From Prediction to Dialogue: Human-AI Collaboration for Nature

We propose a novel human–AI collaboration paradigm that moves beyond black-box species prediction toward interactive, dialogue-based identification. The approach connects domain experts, who provide taxonomic knowledge through diagnostic traits and interpretable annotations, with AI systems that use this knowledge to guide citizen scientists through the identification process. Using a Research through Design methodology, we iteratively design, fine-tune, deploy, and evaluate interactive AI tools, including Large Language Models (LLMs) and Vision-Language Models (VLMs), together with experts and citizen scientists. Rather than simply returning a species prediction, the system explains relevant morphological characteristics, asks diagnostic questions, and adapts its guidance to the user. In doing so, it supports the development of identification expertise and a deeper understanding of natural history. At the same time, the resulting richer observations and annotations can support the development of more accurate, interpretable, and robust AI models, creating a virtuous cycle between human learning and machine learning that benefits biodiversity research, natural history collections, and citizen science communities. Our preliminary results with AI-assisted fossil identification demonstrate the potential of this approach, showing complementary benefits of conversational guidance and structured interaction and motivating the development of adaptive hybrid systems.