# Gemini 3.7 Flash Called 12% of Poisonous Mushrooms Edible **Summary:** In Quesma's 1,040-photo test, Gemini 3.7 Flash called 12% of poisonous mushrooms edible. I would use AI to learn the features, never to decide what to eat. - Canonical: https://markhuang.ai/news/gemini-mushroom-12-percent-edible-errors - Language: en - Author: [Mark Huang](https://markhuang.ai/about) - Published: 2026-09-02 - Section: News - Tags: Gemini 3.7 Flash, Multimodal AI, Mushroom Identification, AI Safety, Computer Vision - Source: [Quesma](https://quesma.com/blog/mushroom-llm-vision/) - License: https://creativecommons.org/licenses/by-nc/4.0/ --- ![A person photographs wild mushrooms while several conflicting mushroom outlines branch from the phone](https://cdn.markhuang.ai/news/gemini-mushroom-12-percent-edible-errors/hero.webp) *One photograph can produce several plausible names. That uncertainty is useful for learning and dangerous when lunch depends on the answer.* [Quesma tested 16 vision models on 1,040 mushroom photos](https://quesma.com/blog/mushroom-llm-vision/) and asked each one for its five most likely species. Gemini 3.7 Flash put the correct species first 61% of the time. More worrying, it named an edible species for 12% of the poisonous mushrooms. That result gives me a clean boundary. I would use a vision model to suggest names and explain which features to inspect while I learn about local fungi. I would never let its answer decide whether a wild mushroom goes into a pan. The model can help long before it is safe enough to make that decision. The stakes show up in hospital data. A [CDC analysis of U.S. cases](https://www.cdc.gov/mmwr/volumes/70/wr/mm7010a1.htm) estimated 1,328 emergency-department visits and 100 hospitalizations associated with accidental poisonous mushroom ingestion in 2016. Among 556 patients in separate claims data from 2016 through 2018, 8.6% had a serious adverse outcome. The CDC's advice is blunt: do not consume wild mushrooms unless an expert has identified them. ## The ranking stops at the safety column Quesma's benchmark used 55 species selected from a Polish list of mushrooms permitted for sale and a list of deadly species. It sampled up to 20 ordinary photos per species from FungiTastic, mostly from the dataset's 2023 test split. The prompt required five Latin binomials and nothing else. The [code, photo list, raw model answers, and scoring scripts are public](https://github.com/stared/mushroom-hunting-llm-bench). Gemini 3.6 Flash led the first-choice result at 64% and included the correct species in its top five 85% of the time. Its poisonous-to-edible error rate was 11%. Gemini 3.7 Flash reached 82% within five guesses, but its 12% dangerous-error rate was slightly worse. Put safety beside accuracy and the leaderboard becomes much less comforting. The mistakes also landed on exactly the lookalikes that matter. Across all model answers in Quesma's species-level analysis, death-cap photos were assigned an edible species 17% of the time. Fool's funnel photos were assigned an edible species 49% of the time. A model does not need to fail often to be unusable as a safety gate. It only needs to sound convincing on the wrong specimen. This is why I care more about the direction of an error than the overall score. I made a similar argument about [model build-offs and failure rates](/news/model-build-offs-need-failure-rates), but mushroom identification makes the asymmetry impossible to ignore. Calling an edible mushroom poisonous wastes a find. Calling a poisonous one edible can send someone to the hospital. ## One photo is not the whole specimen The benchmark is deliberately close to how people use a phone: take one normal photo and ask for a name. That makes it a useful product test, but it is not a complete test of field identification. Location, nearby trees, substrate, smell, texture, bruising, spores, and the underside of the cap can all matter. [iNaturalist's mushroom documentation guide](https://www.inaturalist.org/posts/3531-documenting-mushrooms) recommends at least three photos: the specimen in its habitat, the cap, and the stem and gills. It also notes that photographs alone sometimes cannot settle the species. Quesma tried adding a Denmark location hint and reported that it did not improve these model results. The underlying data are stronger than a random web-image collection. [FungiTastic describes itself](https://bohemianvra.github.io/FungiTastic/) as an expert-verified dataset with roughly 350,000 observations, more than 600,000 images, and over 5,000 species. Its records can include geospatial and climate data, captions, and other views. Quesma's test intentionally removes most of that context. It measures what happens when a general vision model receives the thin evidence a casual user is likely to provide. ## A species guess is not permission to eat Photo-identification software had this problem before the latest multimodal models. In a [2023 study of three mushroom apps](https://pubmed.ncbi.nlm.nih.gov/36794335/), researchers tested 78 specimens whose identities were confirmed by an expert mycologist. The best app correctly identified 49% of specimens. The authors concluded that the apps were not reliable enough, on their own, to exclude a potentially poisonous exposure. Quesma's results are not directly comparable with that study. The apps, specimens, prompts, and scoring methods differ. The narrower lesson is enough: better species suggestions do not convert a camera into a food-safety instrument. Put a polished name, confidence score, and edibility label on one card, and users may read three separate claims as one answer. > **Info:** > > I would remove direct "safe to eat" answers from a general AI identification flow. The product can ask for more views, show several candidates, explain the missing evidence, and direct suspected ingestion to a poison center or emergency service. In the United States, [Poison Control says to use its online tool or call 1-800-222-1222](https://www.poison.org/contact-us) rather than wait for symptoms. ## I would use it as a field guide There is still a good use for this capability. A model can turn a walk into a set of questions: Which parts should I photograph? What features separate these candidates? Which local expert or field guide should I consult next? A five-name shortlist can make unfamiliar biodiversity easier to explore. The interface should leave the uncertainty visible. It needs to say when one photograph is inadequate, keep identification separate from edibility, and make escalation obvious when someone may have eaten the specimen. A refusal buried beneath a confident species card would miss the point. Gemini 3.7 Flash did well enough in Quesma's test to be an interesting learning aid. The same test produced a 12% poisonous-to-edible error rate, which settles the food decision for me. I would ask the model what to look at next. I would not ask it what to cook.