How We Use AI

How We Use AI

Exactly which parts of our data are generated by AI, which aren't, and how we mark them.

Last Updated: September 13, 2026 · Our disclosure under Article 50 of the EU AI Act

The Short Version

Foodashi recipes and recipe photography are generated by artificial intelligence. We do not hide this, we mark it in the data itself, and we tell you exactly which parts of a response are AI-generated and which are not.

We think that distinction matters. A lot of "AI recipe" products present everything as equally trustworthy. Ours isn't, and pretending otherwise would make the genuinely reliable parts of our data less useful to you. So this page is deliberately specific.

Not everything is AI-generated. Nutrition figures are calculated from per-ingredient values, and most of those values come from national food composition databases, not from a language model. Where no measured value exists, an estimate is used and recorded at lower confidence. See the breakdown below.

What Is and Isn't AI-Generated

Every recipe response carries a provenance object describing the origin of each part. Here is what those values mean:

Part of the responseOriginWhat that means
Title, description, steps AI-generated Written by a large language model from a structured prompt. No human authored or editorially reviewed the text.
Ingredient list AI + database The model proposes ingredients and quantities; each one is then matched against our canonical ingredient database using exact, fuzzy, and similarity matching. Names and gram weights you receive come from that database. When no suitable entry exists, a new ingredient entry can be created with AI-estimated nutrition, recorded at reduced confidence.
Nutrition figures Database + estimates Calculated by aggregating per-ingredient nutrient values. Most values come from national food composition databases; some are derived by rule (for example, sugars are zero when carbohydrate is zero), and some are AI-estimated where no measured value exists. Confidence is reported per recipe (see below).
Allergens & dietary tags Rules + AI check Derived from the resolved ingredients using deterministic rules. For ingredients the rules cannot classify, an AI classifier is consulted; it can add an allergen but cannot remove one the rules found. Allergen data is still a starting signal, never a guarantee: see the safety notice.
Recipe photography AI-generated Rendered by a text-to-image model. These are not photographs of real dishes and do not depict real people, places or events.
Meal plans AI-generated The plan structure is generated by a language model. Meals that match a catalogue recipe are substituted in; any that don't are model suggestions.
Barcode product data Third-party Sourced from Open Food Facts under the Open Database License (ODbL), not AI-generated. See our Data Sources page.

On nutrition confidence. Where an ingredient has an authoritative measured value (for example from USDA or another national food composition database) we record full confidence. Values derived by rule are recorded at slightly lower confidence, and AI estimates lower still. The recipe's aggregate (0 to 1) is served as nutrition_meta.confidence by the recipe nutrition endpoint, GET /api/recipes/{id}/nutrition; per-ingredient confidence is not part of the API response. We would rather show you a confidence score than round it up to a promise we can't keep.

How We Mark AI-Generated Content

Under Article 50(2) of the EU AI Act (Regulation 2024/1689), the outputs of a generative AI system must be marked in a machine-readable format and be detectable as artificially generated. We do this in two places, because a flag in a JSON response does not travel with an image file after you download it.

In API responses

Every response that carries recipe content includes an ai_generated flag. Full recipe responses additionally include the provenance breakdown:

// GET /api/v1/recipes/{id} { "title": "Boeuf Bourguignon", // ... "ai_generated": true, "provenance": { "text": "ai_generated", "ingredients": "ai_generated_db_resolved", "nutrition": "database_computed", "image": "ai_generated" } }

The ai_generated flag is present on full recipes, browse cards, comparison results and meal plans. It is never removed by field projection or by your plan tier: if you receive AI-generated content from us, you receive the marker with it. The database_computed value means the figures are calculated from our ingredient database, which includes the estimated values described above.

In image files

Recipe images carry the marking embedded inside the file, so it survives download and redistribution. We write the IPTC standard term for generative-AI provenance into the image metadata:

Iptc4xmpExt:DigitalSourceType = http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia

This is the same controlled-vocabulary term used across the industry, so our images are detectable by standard metadata tooling. We do not place visible watermarks on images. The AI Act requires machine-readable marking, not defaced pixels.

Images generated before 26 July 2026 predate our embedded-metadata implementation and do not carry the IPTC term inside the file. Every image is still marked as AI-generated by the ai_generated flag in the API response that serves it, and all newly generated images carry the embedded metadata. We're telling you the limit precisely rather than letting you assume a whole-back-catalogue guarantee we haven't yet earned.

If You Show Our Content to Your Users

If you display Foodashi content to your own users, the AI Act and other laws may place obligations on you, separately from ours as a provider. We give you the markers so you can meet them; we can't meet them on your behalf. You do not need to review each recipe by hand if you keep these safeguards:

  • Preserve the markers. Don't strip ai_generated or provenance from data you pass downstream, and don't re-encode our images in a way that discards their metadata.
  • Tell your users. Make it clear that recipe content and imagery are AI-generated. A label or a notice in your app is enough.
  • Show nutrition as estimates and allergens as indicative. Remind users to check ingredients and product labels, and never describe a recipe as allergy-safe on the basis of our data alone.
  • Don't present it as medical or clinical guidance. Our nutrition data is not a diagnostic tool and must not be marketed as one.
  • Let users report errors. Give people a way to tell you when something looks wrong.

These duties are also set out in the API License Agreement and the Terms of Service. Higher-risk uses, such as clinical nutrition or food labels, need more; see the Acceptable Use Policy.

Safety Notice

AI-generated recipes can be wrong. Language models can produce inaccurate, incomplete or fabricated content, including incorrect quantities, unsafe combinations, wrong cooking temperatures and missed allergens.

Foodashi is not medical, clinical, or nutritional advice. Do not use this data to make decisions about the treatment or management of any medical condition, and do not present it to your users as capable of doing so.

Allergen data is a starting signal, never a guarantee. If your application serves people with allergies, present our allergen indicators as indicative and make sure users check the actual products they use.

We run automated quality gates over every recipe before publication, including ingredient resolution, nutrition checks, and authenticity checks against known dish definitions, and we reject recipes that fail them. That reduces the error rate. It does not eliminate it, and we will not claim it does.

Models & Regulatory Position

Recipe text, classification, quality checks, and estimates of missing nutrient values are produced using Google's Gemini model family. Recipe imagery is generated using the FLUX model family via DeepInfra. We may change model providers as the field moves; this page will be updated when we do.

Risk classification. Under the EU AI Act, we assess Foodashi as a limited-risk AI system subject to the transparency obligations of Article 50. It is not a prohibited practice under Article 5, and it does not fall within the high-risk categories of Annex III. We do not offer Foodashi as a medical device, a diagnostic tool, or a clinical decision-support system, and the Acceptable Use Policy forbids relying on it as one.

No training on your data. We do not use your API queries or your application's data to train models.

Questions, or think we've got something wrong? Email [email protected]. If you're evaluating Foodashi for a regulated context and need something specific documented, ask: we would rather have that conversation before you build than after.