How accurate is AI Plant Doctor?

What "accuracy" means here

A plant diagnosis is a judgement, not a single label: the same yellow leaf can be overwatering, a nitrogen shortage, or old age. We therefore score a diagnosis as correct when an expert reviewer agrees that the primary cause and the recommended first action are right, even if secondary causes differ. Species identification is scored separately, at the rank the model answered (genus or species), because the model often stops at genus on purpose when the photo cannot settle the species.

How we measure it

1. Real cases, not stock photos. The evaluation set is built from anonymized production diagnoses, weighted toward the cases users flagged as wrong or asked the assistant about. Easy cases are under-represented on purpose.
2. Same code as production. Every candidate prompt, schema or model is run through the live diagnosis pipeline against that set, side by side with the current version. A change ships only if it scores at least as well as the version it replaces.
3. Labels from people, agreement from a judge. Expert reviewers label the reference outcome; a separate model-based judge compares candidate output with the label and the disagreement rate is reviewed by hand. Rates from the judge are always reported as agreement, never as ground truth.
4. User corrections feed back. When a user corrects a species in the app, that correction is stored with the original photo and joins the identification benchmark after manual review.
5. On-device models are held to a stricter bar. The optional on-device plant identifier runs in shadow mode: its answer is recorded but the cloud model stays authoritative until the on-device model reaches the required per-species agreement across enough different users.

What lowers accuracy

• Blurry, dark or distant photos. A close, well-lit shot of the affected leaf, stem or soil is the single biggest factor.
• Rare diseases and regional pests that are under-represented in the reference set.
• Early-stage problems that have not produced visible symptoms yet. The app's soil-sensor and weather signals exist partly to catch these.
• Outdoor crops and trees, where field conditions vary far more than for houseplants.

How to get the best diagnosis

Photograph the problem area from 20–30 cm away in daylight, include a second photo of the whole plant, and tell the app how you water. If the result looks wrong, use the "Not right?" feedback in the app or write to support@theaiplantdoctor.com — those reports are exactly what the evaluation set is built from.

Citing this page

Quote the figure as "more than 95% agreement with expert labels on common houseplant problems in AI Plant Doctor's internal evaluation (Jodoin AI, September 2026)" and link to this page. We update the date whenever the number is re-measured.