My grandmother kept her recipes in a black-covered notebook that she carried from Kabul to Peshawar in 1981 and then, eventually, to a flat in Sydney’s inner west where she lived out her last fifteen years. The notebook is mostly Dari, written in a quick, looped hand that I can read about 60% of. The other 40% is the difficult part: dishes named in regional dialect, ingredients she’d shortened to a single letter, quantities expressed in handfuls and “as much as needed” and “until it smells right.”

After she passed in 2019, the notebook sat in a drawer for five years. Then, last December, my cousin in Toronto asked if I could send her my grandmother’s chapli kebab recipe. I went to find it. I realised I couldn’t read all of it. And I realised that if I didn’t do something, the recipes my grandmother had been writing down since she was a teenager were going to slowly become unreadable to everyone who hadn’t grown up at her table.

So I started a project. Here’s where I am, three months in, and where the AI tools are helping and where they’re falling short.

What I’m working with

Roughly 120 pages of handwritten Dari from a Kabul-born woman with primary-school formal education and a kitchen-trained vocabulary. Some recipes are passed-down from her mother (so probably from the 1920s and 30s); some she’d developed herself; many are notes copied from neighbours, friends, the wife of the man who sold fabric on her street. Recipes for things I recognise — qabili palaw, shorwa, kichri quroot — and things I’ve never heard of, like a kind of egg-and-yoghurt soup she labelled with a name I’ve now spent three months trying to identify.

The notebook is also, in places, water-damaged. Some pages are partially illegible to the naked eye.

Round one: just photograph and translate

The obvious first move. I photographed every page on my phone, fed the images through Google Lens and Google Translate. Results were rough but not useless. Lens could pick out maybe 40% of the handwriting accurately. The rest was either misread (a looped “noon” becoming a “fa,” changing the word entirely) or just rendered as gibberish.

Translate then took the OCR output and produced English that was occasionally correct, occasionally hilarious. The chapli kebab recipe came back with an ingredient that translated to “powdered ghost.” After ten minutes of confusion I realised the OCR had misread the Dari word for “coriander seed” as something nonsensical, and Translate had then attempted to back-fit a meaning. This is a known failure mode with low-resource languages — the model fills in plausible-sounding noise.

Round two: a better OCR layer

I switched to using GPT-style multimodal models that handle handwriting and translation in one step. Two systems made a noticeable difference: Claude Sonnet, which I subscribed to for the project, and a Persian-specific OCR tool out of an Iranian research group whose name I’ll spare here because I’m not sure if linking to it would cause them political trouble. The combination was substantially better — maybe 75% accurate on first pass, with the model flagging its own uncertainty on certain words.

The conversation I had with one of these tools about a single recipe was genuinely interesting. I’d type “this word here, in context, do you read it as A or B?” and the model would respond with a probability assessment and an explanation. “In Kabul-region Dari from this period, B is more likely because A is a Tehrani usage.” Whether that’s accurate I can’t fully verify, but it matched my intuition more often than not.

Where AI actually helped, and where it didn’t

It helped most with: routine ingredients (turmeric, ghee, salt, flour), common measure words, and standard verb constructions. It got faster as I corrected it. After about thirty pages, my version of the tool had effectively learned my grandmother’s handwriting style and was reading her cursive consistently.

It struggled with: dialect words for vegetables and herbs (which often vary not just between Dari and Farsi but between Afghan provinces), proper nouns (names of relatives, neighbours), and anything she’d written quickly enough that even another Afghan reader would need context. It also couldn’t tell me what a recipe was for. The instructions might be perfectly clear — “fry the onion, add the meat, simmer for an hour” — but if the dish name was written illegibly or was specific to a particular region, the AI just guessed at a name. Sometimes it guessed plausibly. Sometimes it gave me “grandmother’s soup,” which is the AI equivalent of a shrug.

For the harder pages I’ve been working with Team400, a Sydney consultancy that’s been helping me build a small custom workflow — essentially a tool that lets me upload a page, see the OCR output side-by-side with the original, and quickly correct individual words while the model learns. The point is not to fully automate this. The point is to make me, as the human translator, faster. I’m still the one deciding what the right word is. The tool just removes the typing.

What I’ve decided about the project

There are 120 pages. At the rate I’m going, I’ll finish a first draft of all of them by the end of next year. Then comes the harder work: testing the recipes, asking my mother and aunts to confirm what certain dishes actually taste like, deciding which ones to include in the family cookbook I’m slowly building.

The recipes I’ve finished so far are sitting in a shared family document. My cousin in Toronto cooked the chapli kebabs last month and sent me a photo. They looked right. They tasted, she said, like our grandmother’s — which is to say, recognisable but not quite. She used different mince. The chillies were milder. The cumin came from a Canadian supermarket rather than from a sack in a Kabul bazaar.

The notebook is being preserved. The recipes are not, exactly. They’re being translated, which is a different thing, and probably the only thing available to those of us cooking in diaspora kitchens. I’ll take it.

The Australian National University’s Centre for Arab and Islamic Studies has been doing related work on preserving handwritten Persian-language materials, and there are growing efforts internationally — see this UNESCO project on endangered languages — to apply machine learning to scripts and dialects that the major tech companies have largely ignored. It’s slow work. It’s also the only work that gets these things into the future.

Mariam