Translating and comparing texts

📋 30-second summary

• Never ask “translate this”: always ask for three versions (literal, idiomatic, free). The gap between the three is the real learning.
• For historical texts, ask for glosses alongside the translation. Verify one gloss out of two on Treccani or the Oxford Classical Dictionary.
• To study a literary text, compare two existing translations with AI’s help: map the differences, then go and read the passages.
• For modern languages, train register: same sentence in formal and informal, and declare which regional variant you’re working in.
• In a translation exam or assignment, AI is plagiarism. Always.

Translation is one of those things where AI “looks perfect” and is instead sneaky. A wrong translation of a line of Catullus, or of an article from Le Monde, isn’t visible at a glance: you only see it when you compare it with other translations, with the original text, or with a dictionary. The fluency of the answer tells you nothing about its fidelity to the text.

This lesson shows how to use AI as a help for translation, not as a substitute. It applies to people studying modern languages (English, French, Spanish, German), to people doing classics (Latin, Greek), to comparative literature, to philology Rigorous study of ancient texts: reconstructing the original from copies, dating them, attributing them, interpreting their historical lexicon. It's the discipline that says 'this line is by Catullus, and here's the textual problem'. . The mechanics are the same: translation is a learning tool, not a finished object you paste into the assignment.

The three-versions pattern

Never ask “translate this sentence for me”. Always ask for three versions with different angles. For a sentence of French prose, a line of verse, a long period:

• Literal: as close as possible to the original text, even at the cost of sounding unnatural in English. It keeps word order, lexical choices, syntactic constructions.
• Idiomatic: smooth English, small syntactic liberties to avoid betraying the meaning. It’s the translation you’d find in a good current edition.
• Free: the idea comes through, but the phrasing changes. More adaptation than translation, useful for seeing where you can “break” the text without losing it.

The difference between the three shows you what you “lose” in translation, and that’s the real learning. People who translate seriously don’t pick the best one: they recompose a version of their own by reading the three. People studying a language use the three versions to surface grammatical structures that a single translation hides (a French participle that becomes a subordinate clause in idiomatic English, a Latin ablative that unfolds into a whole sentence).

The base prompt, to adapt:

“Translate this sentence in three versions: a literal one, an idiomatic one, a free one. For each one, explain in one line the main choice you made.”

The line of explanation matters as much as the translations: it’s where AI shows you its move. If the three versions come back very similar to one another, the text is “transparent” between the two languages (low syntactic density, close vocabulary): not a problem, it’s information. If instead you suspect laziness, ask explicitly “push the differences, even take them to extremes”.

Glosses and historical notes

For texts that need historical understanding (Latin, Greek, Old Italian, Old French, Old English) a modern translation isn’t enough. A word may have no equivalent, or have one but with a different historical meaning. Let in Shakespeare doesn’t mean “permit”, it means “hinder”: “What lets but one may enter?” in Hamlet trips up a modern reader who only knows the current sense. Patroni in a Roman context isn’t the “boss” of modern vocabulary.

The move is to ask AI to pair the translation with glosses: words or concepts a modern reader doesn’t recognize, explained in two lines. Like:

“Translate this line of Catullus for me. For the terms patroni and mos maiorum, add a two-line gloss on the historical value in the Roman context of the first century BC.”

Important caution. Historical glosses are one of the points where AI invents most readily: wrong dates, attributions to authors who never existed, garbled etymologies. It’s not malice, it’s the structural limit of a model that has read a lot but doesn’t query a database while it responds. Verify at least one gloss out of two with a serious encyclopedic source, and prioritize the riskiest: proper names, dates, work titles, etymologies. Generic definitions of well-known concepts you can verify last. Treccani (the major Italian encyclopedia, free online) for Italian and historical concepts, Oxford Classical Dictionary for the classics (your library almost certainly has access).

The check costs two minutes per gloss, and it saves you from making confident claims in an essay about things AI made up.

Without paid encyclopedia access

If your university doesn’t have access to the Oxford Classical Dictionary or other commercial encyclopedias, free alternatives are there. Perseus Digital Library (perseus.tufts.edu) for Greek and Latin: original texts, dictionaries, clickable lemmas. Stanford Encyclopedia of Philosophy (plato.stanford.edu) for ancient and modern philosophy, articles written by specialists. English-language Wikipedia is often denser than the Italian one for the classics. For classical Latin, Castiglioni-Mariotti (Italian Latin dictionary); for Greek, Rocci (Italian Greek dictionary) or LSJ (Liddell-Scott-Jones, on Perseus).

Comparing editions

For a student of comparative literature or classics, reading two translations of the same text often says more than the original text. A translator’s choices (what gets rendered, what gets sacrificed, where they modernize, where they let the ancient grate) are a map of their critical angle.

AI speeds up the mapping. Pattern: give AI two translations of the same passage and ask it to map the differences. Real and verifiable example: the Iliad translated by Richmond Lattimore in 1951 and by Robert Fagles in 1990. The prompt:

“I’m pasting two English translations of the first ten lines of the Iliad: the first is Richmond Lattimore (1951), the second is Robert Fagles (1990). What are the three most important differences, and what do they reveal about each translator’s angle?”

