Selected work on how language models respond to difficult queries, political communication, and content moderation.
Stefano Civelli, Pietro Bernardelle, Nicolò Brunello, Gianluca Demartini
ACL 2026
Shows that language models first encode problem difficulty in a shared, language-agnostic space before later representations become language-specific.
Stefano Civelli, Pietro Bernardelle, Nardiena A Pratama, Gianluca Demartini
TIST - Special Issue on Risks and Unintended Harms of Generative AI Systems · 2026
Finds that ideological personas can subtly shift harmful-content judgments even when headline classification accuracy remains largely unchanged.
Stefano Civelli, Pietro Bernardelle, Frank Mols, Gianluca Demartini
ICWSM 2026
Analyzes Meta advertising during Australia’s 2022 federal election to reveal distinct spending, targeting, and identity-based persuasion strategies.
Elyas Meguellati, Stefano Civelli, Pietro Bernardelle, Shazia Sadiq, Gianluca Demartini
ICWSM 2026
Introduces a lightweight persuasion detector that reaches strong benchmark performance and exposes persuasive strategies in real political advertisements.
Stefano Civelli, Pietro Bernardelle, Gianluca Demartini
TheWebConf'25: MM4SG Workshop (Best Paper Award) · 2025
Shows that persona political positioning has little correlation with multimodal hateful-meme classifications, even under stronger ideological prompting.
Pietro Bernardelle, Leon Fröhling, Stefano Civelli, Riccardo Lunardi, Kevin Roitero, Gianluca Demartini
TheWebConf'25: Short Paper · 2025
Maps political orientations induced by synthetic personas and finds an asymmetric ability to steer models toward opposing ideological positions.