Back to blog
Is Human Translation a Thing of the Past? An Experiment: AI vs. Native Speakers for Multilingual Game Marketing
Digital marketers in the game industry already rely on machine translation tools like Google Translate and DeepL for instant translations of call to actions, keywords, user acquisition campaigns, and other marketing collateral. AI has even reached a stage where it can customize voices for different languages, making multilingual video marketing feasible even for smaller indie teams with limited budgets for hiring native-speaking voiceover actors.
As we see innovation transform the global game industry we all must ask:
What are the strategic opportunities and risks of integrating AI in communications with international players?
Is 'good enough' output from AI translation actually good enough even when speed is a priority?
What about when cost is the main concern? And perhaps the most important question: will technology take translators' jobs?
To estimate the quality of machine translation tools and large language models (LLMs) for game marketing, we ran an experiment comparing AI and human translations. We took several marketing texts from a game publisher and had both human translators (we used Nitro service) and AI software handle them. Then, a neutral third-party linguist judged the results and gave their comments.
Spoiler alerts: One - things got interesting!
Two - the outcome isn't as straightforward as you might think.
Level 1. AI for Italian
We started with English and Italian (Italian represents the Romance language family). After translation by a human translator and a large language model (LLM) we had an independent professional linguist do a comparison.
Here's the source text:
In this immersive adventure, navigating through levels and overcoming obstacles feels as easy as pie. With intuitive controls and clever puzzles, mastering each stage is a walk in the park. Don't be fooled by the simplicity; the game's dynamic twists and turns will keep you on your toes. So, roll up your sleeves, dive in, and enjoy the sweet taste of victory in this delightful gaming journey!
In the table below, you can see the translations:
|
Source |
Human translation |
LLM translation |
|---|---|---|
|
In this immersive adventure, navigating through levels and overcoming obstacles feels as easy as pie. With intuitive controls and clever puzzles, mastering each stage is a walk in the park. Don't be fooled by the simplicity; the game's dynamic twists and turns will keep you on your toes. So, roll up your sleeves, dive in, and enjoy the sweet taste of victory in this delightful gaming journey! |
In questa coinvolgente avventura, attraversare i livelli e superare gli ostacoli è facilissimo. Con controlli intuitivi e puzzle intelligenti, padroneggiare ogni livello è un gioco da ragazzi. Non lasciarti ingannare dalla semplicità; i colpi di scena del gioco ti terranno sempre all'erta. Quindi rimboccati le maniche, tuffati e goditi il dolce sapore della vittoria in questo delizioso viaggio di gioco! |
In questa avventura coinvolgente, navigare attraverso i livelli e superare gli ostacoli sembra semplice come bere un bicchier d'acqua. Con controlli intuitivi ed enigmi astuti, padroneggiare ogni fase è come fare una passeggiata al parco. Non lasciarti ingannare dalla semplicità; i colpi di scena dinamici del gioco ti terranno sempre sulle spine. Quindi, arrotola le maniche, tuffati e goditi il dolce gusto della vittoria in questo delizioso viaggio videoludico! |
And here's how a third-party native-speaking Italian translator assessed both outputs:
"In this case, the preferred choice would be the human translation, because the LLM remains closely tied to the source text and occasionally produces direct translations or so-called 'calques'. However, it's interesting that the LLM captures the essence of the Italian idiom 'semplice come bere un bicchier d'acqua' for the English expression 'as easy as pie' — even I wouldn't have thought of that!"
They added that leveraging different AI and machine translation results can greatly inform and assist with human creativity. In other words, a linguist can review multiple AI outputs and determine what (if any) to incorporate into the final text. As a bonus, a linguist would have a valid first draft that, at the very least, reduces the time required to type the initial translation manually.
Following their own advice, the translator proposed integrating the best of both human and LLM outputs to yield this translation of our sample text:
In questa coinvolgente avventura, attraversare i livelli e superare gli ostacoli è semplice come bere un bicchier d'acqua. Con controlli intuitivi ed enigmi astuti, padroneggiare ogni livello è un gioco da ragazzi. Non lasciarti ingannare dalla semplicità; i colpi di scena del gioco ti terranno sempre sulle spine. Quindi rimboccati le maniche, tuffati e goditi il dolce sapore della vittoria in questo delizioso viaggio videoludico!