AI replies with a structured comparison: Lattimore’s archaizing fidelity and metrical strictness vs Fagles’s modern flow and performance-oriented rhythm, specific lexical choices (Homeric epithets rendered differently), the line’s pacing. Always paste in the passages you want to compare, ten to thirty lines per side. For very well-known works (Iliad, Divine Comedy, Shakespeare) AI has pieces from memory but often gets the exact citations wrong: it invents a hemistich, attributes a line to a different scene. For modern or niche works, pasting is mandatory.

AI doesn’t replace reading: it gives you the map of differences, which you then deepen by reading the passages. The same pattern works for Beowulf in English (Seamus Heaney, 1999 vs J.R.R. Tolkien, 2014 posthumous), or for any classic with two translations distant in time.

For people doing modern languages the pattern holds: two English translations of a Proust novel (Moncrieff vs Davis), two editions of a Sontag essay, two versions of the same Borges story. If only one English translation of a text exists, compare it with a translation into another language you know (French, German, Italian), or with the original text even if you only half-grasp the language: certain junctures show through anyway.

Formal and informal register

For students of modern languages, the same sentence in two registers is one of the most useful exercises. Classic example: an email to a professor.

Prompt:

“I need to write to a professor in English to ask for an appointment. Give me two versions of the same email: a formal one (Anglo-American academic register, never met the person) and an informal one (I’ve met him before, friendly relationship). Underline the three differences that make the register.”

You see where it changes: greeting (Dear Professor Smith vs Hi Prof), vocabulary (I would like to vs I wanted to ask, kindly vs could you), syntax (full subordinates vs broken sentences), contractions (I would vs I’d). And you learn to calibrate based on context.

It also works in reverse, to check naturalness: give AI a sentence of yours and ask whether it sounds like a native speaker or like a textbook.

“Does this sentence sound natural in Parisian French as spoken by people in their thirties, or does it sound like a school textbook? If it’s the second, tell me where it limps.”

Important: always declare the regional variant you’re working in. Iberian or Latin American Spanish (and which country)? European or Brazilian Portuguese? British, American, or Australian English? AI doesn’t guess it, and it mixes: it gives you a “neutral” coger that in Argentina is vulgar, or a British takeaway in an email sent to a university in Boston.

Catullus, three versions and two glosses

Classics student, working on the first line of Carmen 5 by Catullus (first century BC, the most famous Neoteric Latin poetic movement of the 1st century BC (Catullus is the most famous): short, lyrical, personal poetry, in opposition to traditional Roman civic epic. poet): Vivamus, mea Lesbia, atque amemus. She wants to understand the translators’ choices and the historical weight of the two verbs.

👤

You

Translate this line of Catullus into English in three versions: literal, idiomatic, free. Then add two short historical glosses on the terms "vivamus" and "amemus" in Catullus's poetic context (Neoteric poetry, first century BC). Line: Vivamus, mea Lesbia, atque amemus. Translate this line of Catullus into English in three versions: literal, idiomatic, free. Then add two short historical glosses on the terms "vivamus" and "amemus" in Catullus's poetic context (Neoteric poetry, first century BC). Line: Vivamus, mea Lesbia, atque amemus.

🤖

AI

Three versions of the line Vivamus, mea Lesbia, atque amemus: Literal: "Let us live, my Lesbia, and let us love." Keeps the word order, the conjunction atque rendered as "and", the two hortatory subjunctives at the start and end of the period. Sounds spare, but it shows the structure. Idiomatic: "Let us live, my Lesbia, and let us love each other." Keeps the vocative in place, turns amemus from an absolute into a reciprocal. Smooth English, faithful to the sense. Free: "To live, to love, Lesbia: this alone." Steps away from the hortatory subjunctive for a more direct exhortation, sacrifices the vivamus / amemus symmetry in favor of a clipped close. Betrays the form, saves the tone. **Gloss on vivamus: the hortatory subjunctive "let us live" isn't a banality in Catullus. It fits into the Neoteric climate of claiming a private and sensual life against** the civic ethic of the mos maiorum (the discipline inherited from the ancestors, made of public service and austerity). "To live" here has an almost programmatic value: choosing amorous otium as a horizon of meaning, not as a supplement to negotium. Close to Epicurean influences filtered through the Neoteric circle. **Gloss on amemus**: the reading is less unanimous. Literally "let us love" as a hortatory subjunctive, parallel with vivamus. Some philologists, however, read amemus as a collective imperative that draws Lesbia into a shared choice of life, rather than as a simple invitation to an act. The difference weighs on the translation: "let us love each other" (reciprocal act) vs "let us love" (way of being in the world). It's a nuance critical editions diverge on. Three versions of the line Vivamus, mea Lesbia, atque amemus: Literal: "Let us live, my Lesbia, and let us love." Keeps the word order, the conjunction atque rendered as "and", the two hortatory subjunctives at the start and end of the period. Sounds spare, but it shows the structure. Idiomatic: "Let us live, my Lesbia, and let us love each other." Keeps the vocative in place, turns amemus from an absolute into a reciprocal. Smooth English, faithful to the sense. Free: "To live, to love, Lesbia: this alone." Steps away from the hortatory subjunctive for a more direct exhortation, sacrifices the vivamus / amemus symmetry in favor of a clipped close. Betrays the form, saves the tone. **Gloss on vivamus: the hortatory subjunctive "let us live" isn't a banality in Catullus. It fits into the Neoteric climate of claiming a private and sensual life against** the civic ethic of the mos maiorum (the discipline inherited from the ancestors, made of public service and austerity). "To live" here has an almost programmatic value: choosing amorous otium as a horizon of meaning, not as a supplement to negotium. Close to Epicurean influences filtered through the Neoteric circle. **Gloss on amemus**: the reading is less unanimous. Literally "let us love" as a hortatory subjunctive, parallel with vivamus. Some philologists, however, read amemus as a collective imperative that draws Lesbia into a shared choice of life, rather than as a simple invitation to an act. The difference weighs on the translation: "let us love each other" (reciprocal act) vs "let us love" (way of being in the world). It's a nuance critical editions diverge on.

