It often does not take a long time from the moment a new technological discovery is found for its limitations to similarly be discovered, and this has been true on multiple occasions when it comes to machine translation.
With a survey by the European Council of Literary Translators’ Associations dissuading professional translation services from using AI as a primary form of translation or editing texts generated through such software, the long-standing debate has been inflamed again.
The concept of machine translation predates computers themselves, but the first software to provide machine translation was SYSTRAN.
This became the software that powered the first ever online translation tool, Altavista’s Babel Fish, and created the foundations for both the positives of speedy translation and their immense problems.
The problems with machine translation are ones that cannot be completely fixed even with more advanced technology, algorithms and large language models because they are matters of inherently human expression.
This is most commonly seen in works of literature and poetry, where the dilemma surrounding literal translations has always existed for as long as translations have.
There are many expressions and even some words that simply cannot be translated accurately from one language to another, which leads translators to make difficult choices about how to express an idea in a language that lacks a direct analogue.
Machine translation is somewhat infamous for struggling with wordplay, which means that its accuracy must always remain in doubt when translating figures of speech, poetry and even in some cases legal transcripts where an individual uses idioms.
Artificial intelligence can solve some of these issues with more common expressions but brings in new problems such as hallucination, where misleading or outright false information is presented as correct.
The most infamous case of this was an American lawyer who submitted a case brief generated using ChatGPT citing several legal cases that did not exist.
This was generation, rather than translation, but the problem remains that the software simply cannot be trusted beyond the more basic tasks that machine translation is known to successfully do.