Dissertation Defense: Mostofa Najmus Sakib

Title: Understanding Human News Believability and Testing its Robustness against Linguistic AI-based Adversarial Attacks

Program: Computing PhD

Committee Chair: Francesca Spezzano

Committee: Francesca Spezzano, Edoardo Serra, Jerry Alan Fails

Abstract: Social media have become a primary channel for news consumption, accelerating information diffusion while amplifying the circulation of low-quality and misleading content. While prior work has emphasized content accuracy and network dynamics, comparatively less attention has been paid to how humans perceive and believe news, how nuanced linguistic cues shape belief and sharing, and how emerging AI systems can exploit these mechanisms.

This dissertation advances research on online information by centering human news believability as a core perceptual signal and examining its linguistic foundations and robustness under adversarial manipulation. First, we introduce a large-scale human-annotated dataset of perceived news believability derived from 14,535 Twitter comments and show how news-level linguistic features and user-level traits jointly shape belief, further demonstrating that incorporating believability improves automated credibility assessment. Second, we identify linguistic ambivalence, defined as the co-occurrence of conflicting linguistic cues, as a computational signal that distinguishes low-credibility from high-credibility news, changes patterns of perceived believability, and predicts information diffusion across textual, temporal, psychological, and content-related dimensions.

Finally, we examine the robustness of credibility judgments under large language model-driven credibility drift attacks through two complementary studies. First, we develop and evaluate credibility drift attacks, showing that subtle, fact-preserving linguistic modifications can systematically shift predicted believability. Second, through a controlled human-subject study, we examine how these credibility-oriented transformations affect human judgments, finding that credibility-enhancing specificity significantly increases perceived believability, while low-credibility source framing significantly decreases it, and analyzing participants’ explanations to understand the cues underlying their judgments. Collectively, this work establishes human believability as a measurable and predictive signal, highlights the role of nuanced language in belief and spread, and exposes emerging risks posed by AI-enabled credibility manipulation. This research advances knowledge in social computing, with its outcomes published in leading venues including CSCW, ICWSM, WSDM, and TRB.