AI-Powered Translation vs. Human Post-Editing

There have been advancements in AI-powered translation tools, and the technology has proven capable of localizing simple, well-written texts with enough accuracy. Style is another matter, but not always a concern depending on the type of content. However, as efficient as AI translation can be, it is not exactly intelligent yet. Skynet is still a long way off, thankfully!

This means that AI tools cannot make decisions based on each text’s unique characteristics. Humans, on the other hand, are adaptable. From our experience, all of our customers end up needing to localize content that AI cannot handle on its own.

Here are a few examples:

Character Limits

Sometimes, character limits are unavoidable because of the type of content that is being translated. For example, the titles and descriptions used in copy for ads for search engine optimization, marketing materials, software strings, or subtitles for entertainment media.

Here are a few examples:

Character limit: 60

Source  English text: AI tools are everywhere. Make sure yours are streamlined. [57]

1. Google: Les outils d’IA sont partout. Assurez-vous que les vôtres sont rationalisés. [76]
2. DeepL: Les outils d’IA sont omniprésents. Assurez-vous que les vôtres sont rationalisés. [81]
3. Human: Les outils d’IA sont partout. Rationnalisez les vôtres. [55]

Both automated translations end up being too long. Only a human translator can make the appropriate decision as to what to cut out, or how to rephrase in order to stay within the character limit and remain true to the source meaning.

Character limit: 160

Source English text: This AI guide details why and how to integrate generative AI solutions across all departments of your production chain—so you can innovate with confidence. [155]

1. Google & DeepL: Ce guide sur l’IA explique pourquoi et comment intégrer des solutions d’IA générative dans tous les départements de votre chaîne de production, afin que vous puissiez innover en toute confiance. [194]
2. Human: Innovez en toute confiance grâce à ce guide dédié à l’application de l’IA générative dans tous les départements de votre chaîne de production. [142]

Both AI engines suggested the same output which is, again, too long. Since French is a much more descriptive language than English, the translation syntax has to be rearranged to use less causal link words and therefore comply to the character limits.

Erroneous Source

Mistakes happen. Be it because of tight turnaround times, inadequate tools, or simply human error, source texts can sometimes be imperfect.

Typos

Simple typos can, of course, lead to mistranslations.

Here are a few examples:

Source English text with typo: Prior to instillation, remove all incommutable elements.

• DeepL: Avant l’instillation, retirer tous les éléments incommutables.
• Human-corrected source: Prior to installation, remove all incomputable elements.
• Human translation: Avant de procéder à l’installation, assurez-vous de retirer tout élément incalculable.

These typos are only one letter off and do not trigger spellcheckers, which means they can easily be missed at the writing stage. Automated translation tools will always take them at face value because they cannot question the source text, while human translators usually detect such errors pretty easily.

Factual errors

Source English text: Our solution capitalizes on the unmatched processing power of the RTX 5000 Ada Generation from Nvidia, with its 256 GB/s bandwidth and 12 GB GPU memory.

This data is, of course, wrong in the source, and every single AI engine will simply replicate the mistake while translating, providing false information to the customer, and misrepresenting Nvidia technology. The actual capabilities of this specific model are 576 GB/s bandwidth and 16 GB GPU memory.

Cultural Specificities

Idioms

In all types of content, culturally singular elements can come up depending on the locale. Have you ever heard of the Turkish saying “Elinize Sağlık”? It literally means “Health to your hands” and is used to compliment someone who has used their own hands to make something: a meal, a drawing, a poem, pottery, even a great line of code!

The translation in English and many other languages will depend on precisely what the person is being complimented for. “Compliments to the chef” for a meal, but “Amazing work” for a painting, or “You’re a genius!” for a problem-solving coder. When translations depend on context, humans are much more likely to get it right than machines.

Another example: Let’s imagine you’re an e-learning company based in Germany and are creating in-situation exercises for Human Resources. The reader/learner is playing the role of Hans. The exercise states the following:

Sein Boss hat sich nach dem Problem erkundigt. Hans ist in Erklärungsnot.

AI would translate it as something like “His boss has asked about the problem. Hans is in need of an explanation” which is a mistranslation. Erklärungsnot means that the person is losing an argument and feels the need to turn the tide quickly with their own explanation. Hans is the one who must/wants to do the explaining, not the one in need of an explanation.

Pop References

Your company needs to market something to people born between 1970 and 1990 in the West, and you want to hook them in with a couple of pop culture references. There are good chances they’ve seen Back To The Future, aren’t there?

Can AI instinctively adapt famous movie lines to a specific locale without clear context? Let’s see with a French translation:

Source English text: Great Scott! Going from ideation to finished product has never been so quick! It’s almost like time travel!

• Google: Super Scott! Passer de l’idéation au produit fini n’a jamais été aussi rapide ! C’est presque comme un voyage dans le temps !
• DeepL: Superbe Scott ! Passer de l’idée au produit fini n’a jamais été aussi rapide ! C’est presque un voyage dans le temps !

The proper translation for “Great Scott” would have to be either already known by the human translator or sought out online or in the movie itself, which AI obviously cannot do.

FYI, the French version is “Nom de Zeus !”; the German one, “Grosser Gott!”; the Spanish one, “¡Dios mío!”, and so on. All are complete transcreations, and in all cases, “Scott” has got to go!

Conclusion

The examples explored in this article do not offer a complete list of situations where human input proves essential to the translation process – there are many more. AI technology is a powerful tool that (like all tools) can only result in a satisfying outcome when wielded by the right people.

SimulTrans has AI-powered translation solutions that can be integrated into a translation workflow. Whether you need pure AI output for gisting only, professional Human post-editing, AI translations for customer-facing quality content, or pure human translation, SimulTrans is the ideal localization partner.

If you would like to learn more about machine translation topics, download our free eBook: Machine Translation and Post-Editing: Six Steps to Success.