Detecting AI-Generated Maps Using Vision-Language Models with Context-Aware Demonstration Retrieval

Recent advances in artificial intelligence (AI) generative models enable the rapid creation of map-like images by simply providing textual prompts, raising growing concerns about the authenticity and reliability of geospatial information. Detecting AI-generated maps is challenging because synthetic maps can closely mimic the visual style and cartographic elements of conventionally produced maps, making them difficult to distinguish through visual inspection. Recent progress in vision-language models (VLMs) has demonstrated strong capabilities in multimodal understanding and has been widely applied in various visual analysis tasks, including scene interpretation and map-related applications. Directly applying VLMs to identify AI-generated maps remains difficult, though, as subtle visual cues and misleading visual similarities could lead to unreliable judgments. To address this issue, we propose a framework that leverages VLMs with context-aware demonstration retrieval. Specifically, a reference database containing map samples and their corresponding labels (AI-generated or non-AI-generated) is constructed. For a given query map, a set of contextually consistent examples is retrieved from the database to serve as demonstrations for in-context reasoning. Instead of relying solely on pairwise visual similarity, the proposed retrieval strategy considers relationships among candidate samples, enabling the model to access a broader set of retrieval contexts beyond direct pairwise similarity. Experimental results show that incorporating retrieved demonstrations improves the reliability of AI-generated map detection compared with direct zero-shot inference. In addition, the retrieved examples provide more informative contextual evidence than simple similarity-based retrieval, supporting more robust reasoning.