A reader emailed me last week with a photo of a confused-looking seedling and the question every grower eventually asks: “Is this actually the variety I ordered?” She’d bought a Carolina Reaper start from a market in Newtown and her PlantNet app was insisting it was a sweet bell pepper. The leaves did look a bit broad. The plant was about 15cm tall. There’s basically no way to tell, at that stage, what you’re going to end up with.

So I decided to run an experiment. I have 43 varieties in the ground or in pots right now across my Sydney patch. Some are clearly labelled because I’m not a complete chaos agent. Others are mystery seedlings from cross-pollinated pods I saved two seasons back. Perfect test subjects.

The apps I tried

I ran four phone apps over a weekend: PlantNet, iNaturalist’s Seek, Google Lens, and PictureThis. None of them are chili specialists — they’re generalist plant ID tools — but they all use computer vision models trained on enormous image datasets. PlantNet is backed by French research institutions and uses a community-contributed image set, which I trust a bit more than the black-box commercial ones.

Results were a mixed bag. On mature plants with fruit, all four apps could reliably tell me “this is Capsicum annuum” or “this is Capsicum chinense.” That’s the species level — useful but not exciting. Anyone who’s grown chilis for a year can eyeball a chinense leaf versus an annuum leaf. Chinense leaves are usually wider, slightly puckered, and the plants have that distinctive lazy sprawl.

Where the apps fell apart: variety-level identification. A Bhut Jolokia and a 7 Pot Brain Strain are both chinense, both wrinkly orange-red pods, and the apps just shrugged. PictureThis confidently called my Aji Charapita (Capsicum chinense, tiny round yellow pods) a “wild bird pepper” — which is technically not wrong because Charapitas grow wild in Peruvian rainforest, but it’s also not specific enough to help anyone.

What AI is actually good at, right now

The honest answer: it’s good at flagging problems, not at making fine taxonomic calls. I tested Seek on a leaf with what looked like early blossom end rot and it correctly suggested calcium deficiency as a likely cause. I uploaded a photo of one of my chinense plants showing curled, distorted new growth and Google Lens linked me to broad mite damage references within a few seconds. That kind of triage is genuinely useful, especially for newer growers who haven’t seen these problems before.

The CSIRO has been working on more specialised agricultural AI for years — see their Data61 plant phenotyping work — and the gap between consumer apps and research-grade tools is still wide. Consumer apps are trained on what people photograph in gardens and on hikes. Research tools are trained on labelled disease datasets with controlled imaging. Different jobs.

I had a conversation a few months back with a mate who works in tech consulting — his shop, Team400, builds custom AI for businesses — and he made the point that off-the-shelf models almost never solve a specific industry problem. Growers want variety-level ID. Hobbyists want pest detection. Researchers want yield prediction. Each needs its own training data. Until somebody builds a chili-specific model trained on a couple of hundred thousand labelled variety images, the apps will keep telling you your Reaper is a Jalapeño.

Where I actually use AI in the garden

Three places, honestly:

1. Pest and disease triage. Snap, upload, get a second opinion. I still cross-check with the NSW DPI plant health pages before I do anything drastic.
2. Translating overseas growing forums. A lot of the best chinense growing knowledge lives on Brazilian, Hungarian and Trinidadian forums. Phone translation has gotten genuinely good, and I can follow a thread on Trinidad Scorpion cultivation in something close to the original tone.
3. Logging. I dictate notes into my phone after a watering session — “tray 3, two pods set on Yellow Bhut, aphids returning on Aji Lemon Drop” — and let the speech-to-text handle the rest. Saves me about twenty minutes a week.

What I don’t use it for: deciding when to harvest, deciding when to feed, deciding which seedling to cull. Those are pattern-recognition tasks I’ve built up over a decade. The app is the assistant, not the boss.

What to do with that mystery seedling

If you’ve got an unlabelled chili at the seedling stage, here’s my honest advice. Wait. Don’t trust the app at four leaves. Once flowers appear, look at the flower colour — chinense flowers tend to have a slight greenish tint, annuum flowers are stark white. Once pods set, shape and pod-pendant angle tell you almost everything. A pod hanging straight down on a wrinkly-leafed plant with greenish flowers is probably chinense. From there you can narrow further.

And if the app calls your Reaper a bell pepper, just plant it anyway. You’ll know in 90 days.

Marco