"AI recipe app" now covers several quite different products, which makes it hard to judge whether any of them are worth your time. Some are search engines with a chat box bolted on. Some generate recipes from scratch. Some look at a photo of your fridge. They fail in different ways, and it is worth knowing which one you are holding.
This is a straight explanation of the moving parts, written by people who build one of these. We have tried to be honest about the limitations, including ours.
The three things under the bonnet
1. Seeing what you have
The photo step uses a vision model: you take a picture, it detects objects and returns labels with confidence scores. "Bell pepper, 0.94. Carton of milk, 0.88."
What it is good at: things that are visible, distinct, and roughly where you would expect. Loose vegetables, bottles, packaging with clear branding.
What it is bad at, and will remain bad at:
- Things it cannot see. Anything behind something else, in a drawer, or in an opaque tub is invisible. A photo captures a surface, not an inventory.
- Quantity. It may see cheese. It has very little idea whether that is 30g or 300g.
- Ambiguity between similar items. Plain and Greek yoghurt look identical. So do several white powders in unlabelled jars.
- State. It cannot reliably tell fresh from three-days-past.
This is why receipt scanning is often the more useful half of the feature and gets less attention. A receipt is already a structured list with quantities on it. It tells you what entered the house, which is a much better starting point than what happens to be facing the camera.
2. Turning that into a recipe
Once there is a list of ingredients, a language model composes a recipe. This is where the real difference between products lies, because "generate a recipe" can mean two very different things.
The weak version retrieves something close from a database and swaps a few ingredients. You can spot it: the instructions do not quite match the ingredient list, or a substitution has been made that does not work — swapping a structural ingredient in a bake, say, where the chemistry actually matters.
The stronger version constrains generation properly: it respects what you actually have, applies your dietary rules as hard filters rather than suggestions, and produces quantities and times that are internally consistent. It is still generating, so it can still be wrong — but wrong in the way an inattentive cook is wrong, not wrong in the way a broken search result is wrong.
3. Working out the nutrition
Macros are computed by matching each ingredient to a food composition database and scaling by quantity. The databases are solid. The uncertainty is almost entirely in the quantities.
A recipe says "1 onion". Onions vary by a factor of three. It says "a drizzle of oil", and how much ends up in the food rather than the pan is genuinely unknowable. This is not an AI limitation — it applies equally to the numbers printed in any cookbook.
The practical consequence: these numbers are good for direction and magnitude, not precision. That is usually enough, for reasons covered in the guide to macros for home cooks.
Where they genuinely help
- The blank-fridge problem. Being handed one concrete answer beats scrolling twenty tabs when you are tired. The value is in removing the decision, not in the recipe being remarkable.
- Using things up. A system that knows what you own can build around what is about to turn, which is the core of not wasting food.
- Dietary filtering. Applying vegetarian, halal, gluten-free or an allergen exclusion across every suggestion is tedious by hand and trivial to automate.
- Nutrition without arithmetic. Getting a macro breakdown without weighing and looking up each component removes most of the friction from tracking.
- Shopping lists that subtract. If the app knows your inventory, it can list only what you are missing.
Where they still fall short
Being straight about this matters more than the feature list.
Technique does not transfer. A generated recipe can tell you to cook until the sauce coats the back of a spoon. It cannot show you what that looks like. For anything where technique carries the dish — bread, pastry, emulsions, most baking — a good cookbook or a video is still better, and it is not close.
Baking is a weak spot generally. Savoury cooking is forgiving of improvisation. Baking is chemistry with ratios that matter. Treat generated baking recipes with more suspicion than generated stir fries.
Confident wrongness. A generated recipe reads with exactly the same authority whether it is sound or slightly off. There is no tell. Keep your own judgement switched on — if a cooking time looks wrong for the size of the cut, it probably is.
Allergies need a human check. Filtering is useful and worth having. It is not a substitute for reading the label on a packet, because the app cannot see cross-contamination or a reformulation.
Photos are not inventories. Worth repeating, because it is the most common disappointment. Expect to correct the list. Any app that pretends otherwise is overselling.
How to judge one
If you are trying a few, these are the questions that separate them:
- Can you correct what it thought it saw, easily, before it generates?
- Does it keep a persistent inventory, or start from zero every time?
- Are dietary rules hard filters or polite suggestions?
- Do the quantities and instructions match each other? Read one recipe closely.
- Does the shopping list subtract what you already have?
- Can you get your data out or deleted?
That last one is worth caring about. You are handing over a fairly intimate record of what you eat.
Gurt AI does the things above: scan a fridge or a receipt, keep a live inventory you can edit, generate against what you actually own, filter on allergies and diet, and build a shopping list that leaves out what is already in the house. We do not sell your data and you can delete your account and everything attached to it.
We would still tell you to use a cookbook for bread.
The honest summary
An AI recipe app is a decision-removal tool, not a cooking teacher. It is at its best on an ordinary weeknight when the question is "what do I do with this" and the answer needs to arrive quickly. It is at its worst when technique is the point.
Used for the first thing, it earns its place. Expected to do the second, it will disappoint you — and that is a reasonable trade to understand before you download anything, ours included.