What if, instead of taking the direct route from one language to another, we applied the principles taught in translation schools to machine translation?
Rather than translating words verbatim, we were trained to understand the meaning of a sentence, then reformulate that meaning — not the sentence itself — in the target language, taking into considerations the knowledge gap between writer and target readers. This approach ensures the message is clear, relevant, and accessible, even for audiences who may not share the same cultural or conceptual background as the source.
So, why not apply this same logic to Large Language Models (LLMs) tasked with translation or interpreting? Instead of simply asking them to translate a sentence from language A into language B, we could prompt them to follow the reasoning process, we can broadly define it of abstraction, taught in school — leading to more accurate, context-aware translations that truly resonate with the target audience.
Let’s take a very complex example because of the many cultural references and use of figurative language, typical of Italian politicking. In this example, a famous politician, Giuseppe Conte (M5S) speaks at the conventions of the Fratelli d’Italia party and says (14 Dic. 2025, article here):
“M5S mai cespuglio del Pd, no alla sinistra della Ztl”
Some translations done by a conventional MT systems or LLM without any particular approach are:
“M5S never bush the Pd, no to the left of the Ztl” (DeepL)
“M5S never a bush of the Democratic Party, no to the left of the ZTL” (Google)
“M5S will never be a sidekick of the PD, no to the ZTL left.” (ChatGPT, simple query “translate…”)
The first two translations are very literal, with Google correctly explaining the abbreviation “PD” but not “M5S.” The third translation, generated by an LLM with a simple translation prompt, begins to be more comprehensible. However, it is fair to say that they make little to no sense. For the non Italian speaker, let’s be crystal clear: the original sentence is very enigmatic, not at all easy to understand even for the average Italian. It is a typical way politicians speak in Italy (see this article, in Italian), which is convoluted, rich an rethorical figures, and surely a big challenge to translate.
When using a Large Language Model with so called chain-of-thoughts (a series of intermediate reasoning steps), however, the result is completely different:
“The M5S will never be a mere appendage of the Democratic Party, nor will it align with an elitist, city-center left disconnected from the needs of ordinary people.”
The translation becomes verbose but effectively conveys the original meaning to its target audience through explicitation. And if asked to be a little more concise, we obtain a translation which is both precise and short:
“M5S will never be a sidekick to the PD or align with the elitist, city-center left.”
To me, the results are surprising. If you consider that the prompting can be adjusted in infinite ways to account for different translation goals (based on communicative situation, different background between writer/speaker and reader/listener, etc.), it becomes easy to understand the potential of this approach.
In this example, as introduced below, what I did was simply to split down the translation process in several steps. First, I asked to explain the meaning of the sentence. Then I asked to re-formulate the meaning expressed in that explanation into a different language, using few words. And so the translation came out. Following these steps, the process of translation is not done by direct conversion of sentence A into sentence B, but by means of abstraction (the explanation). To be noted is that the process of explanation is done by grounding it into the actual political debate that the model is able to reconstruct, possibly by having access to live information from the Internet. In other words, this process of abstraction can count on infinite and current information, something wich is key for making sense of linguistic acts.
There are several lessons that can be drawn from this example. Firstly, it clearly demonstrates the ability of LLMs to go beyond mere direct translation, instead mimicking the process humans follow when performing complex translations: abstraction. Secondly, it also implies that AI translation does not necessarily require training on translated materials, as is the case with NMT. Finally, given that LLMs do not (at least for now) exhibit any signs of intelligence resembling human cognition, this suggests that such a level of sophistication in translation is achievable even without intelligence. As philosopher Luciano Floridi teaches, this is a noteworthy insight. In fact, it raises a lot of questions about what is needed to be able to perform adequate translations, raising even questions to the very nature of translation itself.
What remains open is determining when to apply one strategy over the other — opting for a direct translation or choosing an adaptation to enhance comprehensibility. This likely ties into the concept of Agentic AI, which refers to AI systems capable of making informed decisions. I have anticipated some straightforward applications of agency in interpreting in this blog post.
How I prompted the LLM
Il contesto è la politica italiana recente. Cosa significa la dichiarazione del politico Conte ad Atreju: “M5S mai cespuglio del Pd, no alla sinistra della Ztl”
La dichiarazione di Giuseppe Conte ad Atreju, “M5S mai cespuglio del Pd, no alla sinistra della Ztl”, sottolinea la volontà del Movimento 5 Stelle (M5S) di mantenere una propria identità politica autonoma, evitando di diventare una forza subordinata al Partito Democratico (Pd) o di allinearsi con una sinistra percepita come elitista.
Il termine “cespuglio del Pd” indica la riluttanza del M5S a fungere da semplice appendice del Pd, ribadendo l’intenzione di non essere inglobato o oscurato da esso. Questa posizione è stata espressa da Conte in diverse occasioni, come riportato da Fanpage: “Il M5s ha una vocazione e un Dna a non recitare la parte di un cespuglio, ma di una forza politica che ha una sua identità ben precisa” .
L’espressione “sinistra della Ztl” fa riferimento a una sinistra associata alle Zone a Traffico Limitato, simbolo di aree urbane centrali e benestanti. Con questa frase, Conte critica una sinistra percepita come distante dai problemi delle periferie e delle classi popolari, sottolineando che il M5S intende rappresentare un’alternativa più vicina alle esigenze dei cittadini comuni.
In sintesi, Conte afferma che il M5S non intende essere subordinato al Pd né associarsi a una sinistra considerata elitista, ma vuole mantenere una propria identità politica autonoma e vicina alle istanze popolari.
Come esprimere questo concetto in inglese in una sola frase?
“The M5S will never be a mere appendage of the Democratic Party, nor will it align with an elitist, city-center left disconnected from the needs of ordinary people.”
Make the translation a bit shorter
“M5S will never be a sidekick to the PD or align with the elitist, city-center left.”