Towards AI-assisted fossil identification

Recent advances in Vision-Language Models (VLMs) offer new opportunities for paleontology, but fossil identification remains constrained by variable image and annotation quality and by subtle or fragmented diagnostic features. We investigate three directions for improving AI-assisted fossil identification: (1) real-time image-quality guidance, (2) conversational AI for citizen-science identification and data collection, and (3) richer localized morphological annotations. We introduce GoldenFossils, a proof-of-concept dataset of expert-validated localized annotations created through an AI-assisted human-in-the-loop framework, and evaluate state-of-the-art VLMs with and without localized evidence. Results show benefits from image-quality guidance, complementary strengths of conversational and form interfaces, and modeldependent gains from localized expert evidence. Together, these findings highlight how better fossil data collection can support more reliable AIassisted identification and research into past biodiversity and evolution.