AI Mushroom ID Gets It Wrong 35% of the Time. That Can Kill You.
A Polish engineer tested 16 AI models on 1,040 mushroom photos. Even the best, Gemini 3.8 Flash, was wrong 35% of the time on a first guess.
Piotr Migdał, a founding engineer at AI analytics firm Quesma, ran 1,040 mushroom photos through 16 AI models in a test published Wednesday. The best performer, Gemini 3.8 Flash, gave a correct first-guess identification just 65% of the time. The worst, Qwen3-27b, managed 13%. More alarming: some of the most lethal species were regularly tagged as safe to eat, meaning anyone relying on AI to go foraging is taking a genuine risk with their life.
What happened
| Detail | Value |
|---|---|
| Study published | Wednesday, September 2026 |
| Total photos tested | 1,040 |
| Species covered | 55 (safe and deadly) |
| Models tested | 16 |
| Best model (first-guess accuracy) | Gemini 3.8 Flash at 65% |
| Worst model (first-guess accuracy) | Qwen3-27b at 13% |
| Death cap misidentified as edible | 16% of the time |
| Fool’s funnel misidentified as edible | 48% of the time |
| Fatal dapperling misidentified as edible | 31% of the time |
Migdał built a dataset of Danish mushroom photos, supplemented it with information on species from Poland and other countries, and fed 20 photos per species through each model. He asked each model to name the most likely species plus four alternatives. The dataset covers 55 species across safe and deadly categories.
Even when models got a second chance, the results were modest. Gemini 3.8 Flash placed the correct species somewhere in its top five 85% of the time. If a user asks a single question and accepts the first answer, they face a one-in-three chance of a wrong call. Qwen3-27b placed correctly in its top five only 24% of the time.
The errors that actually matter
Most misidentifications are harmless in practice: calling a safe mushroom by the wrong safe name changes nothing. The dangerous errors are different in character. Migdał notes that a deadly webcap was identified as a chanterelle, which is exactly the mistake that kills experienced foragers in the real world.
On false positives (calling a poisonous mushroom edible), Qwen3-27b led at 36%. Qwen3-flash followed at 30%. Meta’s Muse-spark-1.2 scored a low 8% false positive rate, but Migdał flagged that as misleading: the model mostly just refused to guess, which happens to be the correct approach in a life-or-death context.
Why does this matter beyond mushrooms?
The mushroom case is an extreme illustration of a wider pattern with general-purpose AI models. These systems are trained to produce a confident, helpful answer. In domains where “I don’t know” is the right answer, that tendency becomes a liability.
For most business uses, a 35% error rate on a first guess is workable. You check the output, iterate, edit. Migdał puts it plainly: “AI slop might be annoying, but it is fixable with a few prompts, or a manual edit. It is much better than having to prompt ‘Do I need a liver transplant?'” The lesson applies well beyond foraging. Any workflow where an AI error has irreversible consequences needs a human checkpoint, not just a confident model response. This is a theme worth keeping in mind as businesses build AI integration into high-stakes processes.
Migdał also made a technical observation worth noting: a single photo is often not enough information to identify a mushroom species at all, so errors are baked in regardless of model quality. The models are being asked to do something that experts with the specimen in hand sometimes cannot do reliably.
Our take
This study is a clean, specific rebuttal to the idea that frontier AI models are approaching expert-level performance on visual identification tasks. Gemini 3.8 Flash is a capable model. A 65% first-guess accuracy rate on a task where being wrong can kill someone is not impressive: it is disqualifying.
The broader pattern Migdał surfaces is important for anyone deploying AI in a business context. Models that decline to answer when uncertain (as Muse-spark-1.2 did) outperform more confident models in high-stakes domains, even if their accuracy numbers look worse on a benchmark. Confidence calibration matters as much as raw accuracy, and most general-purpose models are not calibrated for situations where the cost of a wrong answer is asymmetric.
We cover AI model behavior regularly on the Lumien news desk, and this kind of adversarial, domain-specific testing is far more useful than lab benchmarks. If you are building an AI-assisted tool for any field where errors carry real risk, budget for this kind of targeted evaluation before you ship.
What to do about it
- Never use a general-purpose AI model as the sole decision-maker in any domain where errors are irreversible (medical, legal, safety-critical).
- Prefer models that express uncertainty or refuse low-confidence answers over models that always produce a confident guess.
- Commission domain-specific accuracy tests before deploying AI in a specialized field. Migdał’s dataset is publicly available for adaptation.
- Build human review steps into any AI workflow where a bad output cannot simply be edited and corrected after the fact.
A model that says “I’m not sure” is more useful than one that says the wrong thing with confidence.
Frequently asked questions
How accurate are AI models at identifying mushrooms from photos?
In a test of 16 models on 1,040 photos, the best performer, Gemini 3.8 Flash, got the correct species on a first guess 65% of the time. The worst, Qwen3-27b, was correct only 13% of the time.
Can AI tell if a mushroom is poisonous or safe to eat?
Not reliably. In Migdał's study, some deadly species were tagged as edible up to 48% of the time. Researchers strongly advise against relying on AI for this decision.
Which AI model performed best at mushroom identification?
Gemini 3.8 Flash scored highest with a 65% first-guess accuracy rate and placed the correct species in its top five 85% of the time across 55 tested species.
Why do AI models make dangerous mistakes identifying mushrooms?
General-purpose models are trained to produce confident answers. In a domain like mycology, where a single photo is often insufficient to distinguish species, that confidence leads to wrong calls on deadly lookalikes. Errors are also compounded because some toxic species closely resemble common edible ones.
