AI for meal planning: dinner from the fridge you already have

It is 5:40pm. The fridge holds half a roast chicken, a bag of spinach one day from the bin, six eggs, and yesterday’s rice. You know how to cook all of it. What you don’t have is the idea, and by the time it arrives you will have opened that fridge door three times, as if the answer might have restocked itself.

This post is about the least glamorous thing AI does well, which is also the one you would use every single day: answering “what’s for dinner?” from what you already have. Not a demo, a habit. The method below covers the fridge-photo trick, a week plan that survives contact with a real household, the grocery list, dietary conversions, and scaling. It also covers the failure modes honestly, because an AI once told supermarket shoppers to drink chlorine gas, and the line between those two facts is exactly where you need to stand.

Deciding dinner costs more than cooking it

The deciding is not a small cost. In a May 2026 Talker Research survey of 2,000 US adults, the average respondent spent almost 16 minutes a day deciding what to have for dinner. That is over an hour and a half a week, roughly four full days a year, spent not cooking, not eating, just deliberating. Plenty of actual dinners take less time than the decision that preceded them.

The undecided fridge has a second bill attached. ReFED, the nonprofit that tracks US food waste, estimates the average American spent over $760 on food that went uneaten in 2024. A lot of that is exactly the spinach in the opening paragraph: bought with intentions, waiting for an idea that never came.

Here is why AI happens to be unusually good at this problem. Dinner planning is constraint satisfaction over a small inventory: these ingredients, this much time, these preferences, go. There is no single right answer, the stakes of a mediocre suggestion are one mediocre meal, and the task repeats daily so small savings compound. That profile, boring, bounded, and frequent, is where current AI delivers its most reliable value, far more reliably than the spectacular demos.

Photograph the fridge. That’s the trick.

The move that converts skeptics takes twenty seconds. Open the fridge, photograph the shelf, send the photo to any of the major AI chatbots with one line: “What can I make for dinner tonight from this, in under 30 minutes?” Modern chatbots read images natively, and a fridge shelf is an easy read: it will identify the half chicken, the wilting spinach, the eggs, and propose three real dinners built from them, usually leading with whatever is closest to dying.

For scale on how good a deal your phone camera is here: Samsung’s flagship smart fridges ship a feature called AI Vision Inside, an internal camera that, as of Samsung’s March 2025 announcement, recognizes 37 food items and cannot see into the door bins or the freezer. A phone photo and a general-purpose chatbot recognize essentially anything with a shape or a label, door bins included, on hardware you already own. The appliance industry is building the fridge photo into the fridge; you do not need to wait for it.

Two refinements make the answers noticeably better. First, tell it your pantry baseline once: “assume I always have oil, onions, garlic, rice, pasta, canned tomatoes, soy sauce.” Fridge photos under-represent the boring staples that make dishes work. Second, state the appetite and the clock: “two tired adults, 25 minutes, one pan” produces dinner; an unconstrained ask produces a food blog.

Be fair about what the photo cannot see. Opaque containers are a mystery, the vegetable drawer is closed, and the model will politely guess at the yogurt tub that is actually Sunday’s gravy. The fix is one line of narration alongside the photo: “the containers are leftover chili and cooked rice, and there are carrots in the drawer.” You are not doing the AI’s job for it; you are doing the two seconds of disclosure that turns a decent answer into the right one.

A week of dinners holds up when you brief it like a person

The nightly version is a rescue. The weekly version is where the time actually comes back. The difference between a useless AI meal plan and one you follow is the same thing we found when we planned eight days in Japan with AI: the brief. “Give me a healthy meal plan” returns the average of every wellness blog ever written, all quinoa and optimism. A brief describes your actual household: who eats, how much weeknight time exists, what the kid vetoes, and your leftover strategy.

The leftover strategy is the load-bearing constraint. Ask for one double-batch dinner that covers two nights, and for a Sunday cook that seeds two weeknight meals. Cap novelty at one new recipe a week, scheduled for the weekend, so the inevitable experiment-gone-sideways lands on a night with a pizza fallback rather than a Tuesday at 7pm with a hungry six-year-old. And end the brief with the move that improves every AI planning session: “ask me five questions before you start.” It will ask about allergies, budget, and spice tolerance, and the plan that comes back will be yours instead of the internet average.

Notice what the structure is doing. The 15-minute nights land on the days the plan assumes you are tired, because a plan that demands Wednesday heroics is a plan you abandon by Wednesday. The double batch is scheduled before the night it feeds. And the picky-eater constraint is baked into the dishes rather than bolted on, which is where AI planning quietly shines: “the kid eats pasta, rice, and eggs but nothing spicy” produces dinners where the child’s portion splits off before the chili flakes go in, a trick experienced parents know and no generic meal plan bothers to encode.

