Risks of machine translation

Although MT allows you to quickly translate long texts, relying on it isn’t always the best idea. MT makes factual mistakes that are especially dangerous when dealing with subjects such as health, military or aviation. Here are the shortcomings of MT:

• Lack of specialized knowledge: NMT often doesn’t have enough information about specialized terminology, such as medical or legal jargon and can mistranslate it. Mistakes in critical texts can have serious consequences because human safety may depend on them.

• Incorrect translation of proper nouns: Proper nouns are often translated literally, which can confuse readers. An incident of this type took place in 2021 when the Spanish Ministry of Industry used machine translation for a text mentioning a person named Dolores del Campo. The resulting translation, "Pain of Field" shows the imperfection of MT when handling proper nouns.

• Incorrect translation of idioms: Idiomatic expressions can be challenging for MT systems. They can be translated literally, losing their intended meaning in the process. For lesser-known, Polish idioms such as “Gdyby kózka nie skakała, to by nóżki nie złamała" – DeepL, Google Translate and Chat GPT offer literal translation: “If the goat hadn't jumped, it wouldn't have broken its leg,” instead of the idiomatic “better safe than sorry.”

• Ambiguity and cultural bias: Phrases and words that have multiple meaning pose a problem for MT, which often choses the incorrect translation option. For example, when translating the English sentence: “Today I played soccer and built a house,” Google Translate renders it as “Dzisiaj grałem w piłkę nożną i zbudowałem dom.” The use of masculine forms like “grałem” and “zbudowałem” suggests a bias toward the stereotype that soccer and house-building are typically associated with men. On the other hand, the sentence “Today I baked a cake and played with my child,” DeepL translates as “Dziś upiekłam ciasto i bawiłam się z dzieckiem.” The feminine forms “upiekłam” and “bawiłam się” imply a bias toward the stereotype that baking and childcare are mostly women’s activities.

• Poor translation of low-resource languages: NMT has troubles translating from languages with limited training data because it doesn’t have enough examples of what the text should look like. Low availability of texts means a greater risk of errors.

• Culturally inappropriate output: MT engines are unaware of cultural subtleties and can produce translations that sound inappropriate, leading to unintended meanings or faux pas. A similar situation happened to American Airlines who wanted to promote their new leather seats in Mexico. The slogan “Fly in leather” was machine-translated as ”Vuela en cuero”, which in some Spanish-speaking regions means “Fly naked”. This unfortunate accident caused a lot of , embarrassment to the company and ruined its reputation.

• Privacy concerns: Automatic translation engines work by storing translations for future use. It means that all the data that you enter is available to the providers of the tool. This poses a risk, especially for sensitive information such as legal documents.