AI mushroom identification is a life-threatening gamble, study finds

Using artificial intelligence to identify wild mushrooms is a life-threatening gamble. A new experiment reveals that even top-tier models frequently mistake lethal fungi for edible species, turning a foraging trip into a potential medical emergency.

Piotr Migdał, a founding engineer at Quesma, tested 16 different AI models using 1,040 photographs of 55 mushroom species. The results, reported by The Register, show a staggering lack of reliability across the board. Gemini-3.8-flash emerged as the most accurate model, yet it only correctly identified mushrooms on its first guess 65% of the time. While its accuracy rose to 85% when considering its top five guesses, the margin for error remains unacceptable for high-stakes biological identification.

At the other end of the spectrum, Qwen3.8-27b performed the worst, recording a first-guess accuracy of just 13% and only 24% accuracy within its top five suggestions. The risk was further compounded by high false-positive rates for poisonous mushrooms; Qwen3.8-27b and Qwen3.8-flash saw false-positive rates of 36% and 30%, respectively. In contrast, Meta's Muse-spark-1.2 maintained the lowest false-positive rate at 8%, though this was primarily because the model frequently declined to provide a guess entirely.

The cost of a hallucination

Foraging is a popular pastime in regions like Poland, but it requires deep expertise because many deadly species possess edible look-alikes. In the context of mushroom hunting, an AI "hallucination" is not a mere technical glitch but a fatal error. The study found that deadly species were frequently misidentified as safe: Fool's funnels were mistaken for edible 48% of the time, fatal dapperlings 31% of the time, and death caps 16% of the time.

Migdał noted that these errors are not random, pointing out that a deadly webcap was identified as a chanterelle—the exact mistake that often kills human foragers. "AI slop … might be annoying, but it is fixable with a few prompts, or a manual edit," Migdał said. "It is much better than having to prompt ‘Do I need a liver transplant?’"

Implications for foragers

These findings highlight a critical gap in the capabilities of generalist AI models. While these tools are increasingly used for a variety of identification tasks, they lack the precision required for biological safety. A 35% error rate for the best-performing model means that trusting an AI for a single identification could lead a user to ingest a fungus that causes total organ failure or death.

What to watch

As AI integration into mobile apps and cameras becomes more seamless, the temptation for amateur foragers to rely on automated identification will grow. However, this study suggests that until models can reliably distinguish between lethal and edible species with near-perfect accuracy, they remain a danger to public health. For now, the only safe method of identification remains expert human knowledge and traditional field guides.