Learning Italian With A.I.

From "Spaghetti Bolognese" to "Il Gattopardo" in five months.

Background

I was, I realised, a bit bored. My working days have their longueurs, times where I have to be present, but would prefer not to be involved, and for several years, I had filled them by writing tendentious articles about politics. I felt, however, that I had at least a couple of toes over the line of the Columnists’ Paradox – I had written enough that I no longer needed to be read (my pixel trail was sufficiently long that any fool benighted enough to be interested in my views could easily work them out). Several months of “the government are idiots” articles had sapped my enthusiasm for several more years of the same.

What to do? There is, after all, a limited amount of internet chess anyone can play. Particularly with my rating…

For some reason, call it fate, call it karma, call it destiny, some words of my old Latin teacher, the sainted Mrs Noble, came to mind. Tasked with persuading a bunch of 12-year-olds they wanted to learn a language they would never speak, she fell back on telling us that sticking with the language to A-level would allow us to knock off Italian in six months.

A language? Hmm. That was an interesting idea. Knowing languages is generally a good thing. It would be a challenge. I hadn’t learned anything academic for a couple of decades, so it would be interesting to find out if I still could. If I asked an A.I. to teach me, it would give me some insight into the current state of the technology and it would be there whenever I wanted it, fitting into the constraints of my schedule.

If I was going to learn a language, Italian made a lot of sense. I did French to A-level but carried on with Latin at University. I had the parent of the Romance languages, and one of the family members already. My childhood and degree gave me Modern and Ancient Greek to throw into the mix. They might help too.

When I learned to ride a horse, I set a Point of Sufficient Mastery – a level which told me I had become good enough not to need to get any better. I wanted to be able to get on a horse and ride it, no matter the terrain, not win an Olympic medal. When Dobbin sailed over the (quite low, to be honest) bar, and I stayed on his back, I declared job done.

Having no imminent prospect of being in Italy, a speaking challenge wasn’t going to work but at college I had once declared that I was waiting to read Dante in the original. Age may or may not bring wisdom, but it does bring a degree of realism. Dante was not, I decided, an appropriate PSM. He is too long - I was looking for a project, not a life’s work. Mediaeval Italian would be an impressive thing to learn, but the spin-off benefits would be limited – an Italian who learned Shakespearean English would stand out slightly in The Dog and Duck…

The Leopard. That would do. Modern. Short (at about 240 pages, it straddles the novel/novella boundary). A great work of literature, just the sort of thing that would be an appropriate target.

The Plan

If I was learning Italian to be able to read a book, I was, effectively, treating it like a dead language and so I decided I could learn it like a dead language. I would start by doing the grammar and then I would look to build out the vocabulary. Remembering that, having only done Latin to English at school, doing English to Latin at college had been a bit of an eye-opener (not least to my long-suffering tutor…), I decided it would be useful to build English to Italian into the process from the start. Learning to read and write would give me a solid base if I ever wanted to listen and speak, the latter two, basically, being the former two speeded up.

The Process

One morning in July, I opened ChatGPT, explained my plan and asked it to help. Which it was happy to do (A.I.s are always happy to help). We started with verbs, moved on to nouns and the other in-filly parts of speech. Italian is not a very difficult language grammatically. Where Latin thinks you can do six things to a noun in the singular and plural, Italian has decided that one will be perfectly sufficient (Greek displays a similar phenomenon, leading to the conclusion that modern languages are ancient languages which failed their GCSEs…). My classical background helped in the main; having been taught grammar through it, things like subjunctives were already familiar. But not always – “cenavi” to a Roman would mean “I had dinner (once)” to his present-day descendant, it means “You used to have dinner”. Greek uses “a” as the ending for the first person in the past tense. Italian uses it for the third person. If nothing else, the mistakes I made showed me how deep the grooves left by my education were.

Everything was going swimmingly until about three weeks in, when Chat went a bit funny. Even with my incredibly basic Italian, I realised the answers it was giving were wrong. This seemed a bit of a problem. But one advantage of the current A.I. landscape is that there are countless models available so I explained the situation to Google Gemini and it graciously consented to be my substitute teacher (its words).

I do not have a paid account with Chat (a: I’m cheap and b: locking oneself into an ecosystem which may, due to the rate of change in the sector, be fourth-in-class in a few months doesn’t seem very wise). As a result, it only gave me a limited number of interactions before changing to an older model. If you ask it to produce something for you (a cheat-sheet, for example), it will only let you use the most advanced model and lock you out of the system for a few hours after your limit of interactions has been met. Initially slightly irritating, this turned out to be a blessing in disguise. Spaced Repetition – doing a bit, leaving it, and coming back later – is, apparently, one of the best ways of learning something and Chat had forced me into it, so I carried on doing short bursts with Gemini, even though it has no limit to interactions, each session lasting about 20 minutes.

