1 April 2025 in Research & Politics
A growing body of literature examines manipulative information by detecting political mis-/disinformation in text data. This line of research typically involves highly costly manual annotation of text for manual content analysis, and/or training and validating automated downstream approaches. We examine whether Large Language Models (LLMs) can detect pro-Kremlin disinformation about the war in Ukraine, focusing on the case of the downing of the civilian flight MH17. We benchmark methods using a large set of tweets labeled by expert annotators. We show that both open and closed LLMs can accurately detect pro-Kremlin disinformation tweets, outperforming both a research assistant and supervised models used in earlier research and at drastically lower cost compared to either research assistants or crowd workers. Our findings contribute to the literature on mis/-disinformation by showcasing how LLMs can substantially lower the costs of detection even when the labeling requires complex, context-specific knowledge about a given case.
16 April 2026 in SSRN Electronic Journal
Object detection is increasingly used in a variety of safety-critical settings, such as UAV (Unmanned Aerial Vehicle) search-and-rescue or pedestrian detection in autonomous driving. In such contexts, mis- or non-detected objects can have severe ethical consequences. However, the opaque nature of deep learning models complicates identification of risks, thereby undermining epistemic trust. Explainable AI (xAI) methods have been developed to make these models more human-understandable. This systematic review examines 75 studies presenting xAI methods designed for object detection, assessing both their methodological design and evaluation practices. While current methods offer valuable insights, several challenges hinder trustworthy implementation. Evaluation practices remain inconsistent, relying on fragmented metrics that complicate method comparison. Local explanations are most prominent, but are open to robustness and generalizability problems. Moreover, variation in datasets, model architectures, and evaluation practices introduces bias risk and reduces reproducibility. To advance the field, more standardized evaluation, increased robustness, and generalizable insights are needed, taking into account more complex and specific architectures. These advances will improve the validity and reliability of xAI methods and support trustworthy deployment of object detection in safety-critical contexts.