Why AI recipe apps invent ingredients, and how to spot it
Language models are built to produce complete-looking answers, so they fill gaps. What that means for recipes from videos, and the signs a recipe was guessed.
Ask an AI model for a recipe from a video and it will almost always give you a complete one. That’s the problem. The model is good at producing text that looks like a recipe, and a recipe “looks right” when every ingredient has an amount and every step has a time. When the video doesn’t say, the model is tempted to fill the blank with a typical value.
Where guesses creep in
- Amounts that were never said: “2 tbsp soy sauce” when the cook just poured.
- Cooking times that were never shown: “bake for 25 minutes” when the video cut straight to the result.
- Ingredients that weren’t there: salt, oil or garlic added because most recipes have them.
- The wrong dish: a recipe written from the title of a video that is really a restaurant review, a vlog or an advert.
How to spot a guessed recipe
- Every single ingredient has a neat amount, even for a “splash” style video.
- Times are suspiciously round and the video never showed a timer.
- The recipe reads like a textbook version of the dish rather than what this cook did.
How we handle it
We tell the model to write only what the video shows, says or captions, and to leave anything else as not given. After the read, a separate check asks what kind of video it was and whether parts of the recipe were inferred. Clear non-cooking videos are refused before you’re charged, and doubtful ones are saved with a visible warning instead of a confident-looking recipe.
No system is perfect, so the source video is always one tap away from the saved recipe.
