大语言模型蘑菇识别测试：AI误判危险蘑菇风险凸显

研究人员使用包含2800种蘑菇的FungiTastic数据集对多个大语言模型进行了视觉识别基准测试，结果显示即使性能最优的模型也存在显著的安全隐患1。该测试选取了1040张蘑菇照片进行识别，其中每个物种包含20张照片1。

在测试的各大模型中，Gemini 3.6 Flash的识别准确率最高，达到64%，Gemini 3.7 Flash紧随其后准确率为61%1。Claude Fable 5.1的准确率为53%，成本为$12.881，而DeepSeek deepseek-v4-flash-vision的准确率仅为30%，但成本最低，仅需$2.051。Qwen 3.8 27B模型的表现更为令人担忧，超过三分之一的有毒蘑菇被错误分类为可食用物种1。

更令人担忧的是，即便是准确率最高的模型也存在约12%的错误率，将有毒蘑菇误认为可食用1。具体案例显示，死亡帽（Amanita phalloides）的正确识别率仅为46%，其中17%被误认为可食用的灰伞菌1。致命网帽（Calonarius splendens）也被27%的情况下误认为可食用的金针菇1。此外，Tricholoma equestre在某些国家被认为可食用，而在其他国家则致命，增加了识别判断的复杂性1。研究结果表明，不应仅依赖AI进行蘑菇安全判断1。

Researchers conducted a comprehensive benchmark test of multiple large language models' visual recognition capabilities using the FungiTastic dataset, which contains 2,800 mushroom species across 340,000 observations and 615,000 photographs 1. The study evaluated how well these models could identify 1,040 mushroom photographs, sampling 20 images per species 1.

The results revealed significant performance disparities among the tested models. Gemini 3.6 Flash achieved the highest accuracy rate at 64 percent, followed by Gemini 3.7 Flash at 61 percent, while Claude Fable 5.1 reached 53 percent accuracy at a cost of $12.88 1. In contrast, DeepSeek's deepseek-v4-flash-vision model demonstrated considerably lower performance at 30 percent accuracy, though it offered the lowest operational cost at $2.05 1. The Qwen 3.8 27B model performed poorly, with over one-third of poisonous mushrooms misclassified as edible species 1.

Even the most accurate models posed substantial safety concerns. The best-performing models maintained an approximately 12 percent error rate in categorizing toxic mushrooms as edible, a margin that could prove fatal in real-world applications 1. The death cap mushroom (Amanita phalloides) achieved only 46 percent correct identification, with 17 percent of misidentifications confusing it with the edible gray umbrella mushroom 1. The deadly webcap (Calonarius splendens) was misidentified as the edible enoki mushroom in 27 percent of cases 1. The researchers also noted that Tricholoma equestre presents additional complexity, as it is considered edible in some countries but is lethal in others 1. These findings underscore the critical dangers of relying solely on artificial intelligence for mushroom safety assessment 1.