How does an AI recipe generator work? A look behind the scenes

“Where does the recipe actually come from?”

A question we get a lot in support. The honest answer: it is generated in real time — not fished out of a database. That has advantages and disadvantages, and we'll explain both.

Here's a look behind the scenes at how a recipe comes about when you click “Generate recipe”.

Step 1: Your input is structured

When you type in “I have chicken, rice and peppers”, this is what happens on the server:

1. Your input is combined with your plan context: does the plan include an allergen filter? A language choice? AirFryer mode? A diet profile from your account?
2. If you have a family profile (Standard and up), the allergens and diets of the selected eaters are added.
3. We fetch the titles of your last 10 generated recipes so the AI doesn't repeat itself (the “pasta again?” safeguard).
4. From all of this, a structured prompt is built for OpenAI's GPT model:

Generate a recipe from the following ingredients: chicken, rice, peppers.
Language: English.
Strictly exclude these allergens: nuts, lactose.
Diet style: low-carb.
Cooking method: AirFryer.
Avoid these titles: Pasta Carbonara, Risotto Milanese, ...
Respond as a JSON object with the following fields: title, description, ingredients[],
steps[], cookTime, prepTime, calories, allergens[], tags[]...

Step 2: GPT generates the recipe

GPT (more precisely: gpt-5.6-luna, the current main model — as of September 2026) processes this prompt and returns structured JSON. This is not a database lookup — the AI “invents” a specific recipe from its trained knowledge of cooking, flavour combinations, quantities and cooking times.

What follows from this:

• No recipe is 100 % identical to another — even the same input yields slightly different quantities or steps
• The AI knows cooking times and temperatures (from training on millions of cooking instructions), but in individual cases a value can be off by ±5-10 %
• Allergen detection is very reliable — the AI knows exactly which ingredients contain gluten/lactose/nuts

Step 3: The hero image is generated

While the text recipe is being written, a second AI call runs in parallel to gpt-image-2 (OpenAI's latest image generation model of 2026). The prompt looks roughly like this:

Photorealistic food photography of [recipe title]: warm lighting,
centered composition, top-down or 45° angle, no text, professional
restaurant plating, garnished with fresh herbs.

Every recipe gets a unique image — no stock photos, no generic placeholders. The image shows exactly the dish described in the recipe.

Step 4: Validation + storage

On the server side, the following now runs:

• JSON schema check: are all fields present?
• Allergen sanity check: does the allergen list match the ingredients?
• Hero image → stored in Supabase Storage (EU region) for fast loading
• Quota usage is checked against your plan limit
• The recipe is saved to your history (for the “don't repeat” function)

Total time: roughly 30-60 seconds, depending on server load.

What the AI does well

✅ Finding a suitable recipe from any combination — even unusual ingredient combinations such as “sardines + avocado + rice”

✅ Multilingualism — you can type in English, Hungarian or French and get the recipe back in the language you typed in

✅ Diet adaptation — vegan, gluten-free, keto and low-carb are reliably respected

✅ Allergen detection — the AI knows very precisely which ingredients contain which allergens, and flags them

✅ Scaling — quantities for 1 or 20 people are adjusted correctly

What the AI is less good at

⚠️ Very regional specialities — e.g. Hungarian “halászlé” comes out OK, but not 100 % traditional. Generic international recipes work better.

⚠️ Very precise quantities — if you expect baking-recipe precision (±5 g of flour), the AI can be 10-15 % off. For main courses that doesn't matter; for sourdough bread it matters more.

⚠️ Current trending recipes — if a recipe went viral on TikTok last month, the AI probably can't do it. Search-based platforms are better for that.

⚠️ “Identical reproduction” — if you had a perfect curry yesterday and want it “exactly the same” today, the AI will make a new, similar one. In that case, save your recipe in the Account tab.

Why this suits family diets better than searching

In the classic search workflow (Chefkoch, Google):

1. You search for “vegan gluten-free lactose-free family”
2. The search results rarely tick all three boxes at once
3. You have to browse through 10-20 recipes by hand, checking which ingredients are OK
4. Even then, you often have to make substitutions

In the AI workflow:

1. You select the family members (each with their own diet profile)
2. The AI automatically receives vegan + gluten-free + lactose-free as constraints
3. One recipe is generated that satisfies all three at once
4. Image + quantities + steps included straight away

Time difference: 15-25 min. → 60 seconds.

Privacy & transparency

What the AI does with your data:

• OpenAI contract (as of May 2026): API inputs are NOT used for model training
• Third-country transfer to the USA: safeguarded by EU Standard Contractual Clauses (SCCs) and the EU-US Data Privacy Framework
• No personal data in the prompt: only your input + plan limits
• Every recipe is stored in your private history (Supabase EU)

For details, see § 7 of our privacy policy.

Try KI-Kochhilfe for free — 2 free recipes from your ingredients, no credit card required.