Cooking with AIGuide
Are ChatGPT recipes reliable? What to trust and what to check
Short answer
Partly. ChatGPT is fine for dinner ideas, everyday savoury cooking and quick substitutions, because those dishes forgive small errors. It is not reliable on exact quantities, cooking times or safety. Check meat with a thermometer against USDA temperatures, and use a tested recipe from a trusted source for baking, canning, fermenting and curing.
It's 6pm, you have chicken thighs, half a jar of harissa and a bag of spinach, and you ask ChatGPT what to make. Thirty seconds later you have a recipe that sounds great. The question is whether you can cook it as written. Mostly you can, if you know which parts to check. This page is written by the MySooChef team. Our app uses AI too, so we hold it to the same standard below.
Can you trust a ChatGPT recipe?
You can trust it for ideas and for forgiving, everyday savoury cooking. You can't trust it as written for exact quantities, timings or anything that touches food safety. A stir-fry with a bit too much soy sauce is still dinner. A cake with the wrong ratio of flour to fat isn't, and a jar of beans canned with a made-up processing time can make you seriously ill.
If you've searched Reddit for this, you've seen both sides: people who cook from ChatGPT every week and people who got a dense loaf or a pork chop that was raw in the middle. Those threads are anecdotes, not tests, but they go wrong in the same places we describe below.
Why do AI recipes go wrong?
AI recipes go wrong because nobody cooked them. A language model writes the most likely next words for "chicken harissa traybake". That usually lands close to a real recipe, but nothing checks that the result works in a pan.
- Nobody cooked it. A recipe from a person has usually been made at least once, often many times. A generated one has been made zero times.
- Averaged ratios. When NPR's Morning Edition taste-tested AI chocolate chip cookies against an America's Test Kitchen recipe in 2024, the human recipe won. One of the people involved said the AI version "looked like a really average cookie" [3]. Averaging is fine for a soup. It loses the specific choices that make a tested baking recipe work.
- Confident wrong times. A model states 12 minutes in the same tone whether 12 is right or not. Reports of undercooked meat from AI timings are mostly anecdotal, but the risk is obvious: a time is a guess, a thermometer is a measurement.
- Made-up substitutions. Ask for a swap and you'll always get one, even when the honest answer is "there isn't a good one". In a 2023 study in the journal Nutrition, researchers asked ChatGPT for 56 diets for people with food allergies. A dietitian checked them, and one "nut-free" diet included almond milk [2].
The extreme case is the Pak'nSave Savey Meal-bot in New Zealand. In 2023, users fed the supermarket's recipe bot water, bleach and ammonia, and it suggested an "Aromatic Water Mix" that would give off toxic chloramine gas [1]. People were feeding it non-food items on purpose, so it's not a typical result. Still, the bot didn't stop them.
Where do AI recipes do well?
AI does well where the cook can correct as they go. Ask it for five things to do with leftover rice, a weeknight curry from what's in your fridge, or a way to make a dish less spicy, and the answers are often useful. It can explain why a step exists. It can scale a recipe you give it, or boil a long method down to a short list. NPR's testers found the AI cookies decent, just dull [3]. Dull is fine on a Tuesday.
How much should you trust it for each kind of cooking?
| Task | Trust level | What to check |
|---|---|---|
| Weeknight stir-fry | YesHigh | Salt and sauce amounts. Taste as you go. |
| Baking | NoLow | Every ratio and the leavening. Compare with a tested recipe. |
| Scaling a recipe | PartlyMedium | The maths, line by line. Pan size and time do not scale evenly. |
| Substitutions | PartlyMedium | Whether the swap does the same job (fat, acid, structure). |
| Canning and preserving | NoDo not use | Use a tested recipe from USDA or NCHFP instead. |
| Allergy-safe cooking | NoLow | Every ingredient and label yourself. Hidden allergens get missed. |
| Meat doneness | PartlyTimes only | Internal temperature with a thermometer, not the stated time. |
What should you check before cooking an AI recipe?
Check these six things before you shop. It takes about two minutes.
Read it once, start to finish
Make sure every ingredient in the list is used in the method, and every ingredient in the method is in the list. Generated recipes drop things between the two.
Compare the ratios with one human recipe
Pull up a tested recipe for the same dish from a cookbook author or a test kitchen. If the AI version has twice the liquid or half the flour, trust the human one.
Treat every time as a range
Cook to the cue (golden, bubbling, tender to a knife) and use the time only as a rough guide. Oven temperatures vary and so does the size of your chicken thighs.
