Game LocAIzation – About AI in Game Localization

Ewa Dacko

If it’s the beginning of 2023, everybody and their uncle is talking AI. A few days ago, I participated in the creation of an article where I, along with my fellow mobile gaming experts from Ten Square Games, shared our views on the use of AI in our industry. You can read the whole piece on our company blog, but I think there’s a lot more to say on AI in game localization. Enjoy!

EDIT, March 15th—Things do change fast. I started writing this article a fortnight ago1, and it is already getting outdated. I’ll do my best to add notes with the latest research and data when something relevant changes.

On a personal note, I feel that the term “AI” is being overused and, probably, overhyped. It seems that anything that involves working with Big Data and/or neural networks (machine learning, natural language processing, image and speech recognition—pick your poison) is now being marketed as “artificial intelligence.” And we’ve all watched The MatrixBlade Runner, and Star Wars, right? So, we half-expect shiny robots or mega-computers that can read our minds and conquer new worlds. But the solutions available today—no matter how smart, valuable, and useful—still have no intelligence of their own.

They simply use the AI toolkit. As Louis Perez-Breva of the MIT School of Engineering explains, “Just as a wrench you use to fix your car isn’t a car, the tools for AI aren’t themselves intelligence.”

Most experts agree that we are still far2 from creating the so-called „general AI”—a system that could complete or learn to complete any task that a human can do—and even further from reaching the state of singularity, where the super-intelligent AIs in our lives either become self-aware or achieve an ability for continuous improvement so powerful that it will evolve beyond our control. But I’m sure humankind will get there eventually, and it will most likely happen in our lifetimes.

<pessimist mode=on>Yes,

I do believe that AI will eliminate us as a species, take over the world, and transform the earth to paperclips as efficiently as possible

(or whatever its primary objective was programmed to be).

No alt text provided for this image
Credits and source:

But let’s leave philosophical musings about the future of humanity aside.

Has AI affected my area of gaming?

My main area of expertise is localization—that is, the process of adapting games to the needs of people from different countries, regions, and cultures, with the goal of providing them with the same great player experience. This not only includes translating our games into different languages, but also involves many other areas such as internationalization (designing our games in a way that allows us to localize them), cultural adaptation, compliance with different laws, localization quality assurance, and managing all the processes and resources involved.

In the linguistic industry, AI-based solutions (for example, neural machine translation [NMT] engines like good-old Google Translate or DeepL) have already been known and used for some time. They still can’t replace humans or deliver the translation quality that gamers expect—just google “game localization failures” for some hilariously tragic cases—but they are here to stay and can help us a lot. How? Let me give you some examples.

How does AI help us localize games?

At Ten Square Games, We localize our games into nearly 30 languages, but no matter how fantastic, well-qualified, and versatile my team is, we don’t speak them all (yet). Publicly available NMT engines help us quickly verify that, say, an Arabic or Thai text string matches its English source; or they can help us understand what Hungarian players are complaining about (and why it’s the lack of support in their native language, yet another localization-related issue). Or check if the translated text fits into the space provided by the UI designers. (Did you know that English source text can expand up to 300% in translation, depending on the target language?)

Writing assistance tools like Grammarly, LanguageTool, or DeepL Write help us create good-quality texts in Polish and other languages by catching grammar and punctuation mistakes, offering suggestions for clarity and more creative phrasing, or even proposing that the tone be changed.

As far as I know, most gamedev studios are still reluctant to translate entire games using NMT. But that is slowly changing, and

localization service providers are outdoing themselves in coming up with fancy names for their services, like AI-Powered or AI-Boosted Game Localization (or perhaps just LocAIzation? I hereby reserve the copyright for this word!).

This generally boils down to something called MTPE (machine translation post-editing), where a text translated by a machine is then reviewed and edited by human linguists.

At TSG, we don’t do that. Our games are not text-heavy or rich in narration, and we work in continuous delivery mode (which translates—pun intended!—to dozens of very short strings about entirely different things to be translated every day). At the end of the day:

  1. maintaining the required level of quality and consistency is not possible without extensive human intervention and, most importantly, without using specially-trained MT engines;
  2. having properly-trained MT engines requires feeding them huge amounts of text;
  3. which we don’t have because see point 1.

But the progress in this area is tremendously fast, and if I had to guess,

before 2030 we’ll see the first video/mobile games localized entirely by a machine (or AI, if you prefer)—and localized brilliantly at that.

We also need to remember that good localization must be based on data: not only game assets (audio data, video data, image data, translations), but also player data or financial data. AI tools can help us analyze player behavior and game usage to make better decisions. Do people from the same country use one language or another? Do they switch languages? Is the churn higher in some regions? Does the language version impact retention and monetization? Are there any countries where the number of players is growing to the point where we should consider localizing the game for them? Are there any noticeable anomalies or trends? If we don’t look at this data, we could be missing out on a lot of great opportunities.