Phase I – learning the grammar – took about five weeks and it was time for Phase II – reviewing what I had learned and looking to build out the vocabulary. Every session, Gemini would take me through a point of grammar again and give me a list of words, thematically grouped (colours one day, clothes the next). It was at this point that the exceptions emerged – (“Most of the time you would be right, but in this case…”). Phase I taught me that, like French, Italian uses “to have” to form the past of verbs (j’ai lu/ho letto) apart from those related to movement (je suis venu/sono venuto), in phase II it emerged that some can use either depending on what they are doing. As a rule of thumb, exceptions to rules take just as long to learn as the rules themselves…

Around this time, I persuaded Chat to work again and decided to split the task between the two systems. Gemini would teach me general Italian and, since I was aiming to read a novel, Chat would teach me literary Italian. This meant that I ended up having a third spin through a lot of the grammar which, Spaced Repetition again, further helped to cement it. I would have a lesson with Gemini in the morning, a lesson with Chat in the afternoon and a review session in the evening, both of them giving me a passage of Italian to translate. As a cross-check, I started scrolling through the websites of La Repubblica and Corriere della Sera every evening, which helped gauge my “real-world” progress and added to my vocab (particularly about crime – the Italian media, it seems, really love a crime story).

It was about six weeks into Phase II that Gemini fully came into its own. One day, expecting my usual exercises about grammar and some vocab, it informed me that we had probably taken that approach as far as we could. Why didn’t we just have a chat (in Italian, of course)? A slightly faltering discussion about films then ensued. From then on, it would write texts for me to read and answer questions on, give me the first line of a story and ask me to complete it, give me sentences and ask me where the mistakes were or to rewrite them in my own words. One day it got me to translate into Italian the instructions for putting together and using a coffee pot, another, to translate into English a popular Italian song from the sixties. One exercise (like a comprehension exercise about parmesan) would often tee up a discussion the next day (“Can you think of any similar products in your country? Tell me about them”). Every so often, it would throw in a curveball – “Remember that tricky grammatical construction? Here are five sentences using it, give them to me in Italian.”

I was conscious, because of the way I had approached the matter, that I had effectively skipped GCSE Italian and gone straight to A-level. I could discuss the depopulation of rural villages, but “Can you tell me where the campsite shop is? I need to rent two towels” would have flummoxed me. At this point another fortunate accident occurred. The all-seeing eye of Facebook, observing, no doubt, my frequent trips to Google Translate, had twigged that I was learning Italian so started showing me posts from various “Learn Italian” resources which covered the language one might use, not just the language one might read.

Curious one day about my progress, I asked Gemini which assigned me a ranking on the EU’s schema of language proficiency. Flattered but sceptical about its answer, I explained to Deepseek how long I had been learning and asked it to give me an appropriate passage to translate. Repeating the process until I started to struggle confirmed Gemini’s rating and from then on, every Friday I asked the Chinese site to give me three passages to translate as a sort of mini-test/progress check.

Conscious that the Christmas deadline I had set for starting to read The Leopard was looming, I decided to focus more intently on reading Italian. Gemini gave me adapted passages from the book itself and then Primo Levi’s If This Is a Man and Calvino’s The Baron In the Trees. Chat excelled itself by writing a psychological thriller/horror story which was very good not just as an exercise, but as a story (I still want to know what the thing living in the walls was). Set in the dying world of the 19^th-century Sicilian aristocracy, I decided I probably need some specific vocabulary to tackle my target, so I got Grok to give me a list of 20 “Leopard-like” words every day, slotting it in after lunch. As weeks turned to days and aware that the novel is written in very long, twisty sentences, I got both Chat and Gemini to give me a crash-course in that sort of sinuous prose to get my eye in.

The Final Exam

On the afternoon of Christmas Day, I cracked open The Leopard (probably, I thought, more interesting than the King’s Speech). I had decided to “raw-dog” it. No translation. Just the text, a dictionary and a pencil. At the end of the first page, I thought, “I can do this”. An hour and only ten pages later, I thought, “How long will it take me to do this?”

