Piotr Migdał, a founding engineer at AI analytics firm Quesma, tested how well leading AI models identify mushrooms. He ran 1,040 photos through 16 models, covering 55 species from safe and deadly datasets, cross-checked against a Danish mushroom library. Each model had to name the most likely species plus four alternatives per photo.

The best performer, Gemini-3.8-flash, got its first guess right only 65 percent of the time and reached 85 percent within its top five. Qwen3.8-27b trailed at 13 percent on first guess and 24 percent across five.

The errors were not random. Models called a deadly webcap a chanterelle, the same mix-up that kills human foragers. Death caps were judged edible 16 percent of the time, fool's funnel 48 percent, fatal dapperling 31 percent. Qwen3.8-27b labeled poisonous mushrooms edible 36 percent of the time, Qwen3.8-flash 30 percent. Meta's Muse-spark-1.2 posted the lowest false-positive rate at eight percent, largely because it often declined to guess.

Migdał, who eats mushrooms but only forages with people he trusts, notes that a single photo is often not enough even for experts. His conclusion is blunt: do not eat a mushroom because AI told you it is safe. The benchmark dataset is on GitHub for anyone wanting to test models against their local flora.