Mushroom hunting with LLMs: what can go wrong?

Researchers benchmarked 12 large language models on mushroom identification using the FungiTastic dataset, which contains 615 k photos of 2.8 k species with expert-verified labels. Gemini 3.7 Flash achieved the highest accuracy, outperforming Claude Fable 5 and GPT-5.6-Sol, while Qwen 3.8 27B misclassified over one-third of poisonous species as edible.