“Pronto Translations Details Common AI Errors and Champions the Role of Professional Linguists”
New York, April 17, 2024, At Pronto Translations, a premier translation agency in New York, we’ve employed one of the generative AI programs, Chat GPT, for nearly a year to aid in translation. It has taken months to refine a system that functions effectively. Yet, our experience has led to a significant insight: While Chat GPT or any of its competitors can be used to assist in translation, they cannot serve as standalone solutions. There must be a partnership between a highly skilled translator and the program. Let me outline the reasons:
- Mistranslation: This is one of the highest risks. Even though Chat GPT is quite a bit better than most machine translators at discerning the correct meaning based on context, a mistranslation that we observed recently had to do with a car maintenance and repair manual, in a section having to do with nuts and washers: on every few pages, Chat GPT would suddenly misinterpret the “nuts” as the kind that you eat. Similarly, on several occasions, it misconstrued the washer as a household appliance for laundering clothes.
- Invention of Content: Chat GPT will make stuff up. Chat GPT is relatively good with content that is widely available in its corpus, however, when presented with unfamiliar terms, phrases or concepts, it has a tendency to offer translations based on speculative logic rather than factual accuracy.
To highlight a significant safety risk, consider the consequences of depending exclusively on Chat GPT or any AI-powered translator for translating technical jargon in medical documents. Chat GPT might generate a translation for a non-existent medical procedure based on its interpretation of terms.
Additionally, let’s not forget the case that drew substantial media attention where the lawyers for a man suing an airline in what would have otherwise been a routine personal injury case depended on ChatGPT to prepare a filing. The AI bot created a series of fictitious legal precedents, resulting in the judge fining the legal team $5,000. - Overuse of Trending Terms: Chat GPT tends to overuse words that have recently become popular due to the AI’s corpus receiving updates. This frequent use of trendy terms diminishes the diversity and suitability of language choices in translations, making such terms quickly seem overused. For instance, for at least the last six months, we have noticed that, recently, “pivotal” has been utilized excessively when alternatives like “crucial,” “essential,” “important,” or “vital” might have been more fitting. In our experience translating tech articles, we’ve observed that the AI often overuses “innovative” for a range of descriptors, rendering the text repetitive. Alternating with terms like “new,” “advanced,” “cutting-edge,” or “novel” would have mitigated this issue.
- Challenges with Proper Nouns: We notice that Chat GPT and its rivals has particular difficulties with proper nouns, such as company names, people’s names, and job titles, highlighting a significant limitation of all AI-powered translation tools. If there is one thing you don’t want to get wrong, it’s the name of the company that the article is about or the name of one of its top executives. The issue is especially pronounced in translations between languages with different scripts, like Chinese and English, where cultural nuances and the careful selection of a brand name in the target market are vital for upholding the intended image and respectability.
When a Western company enters the Chinese market, it often adopts a Chinese name that either phonetically resembles its original name or selects one with positive connotations or symbolism in Chinese culture. However, AI, lacking specific knowledge or updates about such adaptations, may revert to literal translations based on a straightforward interpretation of the words or their components. These translations can sometimes be inappropriate, convey unintended meanings, or overlook the cultural significance of the adapted name, leading to confusion or even offense among the target audience.
- Misalignment of Register: Chat GPT may fail to match the formal or informal tone intended for the target audience, potentially leading to cultural insensitivity or miscommunication. A finely crafted document in the source language could unwittingly insult an audience by addressing your material to them in the wrong register, an especially sensitive problem in languages with a strict register hierarchy such as Korean or Japanese.
Even languages as closely related to English as French, German and Spanish present this risk, due to the existence of the informal and formal forms for you (tu / vous; du / Sie; tú / usted, respectively).
Translating a casual blog post into German using formal Sie instead of the informal du would make it seem overly formal and disconnected from the audience. - Exaggeration of Style: In attempts to emulate marketing or promotional styles, AI can overdo the effect, rendering the translation as parody rather than persuasive content. Unless you are a native speaker of the language you translating into, you may never realize these fine points that can make the difference between an excellent written translation, and one that will seem comical to the recipient.
In translating slogans, the AI could turn “Just do it” into an overly enthusiastic and lengthy imperative, losing the slogan’s brevity and punch. - Misinterpretation of Contextual and Cultural Nuances: AI can misinterpret or overlook the contextual and cultural nuances of language use, especially in professional or semi-formal communications. A case in point is the Spanish abrazos or the Portuguese abraços, often used to end informal business communications. While a direct translation to “hugs” might carry the correct literal meaning, it fails to capture the appropriate tone or cultural connotation intended in the original language—a warm, yet professional sign-off. An English reader is likely to find the translation to “hugs” more than just strange. The oversight underscores the challenge AI faces in recognizing and appropriately translating language that relies heavily on context and cultural understanding to convey the correct sentiment, professionalism, or level of formality.
- Inability to Improve Clarity: AI fails to refine or clarify source texts that are poorly written or ambiguous, a limitation that can compromise the quality and accuracy of the translation.
If the original text ambiguously refers to “the bank” without context, AI might not clarify whether it’s a riverbank or a financial institution in translation. - Literal Translations: The pursuit of a “perfect” translation often results in outputs that are overly literal, lacking the nuance and flexibility required for authentic communication.
Idioms like “break a leg” might be translated into another language as literally instructing someone to fracture a limb, rather than wishing them good luck. - Misinterpretation of Mood: Chat GPT will have difficulty figuring out the mood of the original text, and may give the reader the wrong oppression about the intent of the material.
A sarcastic remark in the source language might be translated into the target language without the sarcastic tone, changing the perceived intention. - Introduction of Nonexistent Content: Although it doesn’t happen often, Chat GPT can sometimes “hallucinate“, and put whole sections into the translation that we’re not part of the original text. In translating a concise report, AI might insert a detailed but irrelevant explanation based on related content it has learned, which wasn’t mentioned in the source. In our own experience, we were once commissioned to translate an article about one Nigerian artist in Italy, and found the draft translation included more than one paragraph listing and presenting summaries on several Nigerian artists resident in the country.
- Omission of Content in Long Texts: Particularly with lengthy documents, AI may omit significant information, especially towards the end, potentially altering the message’s integrity. The risk is heightened when translating lengthy legal documents, where key clauses or stipulations towards the document’s end might not be translated, altering its legal interpretation.
These issues are not limited to Chat GPT. They are common problems found with all of the recently AI-enhanced machine translators such as Google Translate and DeepL, but also with Chat GPT’s rivals, such as Google’s Gemini and Baidu’s ERNIE.
Conclusion: Chat GPT and its AI counterparts are powerful “machines” capable of producing workable “first drafts” of translations. However, if the user is not fully fluent in both the source and target languages and skilled in translation, relying solely on machine translation or AI-generated translations for any documentation—particularly for advertising, promotional, or marketing materials—can jeopardize your business or reputation.
When you or your organization have important documents to render in other languages, be they legal documents, employee manuals, marketing materials or a crucial news release, eliminate risk to your or your organization’s reputation by involving professional translators in the process.
Do not hesitate to write or give Pronto Translations a call so that we may assist in your translation project.
Media Contact
Joshua B. Cohen,
Chief Executive Officer, Pronto Translations
Tel: +1 646 984 4073
