Assessing the impact of cultural misinformation deepfakes through AI-based user engagement analysis

Purpose This study aims to propose and evaluate a resource-efficient artificial intelligence (AI)-based workflow for analyzing user stance toward misleading cultural content on social media, using heritage-related deepfake-supported narratives as a case study. The aim is to quantify belief, criticism and neutrality in user reactions, and to examine how these stances contribute to the impact and reinforcement of misleading content over time. Design/methodology/approach The proposed framework combines large language models (LLMs) for zero-shot stance detection with sarcasm-aware correction via multinomial logistic regression. Lightweight, fine-tuned models are then used to classify stance and compute impact indicators by integrating engagement metrics and adjusted scores that account for polarity and sarcasm. Findings Results show that critical reactions are prevalent, yet comments expressing belief can generate disproportionately high engagement, especially in the early life cycle of posts. The hybrid pipeline improves stance classification accuracy over the baseline while revealing persistent errors linked to sarcasm and implicit language. Research limitations/implications This study relies on a single platform, and a limited number of cases, which may restrict generalization and leave residual errors in contexts with complex figurative or highly informal expressions. Practical implications The framework supports the monitoring of emerging misleading narratives, helping institutions prioritize interventions and design corrective communication strategies. Compared with subscription-based or large-scale alternatives, it requires relatively modest computational resources, although its implementation still presupposes technical expertise. Originality/value The work contributes a domain-specific, resource-efficient methodology for assessing the impact of cultural misinformation, focusing on user stance rather than content verification alone. It integrates LLM-assisted labeling with lightweight classifiers and sarcasm-aware correction, tailored to heritage-related deepfakes and ambiguous user reactions, bridging digital humanities and AI-based misinformation analysis.