Use a thermometer for meat, fish and eggs
Ignore any "cook until the juices run clear" advice. These safe minimum internal temperatures come from University of Nebraska-Lincoln Extension's safe cooking temperatures page, which cites the USDA's chart [4]:
Safe minimum internal temperatures (UNL Extension, citing USDA)
- Beef, pork, veal, lamb steaks, chops, roasts
- 145°F (62.8°C), then rest at least 3 minutes
- Ground meats
- 160°F (71.1°C)
- All poultry, including ground poultry and stuffing
- 165°F (73.9°C)
- Fish and shellfish
- 145°F (62.8°C)
- Eggs
- 160°F (71.1°C)
- Casseroles
- 165°F (73.9°C)
Question every substitution
Baking soda and baking powder aren't interchangeable. Almond milk isn't nut-free. If the swap matters for texture, safety or an allergy, look it up somewhere that isn't the same chat.
Stop if it's canning, fermenting or curing
These rely on tested acidity, salt levels and processing times. The National Center for Home Food Preservation "strongly advises using only tested recipes from reliable sources for all home canning", and names the USDA, NCHFP and university extension services as those sources [5].
Is an AI-extracted recipe more reliable than a generated one?
Yes, because a person wrote and cooked the original. Generating a recipe means the AI invents it. Extracting one means the AI reads a recipe that already exists, like a creator's video or a food blog, and writes it out as a clean list of ingredients and steps. The dish has been made. The ratios are the author's, not an average.
MySooChef does the second kind. You share an Instagram, TikTok or YouTube video, a recipe website, a screenshot or a photo of a handwritten card, and the app writes out the ingredients and steps. It's still AI, and it has its own failure mode: misreading. A tablespoon can come out as a teaspoon, a handwritten "½" can become a "2", and an amount said quickly over music can be missed. Results depend on what the video shows and says. Servings, time and calories are estimates when the source never stated them. That's why the app tells you to check an import against the source before you cook. Please do, especially the first time you make something. Our guides to turning a YouTube video into a written recipe and digitizing handwritten family recipes show where to look.
How do you get ChatGPT to use recipes you trust?
Give it your own recipes to work from. An assistant that plans your week from dishes you've already cooked makes far fewer mistakes than one inventing dinners from scratch, because the quantities and times came from your kitchen. You can paste a recipe into the chat, or connect a recipe library. MySooChef has an MCP connector that lets Claude, ChatGPT and other MCP clients read your saved recipes, shopping list and meal plan. Our guide to using your recipes in Claude and ChatGPT walks through the setup. Even then, the assistant can still get a doubled quantity wrong, so read what it hands back.
Sources
- Supermarket AI Offers Recipe for Mom's Famous Mustard Gas, Gizmodo, 10 August 2023
- Robo-diets: can ChatGPT formulate safe dietary advice? (on Niszczota and Rybicka, Nutrition, 2023), News-Medical
- Can AI make better chocolate chip cookie recipes than humans? We taste tested 2, NPR Morning Edition, via KUNC, 6 August 2024
- Safe Cooking Temperatures (cites the USDA Safe Minimum Internal Temperature Chart), University of Nebraska-Lincoln Extension
- Safe home canning: the importance of using tested recipes, National Center for Home Food Preservation
Checked . How we research and write
Questions
Are ChatGPT recipes safe?
Most everyday savoury recipes from ChatGPT are safe to cook if you check them, but the tool is not a food-safety authority. Cook meat to USDA safe minimum internal temperatures with a thermometer (145°F for whole cuts of beef, pork, veal and lamb with a 3-minute rest, 160°F for ground meat, 165°F for poultry), check allergens yourself, and never use an AI recipe for canning.
Do AI recipes actually work?
Often, for forgiving dishes like stir-fries, soups and curries, where you can taste and adjust. They work less well when precision matters. In a 2024 NPR taste test, a human-developed America's Test Kitchen cookie recipe beat the AI-generated versions, which testers found decent but average.
Can ChatGPT write a baking recipe?
It can write one, but baking depends on exact ratios of flour, fat, sugar, liquid and leavening, and nobody has baked the generated version. Compare its quantities with a tested recipe for the same bake before you start, or simply use the tested recipe.
Is it safe to use AI for canning recipes?
No. Home canning needs tested acidity levels and processing times to prevent botulism. The National Center for Home Food Preservation strongly advises using only tested recipes from reliable sources such as the USDA, NCHFP and university extension services.
How do I make ChatGPT use my own recipes?
Paste a recipe into the chat, or connect a recipe library through an MCP connector. MySooChef has one that lets Claude, ChatGPT and other MCP clients read your saved recipes, shopping list and meal plan, so suggestions start from dishes you have already cooked. Still check any quantity it changes.
