Researchers at the ABV-Indian Institute of Information Technology (IIIT), Gwalior, have developed a new artificial intelligence system that combines text analysis, image recognition and fuzzy logic to detect fake news in Indian media with high accuracy.

The system, called F2IND-IT! (fuzzy fake Indian news detection using images and text), was described in a recent paper uploaded to arXiv. The researchers say the project addresses a growing challenge in India, where rapid internet penetration and social media use have accelerated the spread of misinformation.

According to data from the Press Information Bureau, under the Ministry of Information and Broadcasting, 1,575 fake news cases were reported between 2022 and March 2025. The number rose from 338 in 2022 to 583 in 2024. Data from the National Crime Records Bureau also show a 214 per cent increase in fake news cases during the early pandemic period from 2018 to 2020. A 2024 study by ISB and CyberPeace found that 46 per cent of false information was about politics, and over 77 per cent of it spread through social media platforms. Another survey among Gen Z users in Delhi found that 91 per cent believe fake news can affect election outcomes.

To tackle the problem, the researchers designed a multimodal AI model that analyses both the written content of news articles and the accompanying images. The framework uses DistilBERT — a lightweight language-processing model — to understand text semantics, while a convolutional neural network (ResNet-50), which is a deep-learning image recognition system, extracts visual features from photographs. These inputs are then combined using an ‘attention mechanism’ and processed through an adaptive neuro-fuzzy inference system (ANFIS), which produces a probability score indicating whether a news item is fake or genuine.

The model was trained and tested on the Indian Fake News Dataset (IFND), which contains more than 56,000 news articles spanning politics, elections, Covid-19, violence and other topics. According to the paper, the proposed system achieved an accuracy of nearly 98 per cent, outperforming several alternative model configurations in ablation studies.

The researchers say future versions could rely less on manually designed fuzzy rules and instead use data-driven systems capable of dynamically generating their own inference structures during training.

Published on June 1, 2026