We gave an AI a complete picture of a kitchen — 350 items, every one of them sitting on a real shelf — and asked what to make for dinner. It said we were out of dal.

We had four kinds of dal. Moong, urad, chana, and a bag of daliya. All present, all photographed, all in the list it had just read.

The interesting part is not that it was wrong. It is that it was reasoning correctly from what it had been told.

What the AI was actually given

Behind every "what can I make" feature is a block of text describing the kitchen. Ours looked roughly like this:

Moong Dal (Split Mung Beans) (some) — stored in: Shelf [OUT OF STOCK]
Basmati Rice (some) — stored in: Shelf [OUT OF STOCK]
Atta (Whole Wheat Flour) (some) — stored in: Back side [OUT OF STOCK]

Given that, refusing to plan a dal dinner is the right answer. The model was being careful. It was told the dal was gone, and it did not invent a recipe around food it believed was missing — which is exactly what you would want if the label were true.

Where the label came from

Nobody wrote "out of stock" anywhere. The app calculated it.

Stock status was derived from quantity: no quantity recorded meant nothing left. And after photographing a shelf, almost nothing has a recorded quantity — a photo shows a jar of turmeric exists, not how full it is.

So an empty field became a zero, a zero became "out of stock", and "out of stock" became a label the AI had no reason to doubt.

Three rounds of fixing the wrong thing

Here is the part worth sitting with. When the suggestions came back wrong, our first instinct was that the AI needed better instructions.

We told it to prefer the cuisines the household actually cooks. It still suggested pesto pasta. We made that instruction more forceful — "IMPORTANT", "default to these" — and it suggested pesto pasta with more conviction.

Both changes were reasonable. Neither addressed anything, because the problem was never the instruction. The model was being handed a description of an empty kitchen, and no amount of telling it to cook Punjabi food will produce dal from a pantry it believes has none.

We spent two rounds rewriting prompts before checking whether the data reaching the prompt was true.

The actual bug was four lines of plumbing

The kitchen description was assembled by copying each item into a new shape — name, quantity, unit, location. That copy quietly dropped the stock field. It was fetched from the database and lost on the way to the prompt, so the code fell back to guessing from quantity.

Not a hard bug. Just an invisible one: reading a field that was never copied gives you nothing at all rather than an error, so everything downstream carried on confidently with a blank.

Fix that one copy, and the next answer was Moong Dal Tadka with basmati, chana masala and roti.

What this says about AI features generally

The lesson is not "check your data", which everybody already agrees with and nobody does under time pressure.

It is that a language model's output is downstream of context you control, and when it says something odd, the context is the more likely culprit. A model that has been told the pantry is empty will behave exactly like a model that is bad at cooking. From the outside these look identical — confidently wrong answers — and the instinct is to blame the part you can see.

The tell is specificity. A model that is genuinely poor at a task is vague. Ours was precise: it named the dals, said they were marked out of stock, and refused on that basis. That precision was a signal it was working from something, and worth asking what.

A practical test

If an AI feature is giving answers that seem to ignore what you told it, print the context. Not the prompt you wrote — the text that actually arrived, with real data filled in.

Ours would have taken thirty seconds to read and would have shown "[OUT OF STOCK]" beside every ingredient in the house. That is what we should have looked at first, before touching a single word of the instructions.

The AI Said Our Pantry Was Empty. It Was Full.