I fed Inferno Canto 5 into ChatGPT last week. The model spat out a readable translation in about three seconds. It understood the Italian, caught the theological references, even approximated the terza rima rhyme scheme. For a study tool? Genuinely useful. For understanding Dante? That’s where the problem starts.
People are doing this right now. You absolutely can use AI to translate Dante’s original Italian. The question isn’t whether you can — it’s whether you should, and what you’re actually getting in return.
The AI Promise (and the Hidden Catch)
Modern language models like Claude and ChatGPT are genuinely impressive. Feed them medieval texts and they’ll produce something coherent. Most have been trained on vast corpora that include classical Italian literature. That’s a real advantage over earlier machine translation.
Here’s the catch nobody mentions: Dante wrote in Medieval Tuscan, not modern Italian. There’s a difference — sometimes a big one.
Models trained primarily on contemporary Italian texts will handle archaic vocabulary reasonably well. But they’ll miss the weight of unusual grammatical constructions. They’ll gloss over Dante’s intentional linguistic choices. A word like gabbo (mockery, but with specific Florentine connotations) won’t carry its full meaning. The model recognizes the denotation but loses the texture.
You could prompt more carefully. You could specify “Medieval Tuscan” and ask for literal rather than natural translation. Better models, trained specifically on classical texts, would do better. But even then — you’re extracting meaning from a fundamentally different language era. Modern Italian and Medieval Tuscan aren’t the same thing.
AI translations are often surprisingly readable. That’s actually the problem. Readability masks what’s been lost.
What Would “100 Percent Accurate” Even Mean?
Start with the machinery. The Commedia is 14,233 lines of terza rima — aba, bcb, cdc, each tercet reaching forward to seed the rhymes of the next, a chain that never breaks from the dark wood to the final vision. Italian, rhyme-rich and vowel-heavy, makes this look effortless. English is a rhyme-poor language. So every translator makes a founding choice at the door: keep the rhyme and torture the sense, or keep the sense and lose the engine that pulls the poem forward. Dorothy Sayers chose the rhymes and paid for it in contortions. Robert Pinsky kept a slant-rhymed compromise and paid in compression. Allen Mandelbaum and the Hollanders let the rhyme go and paid in momentum. Nobody got to not pay.
Now ask what “100 percent accurate” means against that menu. Accurate to the sound? The syntax? The image, the theology, the register — the way lasciate ogne speranza is simultaneously a legal formula and a curse? Pick two. On a good day, three. There is no translation that maximizes all of them, not because translators have been insufficiently clever for seven hundred years, but because the criteria pull against each other. A 100-percent-accurate translation isn’t a hard engineering target that better models will approach year by year. It’s a category error — like asking for a map at 1:1 scale.
If you want to watch this happen in miniature, take one sentence. Per me si va — the opening of the inscription over the Gate of Hell. I once spent a weekend watching six published translators turn that one nine-line inscription into six different doors. Not six drafts of the same door. Six doors.
The Thirteenth Translation
Here’s the part that took me longest to see. The model has read the tradition — Longfellow, Ciardi, Musa, Mandelbaum, Pinsky, the Hollanders, all of them, plus the commentaries. So when you ask it to translate a canto “fresh,” what comes back is not fresh. It’s the statistical center of the existing translations: a thirteenth translation that is, tercet by tercet, an average of the other twelve.
Averages are useful. That’s exactly why the AI crib is good at what it’s good at. But think about what a real translation is. When Ciardi renders the Gate as “I am the way into the city of woe,” he is committing to an echo of the Gospel of John, and he can be held to that commitment — attacked for it, defended for it, taught for it. And he stays committed: his choices bend in the same American, plain-spoken direction for thirty-four cantos. Longfellow bends the other way, toward Italianate strangeness, for the whole poem. The consistency of the bending is the interpretation. That’s what a translation is — one sustained argument about what the poem means, wrong in a chosen direction for fourteen thousand lines.
The model bends nothing. Each passage drifts toward whatever the local average of the tradition happens to be — a little Ciardi in the vigorous scenes, a little Mandelbaum in the lyrical ones. You get fluency without allegiance. And one more thing, from experience: ask it which translator wrote a particular line, and it may hand you a confident, plausible, wrong answer. I’ve caught models inventing attributions that never existed. A crib that occasionally invents its own sources is a crib you check, every time.
If a translation is beautiful and no one chose its words, who is it faithful to?
What the Crib Is Actually For
None of this means don’t use it. I use it. For untangling a knotted piece of syntax, for glossing an archaic word at eleven at night when no dictionary is in reach, for asking “what does this literally say?” — the model is patient, instant, and never tired of my questions. A reader with only AI and the Italian is far better off than a reader stuck outside the Italian entirely. Used as a crib — a running gloss you interrogate, the way students have always used cribs — it’s the best one ever built.
The trouble starts when the crib gets promoted to translation. A crib answers “what do the words say?” A translation answers “what does the poem mean?” — and answers it by choosing, line after line, what to keep and what to give up. The model is built to avoid exactly that choice. It gives you the reading that offends no prior reading. Dante, of all poets — the man who put popes in Hell by name — deserves better than the reading that offends no one.
The Italians have a pun old enough that it feels like it was waiting for this moment: traduttore, traditore. Translator, traitor. Every translation betrays the original; the tradition’s only defense was to demand that the betrayal be chosen, signed, and consistent — a traitor you could name and argue with. The machine offers the first unsigned betrayal in the history of the art. Something is always lost. The question is whether anyone chose what to keep.
