A DATASET AT A GLANCE...

Nowadays, Artificial Intelligence (AI) models like, Large Language Models (LLMs) and Vision-Language Models (VLMs) have eased up textual and image understanding, respectively.  While LLMs excel in manipulating, generating or summarizing the text with a textual query, VLMs stand a level higher in understanding and creating out text for images!!! Actually, these AI models are trained on previously-available, abundant sources of information to generate a factual text as possible, as the user provides a textual or image query. However, there are also possibilities that a deceiver could use these models to purportedly generate misleading disinformation with a deceptive textual or image query. The examples include generating a contextually-distorted script or fabricating a false content using text, image or text-image pair. Hence, these models, which are designed to exceptionally aid in natural language interpretation and scene understanding are now becoming a forum to generate and propagate disinformation. Do you know that the MIT AI Risk Repository (https://airisk.mit.edu/) has classified disinformation as a systematic threat? So, there is an urgent need for an AI safety benchmark to detect disinformation across text and visual contents to prevent fake news propagation. Shaina Raza, Ashmal Vayani, Aditya Jain, Aravind Narayanan, Vahid Reza Khazaie, Syed Raza Bashir, Elham Dolatabadi, Gias Uddin, Christos Emmanouilidis, Rizwan Qureshi and Mubarak Shah have put forth an AI safety benchmark dataset, called the ‘VLDBench: Vision Language Models Disinformation Detection Benchmark’. This dataset can be accessed from the GitHub repository (https://vectorinstitute.github.io/VLDBench) and it allows to perform multimodal benchmarking of text and image disinformation detection. The various attributes of this dataset are that it involves 31000+ news article-image pairs, falling under 13 unique news categories. Moreover, this dataset was manually verified by 22 domain experts, spending about 300+ hours. Hence, the researchers working to model AI systems with factually-better unimodal or multimodal information generation or interpretation can avail this dataset. Image courtesy: www.freepik.com