Public debate on artificial intelligence (AI) has expanded rapidly, with news media playing a central role in framing its promises, risks, and governance challenges. Yet existing computational approaches often struggle to capture the longitudinal evolution of these frames, as snapshot-based analyses may conflate genuine discursive shifts with analytical artefacts. To address this limitation, we propose a multilayer network approach to temporal frame detection. Drawing on framing theory, media discourse is modelled as a layered semantic network in which nodes represent lexical units and edges encode sentence-level co-occurrences. The framework combines dynamic TF-IWF filtering, multislice modularity optimisation with the Leiden algorithm, and intertemporal coupling to identify evolving framing structures across time. Structural continuity between successive periods is further validated through Matrix Similarity. The methodology is applied to a corpus of 8,453 articles from four English media sources – BBC, The New York Times, The Guardian, and Wired – spanning 2022–2025 across 16 quarterly windows. The comparative analysis reveals source-specific framing signatures, distinct structural responses to major AI milestones, and a shared discursive trajectory whose timing and composition vary systematically with editorial identity.
Temporal Dynamics of AI News Frames: A Multilayer Network Approach
Michelangelo Misuraca;
2026
Abstract
Public debate on artificial intelligence (AI) has expanded rapidly, with news media playing a central role in framing its promises, risks, and governance challenges. Yet existing computational approaches often struggle to capture the longitudinal evolution of these frames, as snapshot-based analyses may conflate genuine discursive shifts with analytical artefacts. To address this limitation, we propose a multilayer network approach to temporal frame detection. Drawing on framing theory, media discourse is modelled as a layered semantic network in which nodes represent lexical units and edges encode sentence-level co-occurrences. The framework combines dynamic TF-IWF filtering, multislice modularity optimisation with the Leiden algorithm, and intertemporal coupling to identify evolving framing structures across time. Structural continuity between successive periods is further validated through Matrix Similarity. The methodology is applied to a corpus of 8,453 articles from four English media sources – BBC, The New York Times, The Guardian, and Wired – spanning 2022–2025 across 16 quarterly windows. The comparative analysis reveals source-specific framing signatures, distinct structural responses to major AI milestones, and a shared discursive trajectory whose timing and composition vary systematically with editorial identity.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
