AI and Data Sources Notice
Where Falose uses AI, what it is allowed to do, and where the data comes from. This notice is published so the use of AI is disclosed plainly, and it forms part of the Terms of Use.
1. Where AI is used
Reading label photographs into numbers and text (a vision model transcribes the pack; a deterministic parser converts units to per 100). Reading product pages from the public web when a product is not in the catalogue (a Google-grounded search finds the pages; a second model transcribes only what those pages state). Drafting explainers about nutrients and additives, which a person publishes. The assistant. Proposing suggestions to the operator, which a person accepts or rejects.
AI never sets a score. The score is arithmetic on the record by the published formula. AI never publishes anything on its own: every record it produces is marked unverified until a person confirms it, and every draft it writes is published by a person.
2. The model and the provider
Google Gemini models on Google Cloud Vertex AI, in the region configured for our account, under Google Cloud's data-processing terms. Prompts contain product data, label photographs, and, for the assistant, your message and the profile fields you chose to fill. Google does not use these to train its models.
3. Web-read records
When you search for a product we do not have and no photograph is available, we may read it from the public web. The record shows the pages it came from, is marked "read from the web, unverified", carries the date, and is excluded from ranked lists until a person at Falose confirms it against a pack. Photographs harvested from those pages are candidates only until a person adopts one. The pages consulted are typically retailer listings and manufacturer sites; we read them the way a browser does, we do not bypass access controls, and we honour requests from site owners to stop.
4. Labelling of AI output
Assistant replies are labelled as AI-generated. Web-read records are labelled unverified and dated. Explainers show who reviewed them and when. Receipts and reels are renderings of a record; the record's source and date are printed on them.
5. Limits and errors
Models make mistakes: they can misread a decimal, a unit, or a row. We check each reading for physical plausibility (energy against macronutrients, per-pack values labelled as per-100, impossible densities) and quarantine what fails. Quarantined records are never scored in public. If you see an error, report it; the Content and Corrections Policy says what happens next.