Your AI got the dandelion right in May. It will get the same plant wrong in August.
The AI didn’t learn the plant. It learned the flower.
Foragers already know this. Beginners don’t. Early success on common, in-bloom plants lulls a beginner into trusting the confident answer — and that confidence is exactly what makes the mistake lethal. The day your phone tells you, with the same 92% certainty, that the plant in your hand is harmless elderberry, it could just as easily be water hemlock — one of the most violently toxic plants in North America. No exaggeration.
Bringing awareness to a serious problem — and giving foragers more to go on than a confidence score.
We don’t claim to know. No software does. We exist to do a better job at educating foragers and surfacing cited information — so the person standing over an unfamiliar plant has more to go on than a single confident-looking answer that can’t reliably tell wild carrot from water hemlock.
Better information. Honest uncertainty. Real citations from established, trusted sources. Tried-and-true forager knowledge from the Foxfire tradition and the people who lived it long before any app existed. That’s the calling.
A working reference, not a one-photo guess.
Behind every identification is a library you can inspect — and it grows every week. The current PlantCraft knowledge base holds:
Every claim keeps its source type — historical public-domain book with a passage citation, or a timestamped creator record — so book knowledge and modern knowledge never get quietly blended together. Browse the manuscript library or the full bibliography any time.
Your own finds live in a private library with a seasonal foraging calendar and full photo history — so you can compare the same plant across months instead of trusting a single frame.
A real-world failure-rate audit of consumer plant-ID AI.
We submitted 89 plant photographs from a single forager's Zone 8b yard to the AI vision models foragers actually reach for — both the everyday ones inside the chatbots in your pocket and the premium frontier-lab models that are supposed to be the smartest available. We asked each of them to identify the species. We then checked every answer against the established reference sources we trust for ground truth — the same vetted foundation, backed by millions of real-world contributions, that PlantCraft is built on. That reference is our truth; the AI models were the test.
wrong
doesn't help
Two examples from the dataset
A photo a beginner might have asked any popular AI chatbot to identify as "Cannabis" was actually Hibiscus coccineus (Scarlet Rosemallow) — our reference sources identified it confidently. A photo that AI-only consensus labeled "Turmeric" was actually Canna indica (Canna Lily). Neither is a deadly example. They're typical examples. The deadly ones are why the study matters.
Why our numbers may look harsher than your experience
It’s true: today’s AI vision is reasonably accurate when the plant is common, in full bloom, photographed in clean light, and showing the part the model has seen most often — almost always the flower. Hand it a dandelion in May and it usually nails it. That’s the experience most casual users have, and it’s why people trust these tools too much.
Foraging doesn’t live in that controlled world. The same exact plant, photographed across its seasonal life cycle, will be identified correctly by AI in one month and confidently misidentified the next — once the flower is gone, the leaves change shape, the plant goes to seed, or it enters a juvenile or late-season form. We saw this pattern repeatedly in our own multi-season testing: same specimen, same camera, same forager — AI confident and correct in spring, AI confident and wrong in late summer.
AI vision is also reliably worse on uncommon, regional, juvenile, damaged, browsed, or hybridized plants, and worse again on the family-level confusions that matter most for safety — the wild parsleys that look like water hemlock, the puffballs that look like a death-cap button, the wild grapes that look like pokeweed. These are exactly the cases a beginner is most likely to encounter and least likely to recognize.
For most uses, that’s a quirk of the tool. For foraging, where a single confidently-wrong answer can end a life, it is not acceptable.
Foragers know this already. Beginners don't.
A model that's right about 1 in 4 times sounds bad on its face. It's worse than that once you realize most foragers identify several plants per outing. Across a 10-plant walk where each ID has a ~75% chance of being wrong, the chance of getting through clean is effectively zero. Across a season, near-certain disaster — if you trusted the AI.
For most decisions, an AI being wrong most of the time is annoying but not catastrophic — pick the wrong restaurant, you get a mediocre meal. For foraging, wrong once with a deadly look-alike is enough. Water hemlock looks like wild parsnip. The destroying angel looks like a young puffball. Pokeweed looks like wild grape. The mistake doesn't unwind. A single confidently-wrong answer can end a life.