Which chatbot you use matters less than the brief; any of the big ones plans a competent week, and our task-by-task comparison covers the differences if you care. The plan is also renegotiable in one line mid-week: “we ate out Tuesday, the chili never happened, replan Thursday and Friday around it.” Replanning, which used to mean the whole system collapsing, is now a sentence.

The grocery list is the quiet killer feature

Nobody buys an AI subscription for grocery lists, and then the grocery list turns out to be the feature they would miss most. From a week plan, one prompt produces a single consolidated list: duplicate ingredients merged across recipes with summed quantities, items grouped by supermarket aisle, and, if you gave it the fridge photo, everything you already own subtracted out. The dumb, error-prone half hour of cross-referencing seven recipes against your cupboards compresses to seconds.

The list is also where the plan meets the budget. Ask it to flag the expensive lines and propose cheaper swaps before you shop, and to keep a rough running total. The numbers it estimates are guesses, not your store’s prices, so treat the total as a shape rather than a fact. But the structural win is real: a week where every purchased ingredient has an assigned job is the opposite of the $760 of goodwill produce composting in the crisper drawer.

Swaps and scaling beat the cookbook

A printed recipe is frozen: four servings, wheat flour, dairy, done. The conversation is where AI beats the cookbook, because your actual question is rarely “what is the recipe” and usually “what is the recipe, except my sister-in-law is vegan and we’re six people now.” Paste any recipe and ask for the conversion with one crucial instruction: keep the method, change only what must change, and say what happens to time and texture.

Scaling has one honest rule: pots scale, ovens are chemistry. Soups, stews, curries, and trays of roast vegetables multiply from two servings to six almost linearly, and AI handles the arithmetic and the pot-size warning without complaint. Baking does not scale linearly, and neither does heat: ask it to triple a chili and a good model will triple the beans but hold back on tripling the chili flakes, because capsaicin compounds faster than volume. If it doesn’t volunteer that caution, ask “what should not scale linearly here?” before you cook for a crowd.

It’s a confident cookbook author, not a chemist

Now the failure modes, because they are real and one of them is famous. In August 2023, the New Zealand supermarket chain Pak’nSave ran an AI meal-planning bot that suggested recipes from whatever ingredients users typed in. Users typed in bleach, ammonia, and water, and the bot cheerfully proposed an “Aromatic Water Mix”, “the perfect non-alcoholic beverage to quench your thirst,” which is a recipe for deadly chlorine-family gas. The bot was not malicious. It was doing exactly what these systems do: producing plausible-sounding text about ingredients with total confidence and zero understanding of chemistry.

The everyday version is subtler. AI recipes read right and mostly are right: a 2025 Utah State University study found participants rated AI-generated recipes comparable to human-written ones on ease, ingredients, and time, though the AI splurged on pricier ingredients, and the researchers concluded that using it for complicated dishes remains “a gamble.” The model has read every recipe on the internet and tasted none of them.

So learn the tells of an invented recipe. A braise that claims to be done in 25 minutes. A step that says “add the reserved marinade” when no step reserved one. Baking ratios that appear from nowhere, because baking is the one genre where a plausible-sounding number produces a brick. When a dish matters, ask the model to cite the tradition it is drawing on and sanity-check the timing against one trusted source. For weeknight improvisation over ingredients you know, none of this is necessary; the stakes are one shrug and a sandwich. Treat it as a confident cookbook author with no sense of taste, and cook accordingly.

And one rule is not optional. If anyone at your table has a food allergy, you verify every ingredient and every label yourself, every time. A peer-reviewed 2023 study in Nutrition tested 56 ChatGPT-generated diets for people with food allergies and found the results generally accurate but with “the potential to produce harmful diets.” Generally accurate is a fine grade for a Tuesday stir-fry and a disqualifying one for anaphylaxis. Use AI to flag likely allergen ambushes (soy sauce carries gluten, pesto carries nuts); never use it as the last word on what is safe.

The prompt kit: pin it to the fridge

The quiet argument of this post is that this, not the spectacular demo, is what AI is actually for. A machine that hands you back four days a year of deliberation and a few hundred dollars of un-wasted groceries, one boring Tuesday at a time, is doing more for your household than any demo reel. The division of labor is clean: it does the deciding, the arithmetic, and the list; you do the tasting, the label-checking, and the cooking, which was never the hard part anyway.

Start with the fridge photo tonight. It is the five-minute version of everything above, it costs nothing, and the worst case is a mediocre frittata. The best case is that the 5:40pm question quietly disappears from your week, and you get to spend those sixteen minutes eating.

Disclaimer: This article is general information about planning everyday meals, not medical or dietetic advice. AI-generated recipes and meal plans can contain errors, including allergen mistakes; always verify ingredients and labels yourself, and consult a doctor or registered dietitian before using AI suggestions for allergies, diabetes, or any medically supervised diet.