<![CDATA[Visual false-context misinformation pairs an authentic image with an inaccurate claim about when or where it was captured, and is prominent in disaster reporting. Existing verification methods either rely on external web evidence unavailable before publication or on text-only analysis that is vulnerable to professionally written false narratives. This study designs and evaluates a proof-of-concept system that extracts Indonesian location and date claims using BERT-based Indonesian Named Entity Recognition and checks their consistency against GPS and timestamp fields in EXIF metadata through great-circle distance and temporal-delta calculations. Evaluation used a controlled synthetic benchmark of 250 samples built from 50 Flickr images with five narrative variants each, split by photo identity so that the test set held 50 narratives from 10 held-out images. Three configurations were compared, namely TF-IDF with linear SVM, an NLP-only rule system, and the proposed EXIF-assisted consistency checker. On this controlled test set, the hybrid configuration classified all 50 samples correctly, reaching 100% accuracy with an exact 95% confidence interval of 92.9% to 100%, whereas the two text-only configurations showed strong class bias with F1-Macro of 54.03% and 35.52%. However, the labels and hybrid features were both defined from the same spatio-temporal deviation, the test set was small and clustered by image, and EXIF integrity was assumed rather than authenticated. The findings therefore demonstrate correct implementation and potential utility for consistency screening, not real-world misinformation verification. External validation on human-written cases, missing or tampered metadata, and cluster-aware statistical evaluation is required before newsroom deployment.]]>
Copyrights © 2026
