Dark pine trees against a mountainous grey background.

Whatever the Wind Brings

LLMs will make translation more expensive in the long run

I've been translating video games for 13 years now and, since the start, people were always looking for ways to replace humans with automatic translation tools. When I started, Google Translate was already widely used and known, and it was heralded for quite a while as "the death of the translation industry". That didn't happen. Then came the Machine Translation, which are basically small databases or neural networks capable of recognizing and reordering words to "make translation faster", also heralded as the death of the translation industry. That didn't happen, but it did create a new type of job: MTPE, or machine translation post-editing, in which a human translator is hired to "edit the little mistakes the machine makes". Unsurprisingly, the machines do a terrible job so, except for a few companies, this tech didn't catch on, but it did make the translation rates worse because it shifted the perspective of how fast a translator can work.

So, now, we have LLMs, also popularly called "AI", a term I dispute and that brings a bad reputation to people researching real possibilities of AI, and that also soils some of previous tech that used the term as an approximation or a controlled simulation of something (remember when we used to blame a game's AI for something? It feels different to say it now). Anyway, it works kinda like Machine Translation, but overpowered and 100x more hungry for power, while also being built over stolen data. And, just like with MT, it's being heralded as the death of the translation industry.

This time they might be right, but for a different reason.

It should be pretty clear by now that every large tech company that's being funded by VC works in a specific way: first they operate at a loss to grow their market share, and then they jack up the prices to start making a profit. We've seen this with streaming services (like Netflix and Spotify) and ride-share apps (like Uber), for example, which destroyed the typical industry that existed and created something that is, overall, worse — sometimes for the consumer, sometimes for the worker/artist. This isn't exclusive to tech, of course: Amazon did the same thing to compete against and collapse traditional bookstores, and Luxottica used a similar tactics against Oakley (to make a hostile takeover). There are tons of stories like that, where these people use money to pressure the world and make it bend to their will.

Anyway.

The hype for the so-called "AI" is the same. Anyone who has the time to read the (very wordy and rambly) Ed Zitron's articles about the industry should know that, by now, these companies operate at a loss of billions of dollars each year while still requiring more investor's money to continue expanding. Some already started changing their pricing, probably in preparation for the price hike sometime in the near future. The fact that these companies are pushing "AI" down our throats against our will — even more if they are useless to us — and without an opt-out button, is a desperate move to try to impose its usage on their users. It's a nice way to fake engagement with a product, at least from what it seems.

While machine learning and language models are useful, they are not adequate for content generation that relies on human interpretation and ingenuity. It is my belief that any translation company adopting "AI" as their core business might face some harsh financial situations in the long-term future.

I mean, why would a company — any company — relinquish control over their own content to third-parties throwing everything inside a "black box"? That's a pretty big financial risk that will require a tighter grip on information they wouldn't be able to really control anymore. Who would really want information about their scheduling, meetings, drafts, and so on, floating inside a software that can, out of nowhere, spill that information out to a competitor by accident? Which leads to another issue: video game localization works under a very strict set of NDAs. How will said company guarantee NDA compliance when they don't control the flow of information anymore?

Like, really, it's insane, to me, to think that an industry who needs tight control over information will simply... put everything into a machine they don't have control over and can never take responsibility for possible leaks that may happen (and I know some funny stories about the extent some companies are willing to go to avoid humans leaking information). It doesn't matter how many guarantees the LLM provider tries to give, there's always a non-zero chance of some confidential text being used as an answer for another user's prompt, even if they never recognize it as a leak. There's not much you can do when everything is a "black box" — although we should call it a black hole, because it absorbs everything and you can never be sure if/when/what will come out on the other end.

It should be noted that, since the LLM is a third-party service in this scenario, the hiring company can't control pricing, access to software, or guarantee development will follow relevant paths to their interest. It would be tying all of its service to another company, from a different field, with different plans on how to keep profitability high.

So the solution to avoid this could be, I imagine, each company maintaining and creating their own LLM. Except that, well, it's very expensive. Not only that, but they would have to create their own set of data to use or do what every "AI" company has done so far: steal it from the internet. I imagine that doing that while also paying engineers hundreds of thousands of dollars a year would be way more expensive than hiring translators. The tech could be optimized, of course, and get cheaper over time — we already seen it happening with the arrival of Deepseek — but it would still require data centers and training data, as far as I know.

Another issue is that, in video game localization, each project is unique. The style, rules, patterns of speech, and so on, can not only be completely different, as well as context and cultural references too — things LLMs are incapable of taking into account due to their own nature. Each project would require a finely-tuned model to produce an acceptable result, and that takes time, money, processing power, and a lot of trial-and-error (because you need to see the results to tune the model). If using a third-party service, an update to the model could break everything; if using their own LLM, the company would have to pay engineers to specialize in each project, so they know how to apply transformers or modifiers or whatever. Not only that, but every language it's being translated into would require this approach.

Meanwhile, you can send to a team of translators just the relevant files, a few documents filled with references, a briefing, and be hands-off until delivery. The result would be faster, more precise, more creative and, in case there's any trouble, someone would take responsibility and fix the issue. And while the mimicking of human writing can get better, the tech will never be able to get consistent results or the finer aspects due to its incapacity to interpret context — something we, humans, do very easily.

For now, using LLMs seems cheaper because the "AI" companies are running at a loss and the services are just passable as good for some markets. When the price hikes starts — just like what happened with every other tech that "disrupted the market" — it is my true belief it will not be cheaper than hiring humans, especially when humans not only are able to provide accountability, but also better and more personalized translations with just a few extra instructions, instead of having to analyze and redo the whole LLM pipeline. And humans will be needed in every case anyway because someone has to fix the mistakes of the machine.

So, what I think it will happen? The same it's already happening: the translation rates have been stagnant since before COVID, with small companies and independent professionals doing their best to push it up to keep with rising costs of living — or, at least, with inflation. Meanwhile, companies pushing "AI"-translated content are pushing the rates down because "it's faster, so you make it up in volume". The thing is, editing machine translation isn't any faster than translating things from the scratch, especially if there are issues being reproduced everywhere that need manual fixing. Just think about how many words you can read per hour. Now think about reading things with your full focus and attention while looking for mistakes. Now think about the time required to edit and adjust words and sentences, and go back and forth in the text to make sure everything is at an acceptable quality. It's impossible to make translation go faster than reading speed.

But it doesn't matter: rates will keep being pushed down by these companies, first to make profits go up, and then to cope with the rising costs of running LLMs or renting their services. If/when the bubble bursts, the damage will have already be done: a whole field of knowledge disrupted, with a technology that now will never truly go away. I know so many people who left the localization industry altogether because they were tired of being mistreated in their own profession, a very specialized field. All the while localization being — in most cases — the cheapest part of video game production/marketing with one of the largest returns of investment: it costs a fraction of a development team's salary to translate a complete game, while opening the product to a whole new market with millions of consumers.

So yeah, companies will be "saving money" now when switching to LLMs, but when the costs rise, it would on par with paying humans translators — if not more expensive. All the while translators are offered fractions of a cent for every word they translated/check/edit. The costs of the whole process stay the same, but the humans earn less than they did the 5, 10, or 15 years before.

Any time you see someone praising the result of a machine-translated text in an official product, how it is as good as something translated by humans, you can bet there was a human working through the night to meet a stupidly short deadline while fixing everything so they would not take the blame for the mistakes the machine made, only because they needed the money to pay the bills.

There are some stories here about it, for those curious, but I've already seen this sort of stuff happening with own two eyes.

#"ai" #llm #localization #rambling