Beyond Content: Integrating Generated User Intent and Planned Behavior Theory for Reliable Fake News Detection

With the rise of generative AI, the boundary between authentic and deceptive content has become increasingly ambiguous, challenging traditional fake news detection methods that rely solely on observable content or propagation structures. These approaches often neglect the underlying psychological motivations driving user behavior, leaving them susceptible to adversarial manipulation. However, as the user decision-making process is inherently unobservable, conventional deep learning models struggle to capture the cognitive mechanisms behind information sharing. To address this, we propose TPB-VAE, a psychologically grounded framework that integrates the Theory of Planned Behavior (TPB) with large language models (LLMs) to infer and encode users' latent intent. TPB-VAE maps TPB constructs into a latent space, making the decision-making process computationally accessible. It employs semi-supervised learning specifically to infer users' latent intent from a small subset of labeled samples, and uses the resulting intents to derive rich behavioral features for more reliable fake news detection. Extensive experiments on four real-world datasets demonstrate the effectiveness and adversarial resilience of our approach.