AI models struggle to correctly identify mushrooms, posing serious safety risks

In a recent experiment, Piotr Migdał, a founding engineer at Quesma, benchmarked 16 AI models against a curated set of 55 mushroom species, including both safe and lethal varieties, by analyzing 1,040 photographs. The leading system, Gemini-3.8-flash, correctly named the species on its first suggestion in 65% of instances and placed the right answer among its top five guesses 85% of the time. By contrast, Qwen3.8-27b managed a mere 13% first-guess accuracy and frequently labeled poisonous mushrooms as edible, with false-positive rates up to 36%.

Notably, deadly species such as the death cap, fool’s funnel, and fatal dapperling were misidentified as edible in a significant share of cases. While most mistakes were harmless, the misclassifications that involved toxic fungi could have lethal consequences. Migdał emphasizes that reliable identification usually requires multiple observations and expert judgment, and he recommends that foragers avoid relying on AI recommendations alone.