It made us wonder: how capable is AI now when it comes to multilingual game marketing? What's the current level of the technology's contextual and linguistic awareness? How does it work with different languages? And, perhaps most importantly, could ChatGPT potentially replace our translators?

That’s why we ran an experiment with game marketing texts that were translated both by AI translation tools and by humans, and a neutral third-party linguist judged the results and gave their comments.

After completing this experiment, and talking with translators who specialize in games and fellow colleagues from the marketing industry, we found 6 common problems they all face when they use artificial intelligence for multilingual content.

In many games, characters have their own unique speech patterns. Take, for example, astrophysics professor Dr. John (game "Kerbal Space Program"), who consistently speaks in a formal tone, uses scientific jargon, and maintains a scholarly manner. Then, there's Albert, a teenager who tends to use nerdy yet colloquial language. It would seem odd for Albert to suddenly adopt overly formal speech, just as it would be out of character for Dr. John to use informal or slang expressions.

On top of that, there are cultures and languages, like Japanese, that put great emphasis on the context and relationship between characters, including their ages, social status, and other factors. For example, dialog tailored for interactions with a young female Japanese character may not be suitable for a middle-aged male Japanese character in the same situation. Using AI translation for such materials will likely lead to the need for retranslations.

The decision to seek affordable translations, even with some compromise to stylistic consistency, should be based upon the assessment of what will satisfy your end consumers, i.e., the players of the game. For example, if occasional shifts in tone or awkward translations don't bother players significantly, it may not greatly impact their gaming experience, in which case, using AI to translate is ok. Maybe, as AI improves its ability to capture the unique voices of characters, this issue may be resolved. In the meantime, consult your players for feedback on translations. You can find some really great insights!

AI systems can vary in their performance across different dialects. They may overlook or misrepresent the richness and diversity of languages, dialects, and accents that are present in the real world. Sometimes, if you don't specify the region, AI systems may even default to the wrong one.

To start, you should prompt for the tool to use the appropriate dialect of the language in question. But, having a basic understanding of the language can help. You'll spot mismatches and dialects. However, if you're not familiar with the language, recognizing the incorrect output would be challenging since the differences between regional varieties can be very subtle.

Spanish, Chinese, French, and even English are top world languages with regional dialects. While you may still be understood in these languages even with the wrong dialect, the question is whether or not it will sound native to the end user.

When discussing challenges in translating marketing materials using AI, one of our translators pointed out that AI-generated translations may seem accurate at first glance but often lack emotion, context, or an understanding of the game's overall ‘vibe’. This disparity has an even bigger impact on marketing, where the end goal is to captivate potential users within the game's world. To do so, translations need to be lively and emotionally appealing so people who see the ad can feel like they're already experiencing the game universe.

When the original message doesn't come across in translation because of some cultural or linguistic nuance, translators can act proactively to introduce creative alternatives, minor adjustments, or transcreation. Because machine translation engines don't yet have the ability to proactively transcreate, it's important to have somebody review the translation of emotionally charged content (like marketing) to see if it matches the tone of voice of the intended audience.

Languages that use honorific speech (e.g., Japanese, Ukrainian, Portuguese, Persian, Russian) often have specific linguistic forms or vocabulary to convey respect, politeness, and formality. When translating into such languages from languages like English, which typically lack explicit honorific systems, AI has a hard time correctly determining the tone of voice that's needed unless there are explicit instructions. 

This one can be even more dangerous than a clumsy word-for-word translation. Since translation engines have evolved to the point where they can produce fluently written texts that read smoothly, it's easier for mistranslations to slip through undetected during proofreading. In some rare cases, automatically translated text might even convey the opposite meaning of the original. So, when a human editor is doing their review, they should make sure to first verify that the machine-generated output aligns with the intended meaning of the original content.

AI tools may not always adhere to character restrictions, leading to truncated or incomplete translations.

This brings us, once again, to seek the trained eye of a human editor to generate or edit a translation that both fits the strict character limits and maintains the intended message and tone

Debates may persist endlessly about the potential for AI to completely supplant humans. Yet, presently, one fact remains evident: for effectively conveying cultural nuances, emotions, meeting specific translation criteria, and localizing into less widely spoken languages, prioritizing human involvement is essential.