In the localization industry, the task that third-party translators perform to refine the source text is known as proofreading and editing. It's a creative process that improves the quality of the phrasing and makes the text more coherent and easier to read. Editors assess the style and structure while also correcting grammar, syntax, and punctuation errors.
Editing can be done on both human and machine translations. While human translations typically require less refinement and can be completed more quickly, thus saving on hourly rates, machine translations might need more extensive editing. It's up to you to compare rates for both services and determine whether you prefer two independent translators for your text or if a machine-human combination better suits your needs.
Level 2. AI for Ukrainian
To take our experiment to the next level we decided to take an advertisement slogan and translate it into Ukrainian (which represents the Slavic language family). Unlike the English-Italian example with idioms, this text was more straightforward. However, this task was more challenging due to the differences between language families.
|
Source |
Human translation |
LLM translation |
|---|---|---|
|
Keep on running, keep on gunning, keep on jumping, keep on rolling. |
Продовжуйте бігати, стріляти, стрибати й котитися. |
Продовжуй бігти, продовжуй стріляти, продовжуй стрибати, продовжуй котитися. |
And you know what's interesting? In this short passage, we've pinpointed two subtle differences between the AI's version and how a human might phrase it.
-
Repetition: Necessary in the target language?
The original text uses repetition of the phrase "keep on" at the beginning of each clause. This repetition creates a rhythmic and emphatic effect, reinforcing the idea of continuous action or persistence:
Keep on running, keep on gunning, keep on jumping, keep on rolling.
The LLM opted to maintain the repetition of the phrase "keep on" ("продовжуй") at the beginning of each clause, which copies the original stylistic device:
Продовжуй бігти, продовжуй стріляти, продовжуй стрибати, продовжуй котитися.
On the other hand, the native-speaking Ukrainian linguist chose to omit the repetition of "продовжуй" and instead used the words "running" ("бігати"), "gunning" ("стріляти"), "jumping" ("стрибати"), and "rolling" ("котитися") as subsequent homogeneous members of a sentence, connected at the end by the newly added conjunction "and" ("й"):
Продовжуйте бігати, стріляти, стрибати й котитися.
Translation vs. Localization
Linguistic technology works similarly to tracing paper: it copies the original structure, rhetorical devices, and verb forms of the source language. On the other hand, a native speaker can intuitively dig a bit deeper to uncover the core message and match it with an accurate stylistic expression in the new language. That's the subtle difference between translation and localization.
-
Singular vs Plural Imperative: Who are we addressing?
Some languages, such as English, don't have verb forms that change based on whether we're talking about a singular or plural subject: I run, they run. On the other hand, certain languages, like Ukrainian, distinguish within the verb when addressing an individual or a group of people.
In this specific context, the phrase "keep on running" could be interpreted as speaking to either an individual or a group of people.
The LLM automatically chose the default singular form "продовжуй" as the translation of the original English phrase "keep on":
Продовжуй бігти, продовжуй стріляти, продовжуй стрибати, продовжуй котитися.
However, the human translator decided to use the plural form "продовжуйте" to address the potential collective audience of players, recognizing that the advertisement likely targets a group rather than an individual:
Продовжуйте бігати, стріляти, стрибати й котитися.
Without further context, there's no definite answer as to which option would work better in an advertisement since either could be suitable. In cases like this, it's important to clarify the context to determine the desired tone of voice and the expected level of formality of the text being translated. When faced with multiple choices, understanding the purpose of the text helps the translator select the more appropriate option.
Statistical Probability vs. Situational Judgment
When there are multiple options for translation and the original language doesn't provide clear gender or verb forms, the LLM usually picks the most frequently occurring choice. A human translator, however, doesn't rely solely on statistics. Instead, they use their understanding of the specific context to land on the correct choice.
