Questions
What this reads, what it will not answer, and what happens to the picture you send.
The two things it refuses
It will not identify a mushroom and it will not identify a pill. Both are asked constantly and both have the same failure mode: a confident wrong answer that somebody acts on. Edible and deadly fungi are separated by features a photograph loses — the gills, the stem base, a spore print, the smell — and pill imprints are reused across countries while generics are made to look alike deliberately. The refusal names a real alternative rather than just declining.
How accurate is it
For printed text in a well-lit, filled frame, very. For handwriting, sometimes. For an object with no writing on it, only as good as whatever detection model is configured, which is why the description layer is opt-in and labelled with the model that produced it. It tells you when it could not read something rather than inventing plausible words, which is the failure mode that actually costs people.
Questions
Why will it not identify a mushroom?
Because getting it wrong kills people, and a photograph does not carry the features that separate an edible species from a deadly lookalike. Take it to a local mycological society, or use a regional field guide with a spore print.
Why will it not identify a pill?
Imprint codes are reused across countries and manufacturers, and generics are designed to look like each other. A pharmacist will identify it for free in minutes, and your country's medicines regulator publishes an imprint lookup.
Does it do reverse image search?
No. It reads what is in the picture; it does not go looking for where else that picture appears.
Does it recognise faces?
No, and it will not. It does not identify people from photographs.
Is my photo stored?
Not unless you are signed in and save it. Scan counting uses a daily-rotated one-way hash rather than your IP address, so usage stays measurable without becoming a location history.