{"slug": "knowledge-vs-wisdom-asking-ai-what-mushroom-is-that", "title": "Knowledge vs wisdom: asking AI “What mushroom is that?”", "summary": "Gemini 3.8 Flash leads in mushroom recognition accuracy, but GPT-6 Astra demonstrates greater wisdom by asking for more photos and issuing disclaimers, according to a test by an unnamed researcher using 1,040 photos of 55 edible and poisonous mushrooms from FungiTastic. The study found that when models were not forced to give restricted answers, correct guesses dropped for GPT-6 Astra from 54% to 32%, Claude Sonnet 5 from 25% to 17%, and Opus 5 from 36% to 28%, yet dire errors with no warning were rare, with only one case by Gemini 3.8 Flash misidentifying brown roll-rim as shiitake.", "body_md": "Gemini 3.8 Flash is the best at recognizing mushrooms, but GPT-6 Astra is wiser. Let’s see why.\n\nIn a previous post on mushroom identification with LLMs, while many models were really good at recognizing a species for a single photo, there was also a concerning number of deadly errors - poisonous species classified as edible.\nI forced the model to return a table of five Latin names of the most likely species, nothing else. I gave no room for mentioning uncertainty, or providing any disclaimers.\n\nHere is a recap of the models that are used in popular AI chats, plus GLM-5.3 Flash that is cost-effective, and the freshly released GPT-6 Astra.\nAs before, we use a subset of FungiTastic, with 1040 photos of 55 edible and poisonous mushrooms, found in Poland, as well as in other parts of Europe.\n\nGemini 3.8 Flash is in the lead.\nWhile Astra is a wonder when it comes to mathematics and puzzles (ARC-AGI-3, virtually or puzzles), it does not mean it is the best model at mushroom recognition. Sometimes intelligence alone is not enough.\n\nChat experience of poisonous mushroom identification\n\nBut what if we ask a regular question, posting a photo to a model, allowing a free-form answer? Do models communicate their uncertainty or risks involved? Does the answer come with a warning?\n\nHere are a few cherry-picked examples of poisonous mushrooms. We highlight text in red to mark potentially dangerous statements (e.g. misidentifications as an edible species) and in green to mark all things that reduce the risk (e.g. fair disclaimers).\n\nSo we see that even in the case of misidentification, often there are disclaimers. Let’s check how common they are, using a dataset of 360 photos containing 19 poisonous species.\n\nWhen a model is not forced to give an answer, its guess is no better - but it comes with a warning.\nTo start with, scores of correct guesses for restricted and free-form responses are similar in most models, with three major differences with GPT-6 Astra (from 54% to 32%), Claude Sonnet 5 (from 25% to 17%) and Opus 5 (from 36% to 28%).\n\nDire errors are rare, fortunately.\nOut of the same 360 photos of deadly and toxic mushrooms shown to 9 models, there was just a single case of an error with absolutely no warning - Gemini 3.8 Flash identified brown roll-rim as shiitake.\n\nThere are two more by Fable 5.1 (chestnut dapperling) but with a warning on identification, and two by Sonnet 5 (yellow knight), again with some notes. Though the latter, as we saw from our previous blog post, is a tricky case, as yellow knight sometimes is classified as deadly, but in some other places as edible.\n\nAn expert would always ask for more photos. And models?\n\nGPT-6 Astra usually asks for more photos (the underside, a spore print, the habitat, another photo), and always when its initial guess was wrong. Gemini 3.8 Flash (while efficient with restricted guesses) asks only in 32% of cases when it is right and, much more worryingly, in 81% of cases when it is wrong.\n\nConclusion\n\nA seasoned mycologist declines to decide on mushroom species, or edibility, from a single photo. We shouldn’t believe ourselves to be smarter just because some recent AI chat told us so.\n\nFortunately, as we tested, if models are incorrect, they mostly make rightful disclaimers and warnings.\nWe shouldn’t disregard these. And even if we get a persuasive result, we shouldn’t blindly follow AI in matters of health, life - not only in mushroom hunting, but in healthcare, psychiatry, legal or serious financial decisions.\n\nFrom the developer side - if you constrain a model so it gives guesses, with no room for warnings - it is your job to maintain that their results are communicated responsibly.\n\nAlso, for tasks heavily based on data, knowledge might be more important than intelligence, as you have seen with this Gemini 3.8 Flash vs GPT-6 Astra. At the same time, wisdom is worth its weight in gold - knowing a model’s own shortcoming and the risk involved.", "url": "https://wpnews.pro/news/knowledge-vs-wisdom-asking-ai-what-mushroom-is-that", "canonical_source": "https://quesma.com/blog/mushroom-llm-just-ask/", "published_at": "2026-09-06 10:00:00+00:00", "updated_at": "2026-09-09 14:42:41.034168+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-safety", "ai-research"], "entities": ["Gemini 3.8 Flash", "GPT-6 Astra", "Claude Sonnet 5", "Opus 5", "Fable 5.1", "GLM-5.3 Flash", "FungiTastic"], "alternates": {"html": "https://wpnews.pro/news/knowledge-vs-wisdom-asking-ai-what-mushroom-is-that", "markdown": "https://wpnews.pro/news/knowledge-vs-wisdom-asking-ai-what-mushroom-is-that.md", "text": "https://wpnews.pro/news/knowledge-vs-wisdom-asking-ai-what-mushroom-is-that.txt", "jsonld": "https://wpnews.pro/news/knowledge-vs-wisdom-asking-ai-what-mushroom-is-that.jsonld"}}