A scan is a hypothesis, not an answer
There is a version of this app that would be more satisfying to use. You point the camera, and a single confident name appears in large type. No percentage, no alternatives, no hedging.
We did not build that version, and this post is about why.
Identification is probabilistic
A plant identification model does not recognise a species the way you recognise a friend's face. It compares visual features against what it has seen before and ranks candidates by similarity. The output is a distribution, not a fact — and the top of that distribution can be wrong in ways that are entirely reasonable from the model's point of view.
A photo of a single leaf, taken in flat light, cropped tight, at the wrong time of year, may not contain the features that separate two species at all. The model will still return a best guess, because that is what it does. Confidence reflects how the photo compares to the training data, not whether the answer is true.
Look-alikes are the whole problem
Plenty of plants have a close visual twin, and in a number of cases one of the pair is toxic. Foraging literature is full of these pairings, and they are exactly the situations where a confident single answer causes harm. The failure mode is not "the app was a bit off" — it is someone eating the wrong thing.
The dangerous scan is not the one with 40% confidence. It is the one with 92% confidence, on the wrong species, shown without alternatives.
So the result screen shows the confidence and the runners-up it was weighed against. If two candidates are close, that is information you need, and hiding it to make the interface calmer would be a design choice with consequences.
What we do instead
- Show the confidence for the top match rather than presenting it as settled.
- Show the close alternatives, so a near-tie is visible instead of hidden.
- Keep the photo with the scan in your library, so an identification can be re-checked later against better information.
- Say on the result screen, every time, that a scan does not confirm a species.
The rule that does not bend
Never eat, brew, apply or otherwise consume a plant on the strength of a scan. Not at 99% confidence. Not when the picture looks obvious. Confirm with a qualified botanist or herbalist who can examine the actual plant — including the parts a photo usually misses, like roots, stem cross-sections and smell.
Herbora is built to make you better informed before that conversation, not to replace it. An app that made you feel certain enough to skip the check would be a worse product, however much better it felt to use.
Herbora is for education only. It is not medical advice, not a diagnosis, and not a substitute for a qualified professional. Many plants are toxic and some interact with prescribed medication.