Mushroom hunting with LLMs: what can go wrong? (quesma.com)

A recent exploration into using large language models (LLMs) like ChatGPT for mushroom identification raises significant questions about the reliability of AI tools in potentially life-threatening scenarios. By leveraging the FungiTastic dataset, which contains over 615,000 photos of 2,800 mushroom species verified by experts, researchers aim to create an accessible identification process for foragers. While LLMs can assist in recognizing edibility and identifying various species based on user-uploaded photos, there is inherent risk involved; inaccuracies could lead to consuming poisonous or deadly mushrooms, as illustrated by the conflicting classifications of Tricholoma equestre, deemed edible in some regions but lethal in others. This development provides an intriguing glimpse into the capabilities of AI in real-world applications, particularly within the mushroom foraging community, where safe identification is crucial. However, implementing LLMs in this capacity employs a zero-shot approach, suggesting that while they can offer preliminary insights, they lack the rigorous validation that traditional machine learning models undergo. As enthusiasts navigate the complexities of mushroom hunting, reliance on AI could necessitate caution, emphasizing the importance of expert verification in augmenting AI's role in food safety and public health.