Image-based diagnosis works well for common, visually distinct problems — nutrient deficiencies, common fungal issues, pest damage — and much less well for anything that looks similar across several causes. Treat it as a second opinion, not a diagnosis.
What the models are actually trained on
Most consumer plant-diagnosis tools are built on image classifiers trained against large labeled datasets of common crop and houseplant problems. That’s exactly why they’re strong on the well-photographed, well-documented issues — powdery mildew, nitrogen deficiency, spider mite damage — and noticeably weaker on anything regional, rare, or that presents differently across cultivars. A model is only as good as the examples it’s seen.
Where it earns its keep
The clearest win is speed of triage: instead of paging through a plant disease guide trying to match symptoms, a photo narrows the possibilities to two or three likely causes in seconds. For a home grower, that’s often enough to know whether to act immediately or wait and watch. For a commercial operation scanning many plants, it’s a genuine labor saver even at imperfect accuracy, because it catches problems a rushed human eye would miss.
The failure mode worth knowing
The riskiest cases are ones where multiple issues look almost identical from a photo — overwatering and root rot present similarly to underwatering in early stages, for instance. A confident wrong answer is worse than no answer, which is why the tools worth using surface their confidence level and alternative diagnoses rather than a single verdict.