Understanding Human Heuristics in Context-Sensitive Image Captioning

Recent studies highlight the context sensitivity of image captioning, where the context in which an image appears strongly influences its caption's informativeness and linguistic style. While AI-generated text increasingly resembles human language, its informativeness and coherence, derived from cross-modal image-text reasoning, may still fall short of content generated by human experts. Given the intertwined nature of informativeness and linguistic style, this study examines news image captioning, a naturally high-context task, to manipulate caption informativeness and assess human sensitivity to such variations. Two experiments (N = 378) and logistic regression analyses reveal that while humans effectively interpret informational cues, their intuition about AI linguistic style often diverges from actual AI language markers. Moreover, humans more readily integrate multiple modalities in preference tasks but rely heavily on linguistic-based strategies for AI detection. These findings underscore the adaptability of human evaluation in image-text systems and suggest informative signals as the more reliable basis for judgment.