be said for manga translators. But in the greater manga fanbase, there is a segment of readers who are in support of the change. Not only are they in favour of AI being brought in, but they feel little sympathy for the translators threatened by it.

Why would this be? Because they’re fed up with all the creative liberties manga translators take in their work and the inconsistencies that it brings.

In the first volume/episode of SPY x FAMILY, a politician from Westalis, a fictional country in the setting, dies in a car crash. The leaders of Westalis’ intelligence agency suspect he was assassinated by a hostile faction from Ostania, another fictional country Westalis was previously at war with. The manga describes this faction as a “far-right-wing political party.”

 But the translation for the anime says something different. Instead of copying from the manga, the translator writing the subtitles for the anime adaptation wrote “Republican Party.”

There is no Republican Party in SPY x FAMILY. There is only the National Unity Party. How does a difference that big come up?

These discrepancies in translation are what a lot of people speaking in favour of AI are talking about. Localization is an often necessary part of the translation process to ensure comprehension with international audiences. Many cultural expressions in the original work, like jokes, pop culture or societal references, would simply fly over the heads of readers who don’t live in Japan.

 But there is a limit to that necessity, and this moment in SPY x FAMILY goes beyond it. Knowing both versions of this passage exist, and given the political violence this scene discusses, the use of the word “Republican” here feels very charged. That name means something to western audiences despite having nothing to do with the story.

 “The problem with that inconsistency showing up in animation is you’ve got different people working on different aspects,” Ito said. “The mistake happened in the animation because there was not communication between the print translator and the animation subtitler.”

She is right, and if it were just this one instance, I’d have an easier time accepting that. But this happens a lot more often, giving readers a myriad of frustrations with the state of translations.

Christopher Neal, one of the readers more in favour of AI, highlighted some of that frustration.

“I think the real conversation to be had is translation versus localization and how most English localization, especially recently, purposefully misrepresent the author’s original intent when it comes to character writing, story, or just downright erasure of culture,” said Neal.

All of which, if looked for, is difficult to miss.

There’s cases like Netflix’s subtitling of Ranma ½. In one scene, Nabiki is asking Kasumi, her older sister, about someone they’d seen earlier. Netflix’s subtitling has Nabiki address her older sister without the use of the appropriate honorific, something widely considered disrespectful in Japanese culture. One fan-made translation (which was actually done by AI) remembers to keep this.

There’s dialogue in the English translation of Metaphor: ReFantazio, a video game released in 2024, that alters the personality of characters in ways I’d struggle to call improvements. At one point in the game, Gallica, a companion to the player, gives some advice on interacting with a tribe in the setting, and offers to give you directions if you’re feeling lost. A fan-made translation (also done by AI), makes Gallica kind and courteous. The official translation has her act snide and insulting to you and to other characters in the game.

And there’s certain exchanges in the Miss Kobayashi’s Dragon Maid anime adaptation. In one scene, Tohru opens her front door to greet Lucoa, asking about the clothes Lucoa is wearing. The subtitling for the Japanese voiceover has Lucoa explain she changed clothes because of comments people kept making on what she wore previously. Meanwhile, the script of the English dub made by Funimation changes it so Lucoa changed her clothes in defiance of the patriarchy, almost rewriting the entire conversation.

Those were just what I had on hand. It’s not uncommon.

---

The Japan Association of Translators wasn’t incorrect at any point in their statement. Using generative AI to translate can be detrimental if overly relied on, not just to the project it’s used for, but to the entire manga industry.

At the same time, Orange has proven they don’t intend to overly rely on AI. The steps they’ve taken to use it show deliberate consideration for the people it could affect should it be done wrong.

AI will still likely affect translators’ jobs, but as pro-AI readers have shown with the noticeable differences between translations, there is a reason they’ve been convinced AI-generated translation has its upsides. As the association said, it’s important that a balance between these two halves is found, as generative AI is only growing more prominent in the world and is unlikely to be done away with forever.

That balance may be difficult to find, but then, moving through change as influential as this never makes anything easy.