Nixo AI Policy

How AI is used

Nixo uses Google Gemini for one job: turning a photograph of a printed ingredient panel into text. It runs when a barcode is not in the open food catalogue and you photograph the ingredient list yourself. The model runs on our servers, never on your device, and our access key is never included in the app. The photograph is sent for the sole purpose of producing that one verdict, and it is dropped once the verdict is on screen.

The verdict is not the model's opinion. The rulebook is a closed, enumerated list of named ingredients that ships inside the app and reads offline, and a verdict is that list compared against the ingredients that were read. Nothing is inferred. This is why every card names the rule it came from, why you can read all four rule families in full on one screen, and why an ingredient we do not carry passes without comment rather than being judged.

The model is given a narrow brief. It is instructed to report only the ingredient list printed on the packet in front of it. It is instructed not to judge the food, not to say whether anything is healthy, unhealthy, safe, toxic or harmful, not to produce a score, a grade or a percentage, not to name or favour a brand, and not to give medical, nutritional, dietary, psychological, legal or financial advice, regardless of what it is asked.

Model output reaches you in one other place, and it is worth naming. On a scan resolved by reading a printed panel, the product's category is inferred from that reading. The category is what decides which product the better in this aisle row names, so an inference sits behind a surface most people would assume is a database lookup. The ranking itself is not a model decision: candidates are filtered by category, by having a complete ingredient list, and by being clean under exactly the rule families applied to your card, and what survives is ordered by the fewest printed ingredients, ties broken alphabetically. The same product in the same rulebook state returns the same alternative every time.

Informational only

Everything Nixo produces is informational. It reports that a named ingredient is printed on a packet and that this breaks one of our published rules. It is a starting point for your own judgement, not professional advice, and it is not a statement about your health. If a result matters to a decision with real consequences, check the packet itself and, where it matters, a person who knows your situation.

You decide

You stay in charge. Nixo reports, and you decide what to put in the trolley. We do not use AI to take actions on your behalf, and there is no step in the app where a model's output is applied without you seeing it: the ingredient is named on the card, the rule it broke is named beside it, and the rule itself is one tap away in full.

Automated decision making

A verdict is about a food, not about you. It does not produce legal effects or similarly significantly affect you, and we do not use it to profile you beyond producing the verdict itself. Nothing in the app scores you for a purpose outside the app, and there is no personal restriction list, so no claim Nixo makes is ever a claim about a person.

Limitations

Reading a printed ingredient panel from a photograph can be wrong. Glare, blur, an angle, a curved packet and small type all reduce accuracy, and a list that came back cut off at either end is treated as unreadable rather than as clean, because a verdict from half a list is the one wrong answer this product must not give. Two runs on the same photograph can differ.

The barcode path has its own limits. The open food catalogue does not hold every product on every shelf, its records can be incomplete or out of date, and a product can be reformulated after we read it. A verdict already in your history is never quietly refreshed, so an old card can be wrong about a product that has changed. Delete the row and scan it again to get a fresh read.

Our rules are a closed list, which is what makes them checkable, and it is also a limit: an ingredient under a name we do not carry passes without comment. We would rather say all of this plainly than let a confident card imply a reliability the method does not have.

Privacy

The photograph you take of an ingredient panel is sent to the model provider only to read that one list. It is not used to train their systems, and we do not send your email address or account identifiers with it. It is dropped once the verdict it produced is on screen, and nothing in your history holds an image. A scan that resolves on the barcode path sends no image anywhere. Our Privacy Policy describes what we store afterwards and for how long.

No exaggerated claims

We do not claim that Nixo is an expert, that it replaces a professional, or that its results are guaranteed. It reads an ingredient list and checks it against rules we publish and you can argue with. It does not audit a supply chain, verify a certification, count a calorie or a macro, or know anything about your body. It is useful in proportion to how honestly it is described.

Changes over time

We change the prompts and the models behind Nixo as better ones become available, which means a panel can be read slightly differently over time. The rulebook is different: it ships with the app rather than being fetched, it carries a last changed date on its own screen, and it cannot change under you between one scan and the next. Where a change materially alters what the app tells you, we will say so in the app rather than let it change silently.