According to a report published on Hacker News, experiments were conducted regarding the ability of LLMs to identify mushrooms.
The experiments utilized the FungiTastic dataset created by researchers in the Czech Republic. This dataset features labels based on expert verification and DNA sequencing.
The subjects of identification included species sold as edible in Poland and a list of lethal poisonous mushrooms documented on Wikipedia.
As a result of the verification, Gemini 3.6, 3.7, and 3.8 Flash demonstrated extremely high identification accuracy. These models are reported to possess superior recognition capabilities compared to state-of-the-art models such as Claude Fable 5 and GPT-5.6-Sol.
On the other hand, the results also showed that identification errors carry significant risks. Misidentifications occurred where poisonous mushrooms were mistaken for edible ones, or vice versa. In particular, cases were cited where highly toxic webcaps were misidentified as chanterelles.
While the details of misidentification vary by model, a 19% error rate was recorded for Muse Spark 1.3.
Source: Mushroom hunting with LLMs: what can go wrong? (Hacker News Frontpage, 2026-09-03)