Can AI Create Recipes? What Works, What to Check

By Alex Fahey, founder of Chop it. Last updated 5 August 2026.

Yes. AI can create a workable recipe in seconds, and for everyday cooking the results are better than most people expect. What it produces is a plausible first draft rather than a tested recipe, and knowing the difference is the whole skill of cooking with it.

Definition. An AI-created recipe is one generated by a large language model such as ChatGPT from a plain-language request, rather than written and tested by a cook. The model has read an enormous amount of cooking text and produces a recipe that follows the patterns of that text.

What AI recipe generation is good at

Familiar territory. Ask for a weeknight chicken curry, a tomato pasta or a traybake and the result draws on thousands of published versions of the same dish. The technique will be sound because the pattern is everywhere in what the model learned.

Constraints. This is where generation beats searching. A recipe database can filter by "vegetarian". A model can handle "vegetarian, one pan, under 30 minutes, no aubergine because my daughter refuses it, using the feta that needs eating". Every constraint you add narrows the answer instead of emptying the results page.

Adaptation. Take any recipe and ask for it dairy-free, cheaper, doubled or slower. The model rewrites rather than re-searches. In Attest's research, 75.9% of UK respondents said they would be comfortable with AI recipe recommendations, the highest comfort score of any AI food task, and this adaptability is a large part of why.

The two failure modes to check

Quantities. A generated recipe can state an amount that looks right and is not. The model is reproducing the shape of a recipe, not calculating one, so a soup might get double the sensible salt or a tenth of the stock. Read the quantities against your own judgement before you shop, the way you would sanity-check a stranger's recipe from a forum.

Timing and temperature. Generated times suit an average of every pan, oven and cut the model has read about, which is to say none in particular. Treat times as estimates and cook to the state of the food. For anything food-safety-critical, meat especially, check temperatures against a trusted reference such as the Food Standards Agency rather than the model.

Baking deserves separate caution. Baking is ratio-driven, and a plausible-looking but wrong flour-to-liquid ratio fails in a way a curry never will. For bread and cakes, prefer a tested recipe and use AI to adapt the flavourings around it.

How to test a generated recipe safely

1. Ask for the recipe with your real constraints in one message.
2. Read it as an editor. Do the quantities look sane? Is any step missing, such as when the onions go in?
3. Ask the model to double-check itself: "Are these quantities right for four portions?" It catches a surprising share of its own slips.
4. Cook it once as written, noting what you changed. It is your recipe after that.

Where the real problem starts

Generation is the easy half. The hard half is that a good generated recipe lives in a chat log, and a chat log buries them. If you cook something worth repeating, get it out of the conversation: our guide to saving recipes from any source covers the practical routes, and Chop it exists because this step should not be a copy-and-paste job. It runs inside ChatGPT and saves what you create as a structured recipe you can plan, shop and cook from.

For the wider picture of planning whole weeks this way, see ChatGPT meal planning: what works and what does not.

Frequently asked questions