As
@JustLooking123 and
@Atma said, it varies a lot, and AI can definitely speed up the process.
I recently completed a sort of translation mod for an indie game called
Late Homework, although I never released it because of a couple of Unity-related technical issues that, to be honest, I never cared enough to overcome.
What counts as "
the traditional way" also depends on what you mean by it. Fan translations, official translations, and localizations (which aren't the same thing) all vary greatly in both workflow and quality.
First of all,
CAT (Computer-Assisted Translation) tools aren't new, and neither is machine translation. Long before LLMs, translators used CAT tools with translation memories, terminology databases, glossaries, and dictionaries to keep terminology and phrasing consistent across a project.
Nowadays many CAT tools can integrate AI providers like Anthropic, Google, OpenAI, DeepL, or even services like Linguee through APIs, allowing you to get contextual suggestions based on your own prompts (e.g. "Translate this 1990s RPG while preserving a slightly archaic tone."). They're powerful tools, but they only take you so far if you haven't decided your translation approach.
That's really the bigger issue. Machine translation has existed for years in video games, fansubs, editorial work, and the translation industry in general, but you still have to decide
how you want to translate.
For example,
Sweet Home (the Famicom horror RPG that inspired
Resident Evil) has multiple English translations.
One follows the Japanese script as closely as possible, while a
nother deliberately adapts parts of the script to match the movie more closely and make certain scenes less obtuse for English-speaking players. The latter project even documented the reasoning behind those decisions. Those are examples of different translation philosophies—roughly speaking,
equivalence-oriented translation versus
Skopos theory (and of course, these aren't the only approaches, nor are they mutually exclusive). One prioritizes preserving the source text, while the other prioritizes communicating effectively with a specific audience.
The important point is that no tool—a dictionary, a CAT program, or an AI model—can make those decisions for you.
You can see the same thing in older Nintendo of America localizations. Many games from the '80s and '90s replaced Japanese jokes or references with American pop culture or puns. Some people find those charming, while others think they've aged poorly outside their original cultural context. An AI assistant can help explain those references or suggest alternatives, but if you lack cultural and linguistic knowledge of both the source and target languages, the result can still feel awkward or forced. LLM quality also varies by language; most models are strongest in English and are still noticeably weaker when handling the nuances of many other languages.
That's also where translation and localization diverge. Translation focuses on rendering the text into another language, while localization adapts it to the target audience's culture, expectations, and experiences.
On the technical side, AI can also be genuinely useful. My own project stalled because I never re-enabled the Unity scene needed to switch languages, not because translating the script was the difficult part. I could probably throw together a quick proof of concept today with a mod loader and ChatGPT or Gemini, but handling the technical side still requires some understanding of the game and a willingness to read documentation. These tools won't do all the work for you, but they can point you in the right direction and save a lot of time.
Ultimately, I think it's less about the tool and more about how you organize your project. If your terminology stays consistent, your character voices remain coherent, you have a clear translation philosophy, and you know who your audience is, then the tool comes second. Looking into
localization kits (lockits), terminology management, or even how professional CAT workflows are organized will probably teach you more about modern translation than debating whether AI is "good" or "bad." Even a well-organized spreadsheet can be enough if you're disciplined.
P.S. The recent wave of
Metal Gear Solid (PSX) modding was initially helped by Claude. People used it to assist with reverse engineering and tool development, and now they've reached the point where they're building their own editors and utilities. I think that's a good example of AI acting as an accelerator rather than a replacement.
