AI models get mushroom ID dangerously wrong. Don't eat what they tell you to

Key points

• Piotr Migdał, a founding engineer at AI analytics firm Quesma, tested 16 AI models on 1,040 mushroom photos across 55 species in a study published in May 2025.
• The best-performing model, Google's Gemini 2.0 Flash, correctly identified a mushroom on its first guess only 65 percent of the time.
• The death cap, one of the world's most lethal mushrooms, was mistakenly identified as edible 16 percent of the time.
• Qwen 3.8-27b, a model from Chinese tech firm Alibaba, labelled poisonous mushrooms as edible 36 percent of the time.
• Migdał's verdict, shared with The Register, is blunt: never eat a mushroom because an AI said it was safe.

Piotr Migdał wanted to know whether a general-purpose AI chatbot could reliably tell a safe mushroom from a deadly one. He spent weeks finding out. The answer should stop you opening your camera app the next time you spot a fungus in the woods.

Migdał, a founding engineer at the AI firm Quesma, built a test set of 1,040 photographs covering 55 species, drawing on a Danish mushroom library and records of safe and deadly species from Poland and beyond. He ran every photo through 16 leading AI image-recognition models, asking each one to name the most likely species and its four next-best guesses.

How badly did the AI models actually do?

Badly enough to worry. The top performer, Google's Gemini 2.0 Flash, got the species right on its first answer just 65 percent of the time. That sounds passable until you remember the other 35 percent were wrong, and wrong answers in mushroom hunting can be fatal. Worth noting: we covered Gemini 2.0 Flash's release as part of Google's rapid model rollout on 2 September, where it was positioned as an everyday-tasks model. Identifying wild mushrooms clearly sits outside that lane.

At the bottom sat Qwen 3.8-27b, built by Alibaba. It named the correct species first only 13 percent of the time.

The errors weren't spread randomly across harmless mix-ups. Migdał found the models kept making the exact mistakes experienced foragers have been warned about for decades. A deadly webcap, which destroys the kidneys over several days, was called a chanterelle. The death cap, responsible for the majority of fatal mushroom poisonings worldwide, was called edible 16 percent of the time. The fool's funnel was misidentified as safe 48 percent of the time. The fatal dapperling, which closely resembles edible relatives, was called safe 31 percent of the time.

Should you use AI to check a mushroom you find?

No. Migdał is direct: "Do not eat a mushroom because AI told you it is safe."

One model, Meta's Muse-spark-1.2, posted a deceptively low false-positive rate of eight percent, meaning it rarely called a poisonous mushroom edible. The catch: it simply refused to guess most of the time. That's actually the right call. Refusing to answer beats confident fiction.

"I am surprised that the top generalist off-the-shelf models were able to identify most cases," Migdał told The Register. "Still, in this case, most is not enough, as errors could have dire consequences."

Experienced foragers follow a practical rule: stick to species with no dangerous look-alikes until you really know what you're doing. That rule applies even more firmly to AI. If a model is wrong one time in three on the straightforward cases, it has no business advising you on the hard ones.

Common questions

Could a specialist mushroom app do better than a general chatbot?

Possibly, though Migdał's study tested general AI models rather than dedicated foraging apps. A model trained specifically on fungi and regional species differences might improve accuracy, but no app has been independently verified as safe enough to stake your health on.

What should you actually do if you find a mushroom you want to eat?

Go with an experienced human forager who knows the local species, or join a local mycology (mushroom-study) club. Failing that, stick to varieties sold in shops, where the identification work is already done.