Machine translation can only really work as a first draft that is then reviewed by at least one human who knows the language. You can tell when translations and captions are done by machine only, because they suck!
The ultimate downfall for machine translation when it comes to game loc is that no matter how advanced the rest of the tech gets or how much better it becomes at covering its tracks when it's bullshitting out of ignorance (which DeepL is especially prone to), the context is rarely self-evident within the raw text itself. As someone who's increasingly hired by some agencies to edit more than translate, if somewhat to my chagrin, I'd even go so far as to say that a game's context is never 100 percent self-contained within those files. LLMs inherently fail at this because their majoritarian approach to translation accuracy can never take into account the "local" realities specific to any project; it's a level of hyper-specificity that makes those large data pools a liability.
It's not just a problem of the machines being unable to play the games themselves to be able to discern that context. These are works belonging in genres in a medium that's decades old. They're all but inevitably in conversation with other games and audiences and discourses that you can't pick up just from processing scripts on their own and you still can't fully discern it even if you play the games themselves. You have to take the material in as an organic response to the circumstances of its creation and the work around it, both within this industry and beyond it.
I know this because when I get hired to proofread and edit other translators' work, I'm so often having to drawn upon my fluency in different genres and their histories to be able to confidently fine tune the translations in the way that they need to be. If, for instance, you want to localize a two-on-two arena fighter, how do you tell an AI to be mindful of the Gundam Vs. arcade games and how those games play when translating tutorials? You can't, because that's contextual knowledge specific to that game that's crucial to parsing its text correctly, but is far too specific for an algorithm to ever bias over the reams and reams of other data that says, "Well, when these words are strung together like this, ordinarily it means this 83 percent of the time." And yet that output would be wrong because it failed to discard that noise and hone in on what's actually relevant to that game and informing its text.
At the risk of sounding haughty, this is not a level of knowledge and discretion that you're likely to get from just playing Japanese games casually or even working in the industry as a localizer in a standard capacity. You have to put in the work to research and experience vast sums of history that never got exported in their time to be able to make those calls. And LLMs will never be able to get there with game localization because to get there, they would have to forsake their biggest selling point and if you have to forsake your biggest selling point, then the rest of the model falls apart in execution and the facade is removed, revealing the same fundamental problems underneath that have always made machine translations a liability for these sorts of professional applications to begin with.
Which is why, as I've said before, I don't fear the technology itself actually coming for my job. The technology will never be able to bring that knowledge and decision making and filtering processes to a game, which is ultimately how I'm able to sell myself as a game translator. The real threat, as it's always been, is the cost-cutters who don't truly understand the industry they're in and will throw us and their own clients under the bus to make a quick buck. That's the battle I'm always fighting and the only way I know how to keep fighting is to just go on doing the work the way I always have and let it speak for itself.