Level 3. AI for Arabic
In our third experiment we chose Arabic language as representative of the Afro-Asiatic language family, which is read from right to left, as the target language, and translated it in 3 ways: human, combined human and ChatGPT translation, and lastly the machine translation engine DeepL.
|
Source |
Human translation |
LLM translation (ChatGPT) |
Machine translation (DeepL) |
|---|---|---|---|
|
In this immersive adventure, navigating through levels and overcoming obstacles feels as easy as pie. With intuitive controls and clever puzzles, mastering each stage is a walk in the park. Don't be fooled by the simplicity; the game's dynamic twists and turns will keep you on your toes. So, roll up your sleeves, dive in, and enjoy the sweet taste of victory in this delightful gaming journey! |
إن تخطي المستويات واجتياز العقبات يبدو سهلًا للغاية في هذه المغامرة التفاعلية الشيقة. فبفضل عناصر التحكم سهلة الاستخدام والألغاز الذكية، يصبح إتقان كل مرحلة مهمةً سهلة. لا تنخدع بالبساطة؛ فالتحوّلات والمنعطفات الديناميكية في اللعبة ستُبقيك متأهبًا وفي قمة التركيز. لذا، شمّر عن ساعديك، وانغمس في اللعب، واستمتع بلذة النصر في رحلة اللعب الممتعة هذه! |
في هذه المغامرة الغامرة، يبدو التنقل عبر المستويات وتجاوز العوائق سهلاً تمامًا كالكعكة. مع وجود controles حدسيّة وألغاز ذكية، فإن إتقان كل مرحلة أمر سهل جدًا. لا تُخدع بالبساطة؛ فالتقلبات الديناميكية في اللعبة ستبقيك في حالة تأهب دائمة. لذا، اثبت أكمامك، غوص في اللعبة، واستمتع بمذاق الانتصار الحلو في هذه الرحلة الممتعة للألعاب! |
في هذه المغامرة الغامرة، يبدو التنقل عبر المستويات والتغلب على العقبات سهلاً للغاية. مع عناصر التحكم البديهية والألغاز الذكية، فإن إتقان كل مرحلة هو نزهة في الحديقة. لا تنخدع بالبساطة؛ فالتقلبات الديناميكية للعبة ستبقيك على أهبة الاستعداد. لذا، شمر عن ساعديك، وانغمس في اللعبة، واستمتع بطعم النصر الحلو في رحلة الألعاب المبهجة هذه! |
Here's what the evaluator, a third native Arabic speaker, says about the texts:
The human translation follows a sentence structure that sounds native to Arabic, while the other two machine translations follow the structure of the original English source.
Additionally, the human translation avoids the awkward concatenation introduced by the machine translations in "المغامرة الغامرة".
One machine translation translates "intuitive" literally, and ChatGPT leaves "controls" untranslated, introducing a typo in the English spelling as "controles".
The machine translation tools translated idioms like "a walk in the park", "roll up your sleeves", "dive in", or "gaming journey", either literally or incorrectly, while the human translator provided matching and grammatically correct translations.
So, the human comes out on top here.
Because English and Arabic are fundamentally different languages, their translation presents unique challenges:
Technology struggles with idiomatic expressions in languages like Arabic; they do better when the idioms match the target language's model precisely such as with Italian.
AI translation tools lack shared collective human knowledge and awareness of cultural nuances. When translations require creative translation, transcreation of quotes, or adaptation of references that are widely known in the source language's culture a machine will struggle.
Puns and linguistic jokes would be much harder to translate and would certainly call for a human's creative decisions.
Conclusion
In 2024, AI translation tools have advanced to the point of generating fluent, well-written translations, especially when translating between languages within the same or similar language groups. These time- and cost-effective tools can be extremely useful for non-creative tasks, where translation accuracy isn't critical. Since some portion of translation in the gaming world falls into this category, AI translations can be viable. Just beware that AI tools usually pick the most frequently occuring choice.
So, while AI is proficient at translating simple non-creative phrases, humans employ the full spectrum of language to translate and adapt content, especially when the text needs to to persuade someone. Marketing content thrives on stylistic expressions, catchy phrases, and wordplay to break through the noise, ad fatigue, and short attention spans of today's average internet users.
Even when we're unaware of it, we instinctively utilize a language's various stylistic tools to convey subtle nuances of meaning. That's the beauty of language. To translate this beauty into another cultural context, we need a human’s native fluency, genuine grasp of the source text, cultural understanding, and ability to identify the best translation to match original linguistic nuance, specific phrasing, and idioms.
The bottom line: AI translation can be used across content types, but when nuance, culture, and emotions are at play, you need to keep humans at the center.
Trust Nitro your translations!
Nitro® is provided by Alconost, a leading provider of translation and localization services worldwide.
