A HYBRID, MULTI-LAYERED AI-POWERED SYSTEM TO COUNTER DISINFORMATION IN EMERGENCY SITUATIONS

During the 2024 floods that affected several regions of Kazakhstan, messaging platforms and social networks were flooded with unverified reports: false claims about dam breaches, inflated casualty figures and outright fabrications spread faster than official bulletins and disrupted the work of emergency response headquarters. This paper presents CrisisStack, a system designed to separate reliable messages from unreliable ones in real time. The architecture is a three-level pipeline: a multilingual LaBSE encoder produces a unified vector representation of the text, one-dimensional convolutional layers detect local manipulation markers, a bidirectional LSTM combined with self-attention analyses the message context as a whole, and a multi-agent protocol cross-checks the neural prediction against external data sources. The model was trained on a combined corpus of 35,281 records from the open WELFake and CrisisMMD datasets using mixed-precision training. On a held-out test set of 3,711 records the architecture reached an accuracy of 84.9% and a macro-F1 of 0.846, outperforming the mBERT baseline (F1 = 0.735) and LaBSE with a linear classifier (F1 = 0.806). 
An ablation study showed that the attention layer contributes the largest single gain among the cascade components. Zero-shot transfer to Kazakh and Russian texts was not achieved: on an open sample of an annotated Kazakh–Russian corpus the model reached 47.3% accuracy, which is reported as a key limitation and defines the direction of future work. The verification protocol was tested under software simulation of external sources; the end-to-end response time on a Tesla T4 GPU was 16.9 ± 1.7 ms.