“AI will take care of it”: will it, though?
Anyone working in translation and language services knows the limits of artificial intelligence.
Those outside the field, however, often turn to machine translation under the impression that it’s faster and more cost-effective.
Let’s be honest: AI is here, it’s evolving, and we can’t pretend it doesn’t exist. But it’s just a tool. A support, not a substitute.
A concrete example: translating in the railway sector
In our work, we often handle highly technical texts, such as those in the railway industry—a niche that demands absolute precision and a thorough understanding of context.
Post-editing proves just how essential human involvement still is.
AI systems make glaring mistakes, often misinterpret the context, and lack terminological consistency.
One telling example: a document describing testing procedures for a pantograph.
🛠️ A pantograph is a device mounted on trains or trams, consisting of articulated arms, that collects electric current from overhead contact line.
The original text included the terms “knuckle trailing” and “knuckle leading”, referring to the orientation of the pantograph’s joint relative to the direction of travel.
🔍 While knuckle literally means “nocca” in Italian, that translation is clearly unsuitable in this context.
Trailing indicates a rear-facing position, while leading refers to a forward-facing one.
To express these concepts in Italian, a descriptive phrase is required, as there are no direct equivalents.
Here’s what machine translation came up with:
• Knuckle leading → “guida delle nocche”
• Knuckle trailing → “strascico delle nocche”
But the issue isn’t just the inaccurate rendering—it’s the lack of consistency.
Elsewhere in the same document, “trailing” is translated as “snodo di trascinamento,” and “leading” becomes “snodo anteriore,” only to be rendered again as “guida delle nocche” a few lines later!
The pitfalls of word-for-word machine translation
Another recurring issue we encounter is the literal translation of idiomatic expressions or technical phrases.
Take this example from German: “Diese erheben nicht den Anspruch auf Vollständigkeit, […]”
Machine translation output: “These do not raise the right to completeness […]”
In German, however, “den Anspruch nicht erheben” actually means “not to claim”.
👉 A more accurate and natural translation would be: “These are not intended to be exhaustive […]”
Another pitfall occurs when machine translation produces a seemingly smooth output that is, in fact, incorrect.
This false fluency is particularly dangerous: a text may sound right while actually oversimplifying or distorting technical or legal content.
But what if this happens in a contract or an official document?
We explore similar risks in our articles about historical translation errors and mistakes in the medical field: today, the unqualified or non-professional translator might very well be artificial intelligence.
So should we reject AI altogether?
Absolutely not.
As mentioned, artificial intelligence is part of our reality, and anyone working in communication must understand it thoroughly in order to make the most of it.
AI can be useful for very simple texts or for a first general understanding—but even then, critical thinking is essential.
And what about the workload?
No, it doesn’t shrink by half just like that. However, the time it takes can be reallocated: if the first draft is faster, the translator can devote more time to extra-linguistic skills, such as:
• Terminology research
• Legal and regulatory cross-checking
• Cultural adaptation of the text
The real challenge—and our conclusion
The real challenge is helping people outside the industry understand that human input is not “second best.”
On the contrary—it enhances the final quality and ensures coherence, accuracy, and accountability.
We firmly believe that AI only truly works when guided and reviewed by professionals.
Have you ever come across serious mistakes in a machine-generated translation?
Share this article and let us know!