And what about ChatGPT, notorious for its “mansplaining-as-a-service” tendencies? 😉 It can be a great tool, but it’s still a tool and just a tool, not a magic wand that will solve all our problems and make our players happy. (Though sometimes I wish it were just that.)

Still, it does offer many valuable features. We’ve experimented with asking the chat to create descriptions for animals or events in our games, and it’s done surprisingly well. So, maybe in several months, we’ll let the chat create the narrative for our new title?

Have you heard of a mod for Mount and Blade: Bannerlord, a strategy action role-playing game developed by TaleWorlds Entertainment? The mod’s author fed ChatGPT knowledge about the game world and allowed all characters in the game to talk using it. Instead of being presented with only a few dialogue choices, players can type prompts on their keyboard and have free conversations with NPCs about everything from the art of blacksmithing to the intricacies of politics.

Of course, it can’t compete with brilliant dialogues created by a team of skilled writers, but it can be a great way to make the game more immersive and provide a more engaging experience for players – quickly and at little extra cost.

And yes, ChatGPT does offer translation capabilities, but there are much better MT engines available on the market3.

No alt text provided for this image
/imagine: illustration for an article about translating video games with ChatGPT (Midjourney)

What’s even more important, ChatGPT is now very English-centric (no wonder, it’s been trained mostly on texts written in English, with all their biases and presumptions, and as a result, it’s very much embedded in American and Western European cultures). And nearly two-thirds of all gamers are not native English speakers! Just like translation does not equal localization, multilingualism does not equal multiculturalism. So,

the GPT model can (somehow) translate a game into several languages but does not really “know” other cultures—something a good localizer must always take into account!

Such a “monovision” is actually the weakness of all NLP-based tools: they are only as smart as the corpora they have been trained on. That’s why they are burdened by a considerable cultural or gender bias (you can read a sad story about ChatGPT writing performance feedback, or section 6.2 in this very scientific article on training the GPT model).

And that brings us to another question to which there are no simple or straightforward answers: what are the ethical or legal considerations regarding the use of AI when developing games?

Should we regulate it, and to what extent?

For example, should MT engine developers add mechanisms to eliminate linguistic bias? Or maybe assume that the language of machine translators or writing assistants is supposed to reflect reality as closely as possible, that is, the—biased and imperfect—language “spoken” by the Internet? Should such a bias be covered by anti-discrimination laws?

And what about copyright? Who’s the author of an in-game story or dialogue? The AI? The company that trained the model? The prompt operator? The final editor and approver? Who should be credited for such content?

Does the concept of intellectual property, copyright, and royalties withstand the clash with the reality of machine-generated text being analyzed by other machines to enable yet other machines to generate more text?

Does it leave enough room for human creativity, and where are its limits anyway?

And finally, who should be held accountable if something goes wrong?

Mira Murati, CTO of OpenAI (the company behind ChatGPT or DALL-E), believes now is the right moment to start talking about regulating AI, and I agree with her.

Where do I see myself five years from now?

I think I can generalize it to most AI-in-gaming applications: for now, the best course of action is to stick with the “human-in-the-loop” concept. It’s a model that requires human interaction at some stage(s) of the process, be it post-editing machine translation, adapting the content to another culture, adding some finishing touches to images created by Midjourney, or just final quality assurance. For my team today, the “human” part means using the tools we currently have and constantly looking for new ones that could make our jobs easier. Despite all the technological advances and AI development, the core of our work remains human.

As I mentioned above, there are also some big legal and ethical questions that need to be answered. And, as a huge fan of conspiracy theories, I’m not sure that we’ll like the answers to those questions. Who’s going to provide them–regulators, big corporations, scientists, or ordinary people? Will they focus on the welfare and future of humanity, or maybe on the interests of big business?

But let’s get back on track. Our localization team is keeping a close eye on all of this and will definitely be exploring any AI-related possibilities that (will) arise. After all,

it is fascinating to watch the world of game development and game localization change as we speak, and to be part of this revolution.

Some experts predict that “we need to forget about the ‘five years from now’ perspective, as the translation and localization industry will undoubtedly change in the next 6–10 months, not years.”

But will it really? That remains to be seen.

  1. It was originally published on LinkedIn []
  2. March 23rd – Or maybe not so far after all? Microsoft has just announced that they „ignited the sparks of AGI.” And should we take this statement lightly from a company that created their first virtual paperclip back in 1996? []
  3. March 15th – The latest research shows that ChatGPT performs on par with the best stock MT engines, at less than 10% of their cost 😱 []