Reassured by my silicon senseis that a) the opening is very dense, b) the prose opens out as the book continues, c) despite the “training” I had done, I would still adapt to the style and d) even Italians have to read it slowly, I carried on. Swapping the dictionary for Google Translate helped, the prose did open out and I probably did get my eye in a bit more, so the pace did pick up somewhat and on the 2^nd of January, The Leopard joined the (disappointingly) short list of great books I have finished.

Observations

Since the A.I.s got me to my PSM a couple of weeks ahead of my teacher’s schedule, I have to conclude that they are, at the moment, an effective way of learning languages. They were there when I needed them to be, unfailingly patient and encouraging (even when I got everything wrong, I was still producing ottimo lavoro) and, not unimportantly, free. The only cost I incurred in the whole process was the price of my copy of The Leopard and I used that not to learn Italian but to tell myself that I had learned Italian.

They are not, however, flawless. After a certain length of thread, the quality of the output tends to deteriorate markedly, necessitating a fresh one. Where exactly this point comes is more of an art than a science. Gemini is the only A.I. to display how many tokens have been used; on it, after 180,000 or so, there be dragons. For the others, it is a matter of waiting for gobbledygook to be produced. For those learning a language, periodically asking how one stands compared to an external benchmark like the E.U.’s functions as a “save point” in a game, allowing starting a new thread to go smoothly. In other subjects, where no such framework exists, it would be a matter of mechanically telling a new thread what one had learned.

Several of the models (Chat, Grok and Deepseek) have slightly one-track minds. No matter how many exercises on pronouns I had done, Chat would suggest another, Grok decided that, having learned some words beginning with “s”, I only wanted to learn words beginning with “s” and Deepseek, having given me a wistful, slightly philosophical passage, decided I only wanted wistful, slightly philosophical passages. When I pointed this out, they were more than happy to change tack (Deepseek responded with a tendering document for the removal of asbestos from a school), but they do sometimes need a nudge.

No matter how good their teaching, it is still, unfortunately, up to the student to do the learning. They may slightly speed up the process by always being available, but the information they offer goes into the human mind at the same rate it would with a human teacher. In one way, they may actually slow learning slightly. Just looking at a list of words on a screen means one misses out on the well-attested memory-boosting effects of the physical action of writing. I could, obviously, have copied them down, but convenience (by which I mean laziness) meant I just scrolled through them on my laptop. English-speakers already know, without being aware of it, quite a lot of Italian – “memoria” is unlikely to be anything other than “memory” – but there is often a vowel shift (“To remember” is “ricordare” which looks sufficiently similar to “record” that one can guess when reading but one needs to remember when writing that the English “e” has become an “i”). Going offline would probably help cement the differences.

My Italian is not perfect by any means. A recent TV documentary with a couple of Italian talking-heads (of which I understood about 80%) suggested that learning to read and write has bled into my ability to listen and comprehend, but there is still work to do. This, I think, is to be expected given the approach I took. Had I chosen a different route, I am sure the A.I.s would have got me to 100% (Gemini offered, for example, to role-play an interaction in a restaurant) but with the side effect that my knowledge of the dress of nineteenth century Sicilian aristocrats would obviously have been poorer. To learn is to choose.

Reflection

A few days before I opened The Leopard, an irony struck me. I had spent several months using an A.I. to learn how to do something an A.I. would happily do for me. Reading literature in the original has always been a luxury pastime – there are plenty of translations of The Leopard, as well as a film and mini-series had I just wanted to know the story. It was not just that I did not need Italian to read the book, but that now I didn’t need Italian at all. If I were in Italy, my phone would happily translate anything I saw or heard, and anything I wanted to say. There was no necessity to know the language at all. It was, in effect, a party trick I had spent several months learning. Nice to have, but in no way necessary.

My mind turned to a series of experiments conducted by John Calhoun in the fifties and sixties. Endeavouring to provide the perfect environment for rats, he discovered that rather than the population exploding as he expected, at a certain point, his rodents started to isolate themselves, concentrating on grooming to the extent that he termed them “the beautiful ones”. In a way I had done the same thing. I had improved myself, but with no real purpose other than improving myself. Reading a novel about the past had perhaps given me a glimpse of the future, for if A.I. continues to follow the projections of its promoters, it will not leave much else for people to do. It may allow us to become more accomplished individuals, but leave us without any real-world accomplishments.

The Leopard is the story of a Sicilian aristocrat watching the inevitable death of the world he knew. I had enjoyed it. But I felt no triumph when I finished it because the way I came to it gave me the uncomfortable sense that I was doing the same.