Replay

Worth noting: the three translations show different choices, one gloss solid and verifiable (the one on vivamus in the Neoteric and Epicurean frame) and one with an uncertainty AI declares explicitly (the nuance on amemus). That declaration of uncertainty is the good sign: it tells you where there’s debate, and where you have to check yourself before backing a reading. The gloss on mos maiorum, before citing it in an essay, you check on Treccani.

When the AI doesn’t flag uncertainty but is wrong

The good sign is when AI flags where there’s debate (like the gloss on amemus). The problem is when it doesn’t, and gets it wrong anyway: you aren’t expert yet, and you don’t have the signal to stop. Three countermeasures: always ask which source (author, manual, dictionary) the gloss comes from, then check the source; ask the same gloss to a second AI (Claude vs ChatGPT vs Gemini) and compare; for every gloss that will end up in an essay, the encyclopedic check isn’t optional.

Linguistic traps

Three types of trap AI often falls into, and where you have to keep watch.

False friends. Italian attualmente doesn’t mean English actually (it means “currently”). German Gift means “poison”, not “gift” as in English. Spanish embarazada is “pregnant”, not “embarrassed”. AI translates word by word and often stumbles. If you’re translating into a language you barely know, check the key terms in a serious dictionary (Treccani, Larousse, Duden, Oxford).

Historical registers. A sixteenth-century English expression doesn’t mean what the same expression means today. Let in Shakespeare covers a semantic field we’d now call “hinder”, not “permit”. Same problem with Middle English (Chaucer), medieval French, any text more than two or three centuries old.

Dialects and regional variants. Iberian vs Latin American Spanish. European vs Brazilian Portuguese. British vs American vs Australian English. Standard vs Austrian vs Swiss German. Always declare the variant to AI: if you don’t, it mixes, and you end up with a “neutral” email that in a specific country sounds strange.

AI in a translation exam is plagiarism

In an exam, a translation assignment, or an interpreting test (Latin, Greek, modern languages) AI isn’t a support tool: it’s a forbidden substitute. Same thing as handing in a classmate’s translation. There are no gray zones on this point.

Where there is a gray zone instead: using AI to study (the three versions, the glosses, the comparisons) is studying. Using AI to prep the specific passage you think will come up on the exam is preparing plagiarism: at the real test, on a new passage, you can’t translate anymore because you’ve trained the “what” and not the “how”. The wider ethical discussion is picked up by Ethics of studying, the last lesson of the module.

What NOT to do

Don’t hand in an AI translation in a translation exam. It’s plagiarism, and falls under the same sanctions as any other form of plagiarism.

Don’t study a text using a single translation. This whole lesson is about comparison: three angles for the same sentence, two editions side by side, two registers. A single translation hides the choices the translator made, and gives you the false certainty that “the text says this”.

Don’t trust historical glosses without an encyclopedic check. They’re the point where AI invents most: dates, attributions, etymologies. Treccani, Oxford Classical Dictionary, or a reference history of your field: two minutes per gloss, and you don’t sign off on an invention.

Don’t use AI to translate into a language you don’t know at all. If you have to send an email in Russian and you don’t know a word of Russian, AI gives you back an email in the wrong register without you noticing. For uses that matter (formal letter, application, publication), always have a native speaker review it.

Check what you’ve understood

📝 Check-in with your AI

Four questions on comparison

5-7 min of check-in

Paste the text into your AI. It'll ask four questions on the three-versions pattern, historical glosses, and edition comparison.

Copy the text below and paste it into your favourite AI (Claude, ChatGPT, Gemini, or another). Or download the file and give it to your AI as an attachment.

Copy the text Download Open

What comes next

For translations as for everything else, the background problem holds: AI can be convincing even when it’s wrong. The next lesson zooms in on this, in a study context: how to spot it when you DON’T already know the right answer.