How Nosy's AI works
Last updated 14 October 2026
Nosy uses AI to read photos of your pet and to answer questions in the chat. This page explains what the AI does, what it never does, how we keep it safe, and where its limits are. It is written for owners, Apple's App Review, and the transparency duties of the EU AI Act.
1. What the AI does
For a photo or a symptom you describe, Nosy gives a working assessment:
- How soon a vet should see it — one of five levels: Clear, Observe, Vet soon, Vet within 24 h, Emergency.
- What is visible, in plain language.
- What it most likely is — the explanation the visible evidence best supports, with the alternatives ranked as most likely, possible or less likely. Confidence is stated in words, never as a percentage.
- How a vet would confirm it — the exam or test that settles the question, and what waiting would cost.
- Care steps that are safe at home while you arrange a visit.
It is framed as a working assessment, never as a confirmed diagnosis. Every report carries the line “AI working assessment — not a diagnosis, and it can be wrong. Check with your vet before acting on it.” The chat is labelled AI assistant and shows “Nosy is an AI, not a vet — it can make mistakes.” under the message box.
2. What it never does — enforced in code, not only in prompts
| Never | How |
|---|---|
| Doses or prescribes medication or supplements | A server-side filter removes amounts (mg, ml, units, and household amounts like “half a tablet”) from every report and chat reply and replaces them with “Dosing must come from your veterinarian.” Food amounts come from published formulas (RER = 70 × kg0.75 × life-stage factors), not from the AI. |
| Plays down an emergency | A fixed list of red flags — toxins such as chocolate, xylitol, grapes and raisins, antifreeze and rodent poison; seizures; breathing trouble; bloat; collapse; trauma; blood in vomit, stool or urine; urinary blockage — forces the Emergency level whatever the model said. These rules can raise urgency, never lower it. |
| Calls a blurry or unreadable photo “clear” | A photo the model can't assess never comes back as Clear. |
| Puts safety behind the paywall | The emergency, poison and “Something's off” screens are fixed guidance written from published veterinary sources (Merck Veterinary Manual, AAHA, ASPCA Animal Poison Control), not AI output. They are free, and a message describing an emergency is answered even when the chat allowance is used up. |
| Looks at people | Nosy only reads pets. If a photo or record is about a person, it says so and stops. |
3. Escalation first
- When the evidence is uncertain, the model is told to choose the more urgent level.
- From Vet soon upwards, the report always offers nearby vets with call and directions.
- For Emergency, Nosy tells you to contact a vet or an emergency clinic now; home care is never offered as a substitute.
4. Models and accountability
- Photo reports, lab results and the vet records you add under Documents use Claude Sonnet 5.5; the chat uses Claude Sonnet 4.6; Live guidance in the camera, food and product labels, documents photographed with the scanner, chat titles and care suggestions use Claude Haiku 4.5 — all made by Anthropic.
- Before any photo, message or record is sent to Anthropic, the app explains this and asks your permission. You can turn it off in Settings › AI processing. Details: privacy policy.
- Email addresses and phone numbers are removed from your messages and notes before they reach the AI model.
- Live guidance: while the camera is open for a close-up check, a small preview goes to the AI every few seconds to coach the framing (“get closer”, “more light”). Previews are not stored.
- Every report is logged with the model used, a hash of the prompt and the full output, so any assessment can be reviewed later.
- AI keys and calls live only on our server; the app can't bypass the safety filters.
5. The Health Index
The 0–100 Health Index on Home is computed by code, not by the AI, from what Nosy actually has on record: the urgency of recent checks for each body area, open concerns, lab values against species reference ranges, the weight trend, logged vitals and whether preventive care is up to date. It can only see what has been checked — an area never photographed counts as unknown, never as healthy — and it is not a vet exam. Breed and age risks are shown as “worth screening for”; they never move the number.
6. Limits, by kind of check
| Check | What it can reasonably assess | What it can't |
|---|---|---|
| Body, eyes, ears, teeth, gums, nose, paws, skin | What is visible: discharge, redness, swelling, lesions, tartar, coat and body-condition cues — and how soon a vet should look | Anything not visible, pain, internal illness. Similar-looking lesions can have very different causes; cytology, biopsy or an exam names them |
| Stool | Consistency, colour, visible blood, mucus, worm segments or foreign material | Most parasites and infections are microscopic and need a fecal test; “looks normal” doesn't rule them out |
| Food and product labels | Reading the label: adequacy statement, ingredients, calories | Testing the product; contamination or recalls |
| Vet records and lab reports | Reading the values and dates; lab values are compared with species reference ranges by code | Misreads are possible on poor photos — the original document stays authoritative; each lab has its own reference ranges |
| X-ray and imaging photos | An orientation read: region, view, positioning, exposure, and plainly visible structural findings, with a working explanation | A photo of a film loses detail; the interpreting vet's read is authoritative |
7. In one paragraph
Nosy gives owners an honest, committed working assessment of what is visible and how soon to involve a vet. It never doses or prescribes, it can only raise urgency for red flags, its safety screens are fixed and free, every output is labelled as AI that can be wrong, and owners decide — with their vet — what happens next.