A software engineer at Quesma has found that leading AI models are unreliable and potentially dangerous for identifying wild mushrooms, with even the best-performing model correct only 65% of the time on a first guess. The finding matters because misidentifying toxic species as edible can be fatal, and the errors were not random but included several models confusing deadly mushrooms, such as the death cap and deadly webcap, with safe, popular-to-forage look-alikes.

Piotr Migdał tested 16 AI models, including versions of Gemini and Qwen, against 1,040 photos covering 55 mushroom species from a Danish dataset plus known dangerous species. Gemini-3.8-flash performed best, correct 65% of the time on the first guess and 85% within its top five, while Qwen3.8-27b performed worst, correct only 13% of the time initially. Dangerous misidentifications included the death cap being called edible 16% of the time and the fool's funnel 48% of the time, with Qwen3.8-27b wrongly labelling poisonous mushrooms as edible in 36% of cases; Migdał concluded that a single photo is often insufficient for safe identification and that experienced human judgement remains essential.