The short answer: our on-device model gets the species exactly right 56.5% of the time. That number is worse than the ones our competitors advertise, and we are publishing it anyway, because the alternative is to quote a figure nobody can check.

The longer answer is more interesting, and it is the reason the app is still useful. The model does not have to answer. It only commits when it is confident, and we measured what happens when it does.

The numbersn = 38,299

Top-1 accuracy

56.5%

exact species, every image

It answers

32.6%

of the time, at 70.0% confidence

And is right

90.3%

when it does answer

That third number is the one a user actually experiences. The model declines most of the time, and hands the photograph to Pl@ntNet and Plant.id instead. When it does commit on its own, it is right 90.3% of the time.

A model that answered everything at 56.5% would be worse than useless, because you could not tell the good answers from the bad ones. Knowing when to shut up is most of the value.

Methodreproduce it

Model	aiy_plants_V1 (TFLite, 2,102 classes, uint8)
Preprocessing	centre-crop, bilinear 224, uint8 RGB, TTA over horizontal mirror
Ground truth	iNaturalist research-grade community ID
Images	38,299 across 2,040 species, licensed CC0, CC-BY, CC-BY-SA
Threshold	70.0%. Below it, the app defers to Pl@ntNet and Plant.id.

The evaluation images are openly licensed and their photographers are credited on the species pages that use them. The full result set is downloadable below. The preprocessing replicates the app exactly, including the test-time mirror averaging, because a benchmark of a model we do not ship would tell you nothing.

What it confuses 1,741 pairs

These are not guesses about what people might mix up. Each row is a pair of species that this model genuinely mistook for one another, and how often.

Actually	Called it	Rate
Solanum americanum	Solanum nigrum	16/20
Toxicodendron rydbergii	Toxicodendron radicans	16/17
Dichelostemma congestum	Dichelostemma capitatum	16/19
Nymphaea alba	Nymphaea odorata	14/20
Pinus monticola	Pinus strobus	13/20
Solidago altissima	Solidago canadensis	13/19
Typha angustifolia	Typha latifolia	12/20
Opuntia engelmannii	Opuntia littoralis	12/19
Opuntia phaeacantha	Opuntia littoralis	12/20
Impatiens noli-tangere	Impatiens pallida	12/19
Zephyranthes drummondii	Zephyranthes chlorosolen	12/20
Ulmus rubra	Ulmus americana	12/20
Callirhoe pedata	Callirhoe involucrata	12/20
Sisyrinchium montanum	Sisyrinchium bellum	11/20
Conopholis alpina	Conopholis americana	11/17
Calystegia macrostegia	Convolvulus arvensis	11/20
Verbascum virgatum	Verbascum blattaria	11/19
Conocephalum conicum	Lunularia cruciata	11/19
Tsuga heterophylla	Tsuga canadensis	11/19
Viola adunca	Viola sororia	10/17

Hardest to identifyby our own model

Species this model gets wrong most often. If you are photographing one of these, take the picture and then check it against a key.

Why publish thisthe obvious question

Because a number you cannot check is not a number. PictureThis advertises 98% accuracy and publishes no methodology, no dataset, and no error analysis. We have no way to evaluate that claim, and neither do you.

Fine-grained plant identification from a single photograph is a genuinely hard problem. A juvenile Monstera deliciosa and a pothos are close to indistinguishable without a petiole. Two cottonwoods differ by leaf margin. Anyone claiming near-perfect accuracy across tens of thousands of species is either measuring something much easier than what you are doing, or not measuring at all.

So we would rather tell you what the model actually does, show you where it fails, and let the app hand the hard cases to engines that do better. That is also why the app reports its confidence instead of hiding it.