Vision-language models detect real emergencies reasonably well, but they often overreact by misclassifying safe scenes as dangerous. This post shares results from the VERI benchmark, which measures this limitation using visually similar emergency and safe scenes with different meanings.