What ChatGPT Gets Wrong About Cooking

A lot of home cooks have tried asking a large language model what to make for dinner. The results are impressive at first — it generates complete recipes, offers substitutions, explains techniques, and sounds authoritative. Then you actually cook from it and start noticing the gaps. The onions aren't caramelised in 5 minutes. The dish calls for chicken thighs and you only have breasts. The sauce needs adjusting but the AI can't taste it. The enthusiasm cools.

I've had the same experience, and I cook professionally. Here's an honest breakdown of where AI cooking advice works, where it doesn't, and why the gap matters for actual weeknight cooking.

What AI Does Well

It's worth starting here, because dismissing AI cooking tools entirely would miss what they're genuinely good at.

A large language model has absorbed an enormous amount of culinary content — recipes, technique explanations, food science, regional cuisines, flavour pairing principles, kitchen chemistry. Asked a specific culinary question ("what's the difference between braising and stewing?" or "what does cream of tartar do in a meringue?"), it gives a solid, accurate answer faster than a cookbook search.

It's also useful for substitution questions when you're stuck. "I don't have buttermilk — what can I use?" is exactly the kind of specific, answerable question where a language model performs well. It knows the functional role of buttermilk in a recipe (acidity + protein) and can suggest a reasonable substitute (milk + a teaspoon of vinegar).

And for someone who has never cooked a dish before and wants a structural recipe to follow, AI generates plausible instructions for most standard dishes. The instructions are usually correct in the aggregate even if the specifics need adjusting.

What AI Consistently Gets Wrong

It doesn't know what's in your kitchen

This is the fundamental problem. When you ask an AI "what should I make for dinner," it generates a recipe from its training data — not from the contents of your fridge and pantry. The recipe might call for fresh herbs you don't have, a cut of meat you didn't buy, or equipment you don't own. It's giving you a theoretical dinner rather than a real one.

You can partially solve this by describing your available ingredients — "I have chicken breasts, some wilting spinach, garlic, and pantry staples" — but this is an imprecise workaround. You're describing what you think you have, filtered through your own memory of what's in the fridge, rather than working from your actual inventory. The AI version of your kitchen is only as accurate as your description, which is almost always incomplete.

This is why photo-based ingredient recognition changes the equation. If you'd like to see how a tool actually built around your real pantry works, the best AI recipe generator comparison for 2026 covers the field honestly.

It cannot taste your food

Half of cooking is tasting and adjusting. Not following the recipe — deviating from it intelligently based on what's actually happening in the pan. The salt level needs to go up. The acid is missing. The dish needs another minute. These decisions require sensory feedback, and an AI has none.

When a recipe says "season to taste," that instruction is directing you to use judgment that an AI cannot exercise on your behalf. When it says "cook until golden brown," it means something specific that a camera-less language model cannot verify. The recipe text describes a cooking process; the actual cooking process is happening in your kitchen, in real time, with variables the recipe cannot know about.

Knowing how to make these in-the-moment adjustments is one of the core skills that separates home cooks who can improvise from those who need a strict recipe. The guide to how to taste food while cooking covers exactly this — the sensory adjustments that no AI can make for you.

It has overconfident timing

Recipe timing is one of the most unreliable elements in any recipe, and AI-generated recipes are no different. "Caramelise the onions for 10 minutes" appears in recipes with alarming frequency. Real caramelisation — onions going from raw to sweet, jammy, golden — takes 35–45 minutes minimum at the temperatures most home cooks use. Ten minutes produces soft, translucent onions, not caramelised ones. These are different things.

AI generates timing based on patterns in its training data, which includes many poorly tested recipes from sites that optimise for appealing claims ("ready in 20 minutes!") rather than accurate reporting. It inherits those inaccuracies. If you're a newer cook following AI timing estimates, you'll frequently undercook complex components and overcook simple ones.

It makes implicit equipment assumptions

AI-generated recipes assume a kitchen equipped with the tools needed for each technique. They don't ask whether you own a cast iron skillet before recommending one for a sear. They don't verify you have a blender before generating a pureed soup. They don't know whether your oven runs hot, whether your stove is gas or electric, or whether your "large pan" is a 10-inch or 14-inch surface.

All of these variables matter. A sear requires adequate surface area and heat retention; the wrong pan produces steaming rather than browning. An oven that runs 25 degrees hot will overcook anything with a precise temperature requirement. A recipe calibrated for gas will undercook on a slow-to-respond electric coil. AI recipes ignore all of this.

It doesn't remember last week

A recurring problem for people who use AI for meal planning: each conversation starts fresh. The AI doesn't know what you cooked Monday, what you have leftover, what proteins you bought but haven't used yet, or what produce is wilting in the crisper. It gives you meal ideas disconnected from your actual kitchen's current state.

A useful meal planning tool needs continuity. It needs to know that the chicken you mentioned on Sunday is still there on Thursday. Without that memory, every AI dinner suggestion is a fresh theoretical meal with a fresh theoretical shopping list, rather than a practical dinner built from what you have.

The Recipe Confidence Problem

One structural issue with AI cooking advice: language models are trained to produce confident, complete answers. They generate a full recipe with precise measurements, specific temperatures, and authoritative-sounding instructions whether or not those specifics are well-calibrated. A recipe that says "roast at 425°F for 22 minutes" sounds precise. But that precision may be pattern-matched from training data rather than validated by cooking.

In culinary education, a cook learns to treat recipe times and temperatures as starting points, not destinations. A recipe is a map, not a GPS route. The actual cooking requires adjustment based on what's happening. This is the skill that AI cooking advice cannot build in the people following it — because it produces recipes that sound definitive rather than teaching the adjustments that real cooking requires.

For the foundational skill of reading a recipe the right way, see how to read a recipe like a chef — specifically how professional cooks approach recipes as frameworks rather than scripts.

Where the Gap Matters Most for Home Cooks

The mismatch between AI cooking advice and real cooking experience is most painful at the point where the recipe breaks down. You're mid-cook, something has gone wrong, and you need to make a decision: add more liquid, turn down the heat, switch technique entirely. An AI cannot help you here because it doesn't know what's currently happening in your pan.

The home cooks who successfully use AI tools for cooking are those who have enough baseline culinary knowledge to critically evaluate what the AI produces, adjust the timing from experience, and improvise when the plan meets the stove. Newer cooks who follow AI recipes literally — treating the timing estimates as fact and the instructions as infallible — have a harder time.

The guide to why food tastes bland covers the real-time tasting and adjustment skills that no recipe can substitute for. And the NowCook vs. other tools comparison explains specifically how a purpose-built cooking tool differs from a general-purpose language model.

A More Useful Frame for AI in the Kitchen

AI cooking tools work best as assistants to a cook who already knows what they're doing — reference tools for specific questions, not replacement cooks. Ask them for ingredient information, technique explanations, and substitution options. Verify timing claims against your own experience. Treat generated recipes as rough drafts rather than final instructions.

What they can't do: see your actual ingredients, taste your actual food, adapt in real time, or maintain continuity across your cooking week. Those are the parts that require either your own skill or a purpose-built tool trained on your specific kitchen context.

NowCook approaches this differently. The photo scan reads your actual pantry, not your description of it. The meal suggestions are calibrated by chefs who cook for real kitchens, not trained purely on internet recipe data. And the suggestions are built around what's in front of you — the chicken that needs to be used, the half-used can of tomatoes in the fridge, the pantry staples that turn them into dinner. The use cases page shows exactly what this looks like in practice. A 14-day free trial with no credit card required gives you two weeks to test whether it closes the gaps that AI cooking advice leaves open.