Meal planning has two halves. There is the admin: deciding what to eat, writing the list, avoiding the same dinner three nights running, using up half a cabbage. And there is the cooking: timings, temperatures, quantities, whether the chicken is done.
AI tools are good at the first half and unreliable at the second. Split the job along that line and they earn their place.
Where it genuinely helps
Turning a plan into a shopping list
Give it five meals and it produces a consolidated list grouped by aisle, adding up the onions across recipes. This is dull work a computer should be doing.
Working backwards from the fridge
"I have half a cabbage, three eggs, feta and rice, and twenty-five minutes." This is its strongest use, because it is a matching problem rather than a knowledge problem, and you can judge any suggestion before you cook it.
Rotation and variety
Households drift into six dinners on repeat. Asking for a fortnight of meals that share ingredients but not shapes breaks the loop, and the sharing keeps waste down.
Scaling and adapting
Doubling a recipe for eight, or converting cups to grams. Check the arithmetic on anything critical, especially baking, where scaling is not always linear and eggs do not divide neatly.
Where it goes wrong
Invented recipes that read fine
The output is fluent, which is the problem. A generated recipe can have a plausible structure, plausible steps and a ratio that produces a wet, flat cake. Nothing in the writing signals that it has never been cooked.
Baking ratios
Baking is chemistry with narrow tolerances. Flour to fat to sugar to liquid has to be right. This is where generated recipes fail most often and most expensively, since you find out an hour later.
Food safety
Treat any AI statement about cooking temperatures, cooling times, canning, fermenting or preserving as unverified. These are the areas where being wrong makes people ill. Check the official guidance for your country instead.
Confident invention
It will attribute a dish to a region it does not come from, describe a traditional method nobody uses, or give an ingredient a property it does not have. Fluent and wrong is its default failure mode.
How to use it without getting burned
- Ask for a plan, not a recipe. Let it decide what to cook on Thursday. Get the method from a source that has actually cooked it.
- Give it your constraints. Equipment, time, allergies, what people refuse to eat, what is already in the cupboard. Vague questions get generic answers.
- Sanity-check quantities. Does 200g of rice for four sound right? Does that cake have enough flour for that many eggs? If a number looks odd, it is.
- Never take its word on safety. Internal temperatures, cooling, preserving, ferments: verify elsewhere.
- Use it to adapt, not to originate. "Make this recipe I already trust vegetarian" is a far safer request than "invent a recipe".
A worked example
Ask for five weeknight dinners for two, thirty minutes each, no pork, one vegetarian, sharing ingredients so nothing goes off, with a shopping list grouped by supermarket section and a note on what to prep on Sunday.
That is a scheduling question, and the answer will be useful. Then cook each dish from a method you trust, or from your own head. The planner handles the week; you handle the pan.
The honest summary
Use it as a kitchen secretary. It is quick, tireless, good at lists, and it will keep the fridge from turning into a compost heap. It is not a cook, it has never tasted anything, and the moment it starts telling you how long to cook chicken you should be reading something else.