Why Human Translators Are Still the Best Language Partners for Publishers

I have translated nearly forty books published worldwide. My language pair is French (the “source” language) to English (the “target” language), so I make specific references below to these languages, but the points would apply to all language pairs.

AI does not recognize important errors in the source text. Yes, AI can correct grammatical errors in the target text and help the translator write well-constructed sentences much faster, but grammatical errors in the target text are not the most important potential errors in a translated text because, once translated, the text undergoes normal copyediting and proofreading processes just like any authored text (although good translators meticulously proofread their own work to cut down on that effort).

The critical errors that AI cannot find are factual errors in the source text, and these errors may become part of the target text without a knowledgeable and skilled translator. As a translator, I find factual errors in the source text on a regular basis, and I report these to the English editors, who then report back to the French editors to confirm the facts with the author. This often results in updates to the source text. For example, I recently found an error in a source text regarding the length of a river. AI would simply translate what is written, not question it.

AI does not track the entire text for inconsistencies. As a translator, I carefully choose specific terms or expressions. Once I choose a term or expression in context, I ensure its consistency throughout the entire manuscript. This approach helps convey the manuscript’s message more clearly. In cookbooks, each ingredient should be consistently represented throughout the book, despite having multiple possible translations. AI may translate a term one way in one section of text and another way in a different section, creating inconsistencies.

AI may not understand abbreviations in context. To continue with cookbooks as an example, many chef-authors abbreviate ingredients or measures, and they may abbreviate them differently from other chef-authors or across their own set of recipes within the same book. AI does not pick up on this and can make errors in its assumptions.

Here are two examples of frequent errors I see machine translation tools make due to abbreviations:

The French term for lemon is citron, and for lime is citron vert. In a recipe’s ingredient list, the chef will note citron vert but in the method may choose to abbreviate the word to just citron for ease. AI does not know the chef chose to make this abbreviation, so it translates the ingredient as “lemon” instead of “lime” in the instructions.

“Teaspoon” (cuillère à café) in French might be abbreviated as “cac,” “c.a.c.,” and, on occasion, “cc,” depending on the French author’s style. And these abbreviations may occur differently across recipes within the same manuscript. Although machine translation tools can pick up on the correct translation for “teaspoon” when the entire word is spelled out in the source text, they will not, in context, understand the abbreviations of “cac,” “c.a.c.,” or “cc” and make a mistake. By the way, “cc” could also mean a measure of milliliters.

AI has a limited understanding of idioms, puns, and wordplay. Here is one example among many: The French term chinois is a current cooking term for a particular food strainer. These usually represent the conical strainers that you often see professional chefs use. This term in French literally means “Chinese” because the strainer is shaped like the traditional conical hats common in Chinese culture. I have seen machine translation tools translate this term as “China man,” or even provide instructions of “go to China” rather than “strain” as a step within recipe instructions when they encounter this word.

AI is not a hands-on SME. AI gathers information—both correct and incorrect—from the internet and without practical experience. Does AI recognize when a recipe has a potential error?

I once noticed that a recipe intended for a 3-star Michelin chef’s cookbook that I was translating included a quantity of milk that I believed simply would not work. I determined this from years of experience in the kitchen working with recipes. I tested his recipe (though not asked to), and I was right. I informed the French editors, who informed the French chef, and he confirmed that the recipe did indeed need twice the amount of milk he had noted. My subject matter expertise caught an error that AI could never have recognized.

French chefs, for ease, may abbreviate levure chimique (baking powder) simply as levure, but levure typically means yeast, not baking powder. A skilled translator with the right subject matter expertise will recognize the shorthand used by the chef based on the nature of the recipe. If there is any doubt, the translator queries the editor to query the chef to avoid an error in translation.

Would you want AI to diagnose your illness and assign your treatment based on information it gathers from the internet, or would you rather have a human doctor with years of combined subject matter expertise, proper motivation, intuition, and training in the field do that?

Editors may not have post-translation language expertise. When AI does make an error in the translation, how will editors who decide to rely on AI rather than human language experts and who do not speak the source language know the error was made? Machine translation tools, because their expertise is one-dimensional, make frequent mistakes. Here are two examples where errors creep in:

In French, the gender of a person can easily be mistranslated if there is not enough context around it. As a simple example, the words son bras, could mean “his arm” or “her arm,” but notice the spelling is the same thanks to French grammar rules. The translator must know the correct gender within the context of the text. I have often seen machine translation tools incorrectly identify the gender of a person in these cases and then carry that mistake throughout the entire text.

In French, fruits secs may mean “dried fruits” or “nuts.” In the context of a recipe, either translation may sound correct and be perfectly plausible, but the source recipe has to be studied and understood to know which is correct.

AI cannot handle post-translation text edits. It is common for an editor to send me new or updated text from the French authors after I have completed a translation. This typically occurs when the source-language version of a book is being published simultaneously with the target-language version. As a translator, I must review the French edits and determine whether any changes in the source text will affect my previous translation. Often, the corrections to the source text are purely grammatical, which means they do not impact my translation, and no adjustments are needed on my part. However, most often, the new text does require changes in the target text. In such cases, I must carefully incorporate the new text into my translation while ensuring the message remains intact.

What if new edits to the source text change the context or direction of the text in a broader scope? This means that I might have to revisit various parts of the manuscript and make edits in my translation that were not otherwise anticipated.

Good translation goes beyond linguistic expertise. It involves iterations, discussions with editors, queries to authors, reflection on word choice, and the application of the human experience, but only if the intention is to provide the best possible translation. This can be achieved even within the tight deadlines often encountered in book publishing. To achieve this, skilled translators will embrace AI as a tool for increasing efficiency of work, not as a substitute for language expertise and hands-on experience within their specialties.

Although machine translation tools are intelligent, they cannot match the cumulative intellect of humans, which is shaped by years of study, practical experience, intuition, and instinct—factors that all contribute to a human translator's effectiveness to make the translated text the best it can be, thus making the human translator a publisher’s best choice as a language partner.