The pattern is worse than the rate. AI rarely hedges — when wrong, it's confidently wrong, often with high-percentage certainty scores that read like authority. Different AI models are frequently wrong on different plants, and asking one chatbot then asking another and getting the same answer does not validate the answer. We saw cases where the entire AI panel agreed on the same wrong species at high confidence. Cross-checking AI against AI does not catch this failure mode — it can mask it.
Trusted sources for the identity. AI only for the jobs AI is actually good at.
We do not ask consumer AI to identify your plant. We vetted numerous well-established reference sources — backed by millions of real-world contributions — and built our own system on the ones we trust for ground truth. Then we put AI to the narrower jobs it’s genuinely good at: asking four of the best vision models whether they agree, flagging dangerous look-alikes, and surfacing other useful facts about the proposed ID. How we weigh and combine those signals is our own recipe.
Trusted reference sources establish the species
Instead of trusting a single chatbot's guess, we anchor the identity on established reference sources backed by millions of real-world contributions. In our study, this is the ground truth that consumer AI got wrong roughly three quarters of the time.
A four-model caution panel checks look-alikes and agreement
The caution panel takes your photo together with that identification and asks four independent AI models a narrower question where AI is genuinely useful: what dangerous look-alikes exist, could this plausibly be one — and do they agree it’s the identified species? Agreement raises confidence; every dissent stays visible.
RED / YELLOW / GREEN caution rating
RED: dangerous look-alikes flagged that visually resemble your photo. YELLOW: look-alikes exist but are distinguishable. GREEN: no dangerous look-alikes detected by our verifiers. Rating is independent of identification confidence.
Side-by-side comparison
Your photo next to a known-good reference for the identified species. You make the final visual call — we never hide the reference and ask you to trust us.
Out-of-specialty warnings
For fungi, mosses, lichens, and algae — categories outside our plant specialty — we tell you so, explicitly. We don’t pretend to identify them. We’ll tell you what it looks a lot like and require you to consult a specialist.
Hand-curated education guides — old, tried, and true
Tinctures, preservation, recipes, cultivation, grafting — drawn from the Foxfire series and other old tried-and-true forager sources. Hand-curated by foragers, never chat-bot output. Browse the guides →
We never claim to know. We show you what was found.
Most tools optimize for the user feeling sure. We optimized for the user being honest about uncertainty — because that’s the part that keeps foragers alive.
Every identification on PlantCraft AI shows you the same thing we see: the vetted source’s top species and confidence, the per-AI verifier agreement on look-alike risk, the dangerous look-alikes the AI flagged, the reference photo to visually compare against your own. We don’t round up. We don’t simplify. If two sources contradict each other, you see both.
We are not a substitute for a human expert. No software currently is. We are a sharper, more honest tool than what most foragers walk into the field with today — and we tell you exactly how sharp we are at every step. Use this alongside a qualified forager, mycologist, or botanist. Never instead of one.
The beta and the education guides are free. The identity we anchor on comes from established sources backed by millions of real-world contributions — a foundation decades in the making. We’re the layer that combines those trusted sources with focused AI checks and surfaces the result side-by-side with what generalist AI thought — more honestly than the chatbot in your pocket.
PlantCraft, plainly explained.
Does PlantCraft guarantee an identification?
No. No identification system is a guarantee. PlantCraft is educational decision support. Always consult a qualified human expert before consuming, applying, or otherwise using a wild plant.
How does PlantCraft identify a plant?
A specialist botanical reference identifies the species first. A four-model caution panel then takes your photo plus that identification and checks for dangerous look-alikes and model agreement, followed by a side-by-side reference photo you compare yourself. AI never overrides the primary identification.
Where does the knowledge come from?
The current library holds 636 species, 30 public-domain botanical books, 5,882 book-grounded claims with passage-level citations, 1,027 timestamped creator records, and 21 cited recipes. Historical book claims and creator claims keep distinct provenance.
Is PlantCraft free?
Yes. The current PlantCraft web app and the 16-guide practical library are a free open beta.
Try it on something growing where you are.
Free. No signup. Works on phones and desktops. Bring a photo of something growing in your yard, a trail, or a parking-lot crack.