An Analysis of Gpt-4O's Capability to Detect Incongruity in Double-Bind Scenarios With Text and Images

This study evaluates the detection of Incongruity in double-bind scenarios combining text and images, utilizing OpenAI's GPT-4o. The research focuses on three aspects: (1) GPT-4o's capability to detect Incongruity, (2) the impact of image polarity (positive vs. negative impressions) on detection, and (3) alignment between GPT-4o's assessments and human evaluations. Images with both positive and negative valence, along with corresponding texts reflecting similar positive or negative valence, were presented in various combinations. GPT-4o assessed likability, mood, and Incongruity using a 7-point scale. Results indicate that GPT-4o effectively identifies Incongruity in multimodal inputs, with performance closely matching human evaluations. However, inconsistencies were noted in judgment trends influenced by image polarity. Despite these variations, GPT-4o predominantly prioritized textual cues over visual data in detecting Incongruity. These findings highlight the promise of multimodal AI systems in recognizing complex emotional and logical inconsistencies, while also underscoring limitations caused by image-induced biases. Future research should address these limitations by leveraging diverse datasets and real-world scenarios to enhance the robustness and generalizability of such systems.