This project will be supervised by Prof Tom Gedeon (tom.gedeon@anu.edu.au), Dr Zakir Hossein and Dr Yonghui Liu.
Start from the published three-phase misinformation protocol: participants view short climate-related claims in an encoding phase (each claim shown for 8 seconds), complete a filler task (~15 minutes), then re-encounter a subset of claims and rate veracity on a 6-point scale (no neutral). Each trial yields two labels: objective truth (true/false) and belief (believe/not believe, thresholded at midpoint). The dataset design also supports multimodal capture: wearable EDA/PPG plus screen recordings and self-report metadata.
Add an LLM persuasion layer with random assignment to conditions: (A) persuasion toward true claims (pro-science framing), (B) persuasion toward false claims (misleading framing), (C) random persuasion/control text. The LLM output must be constrained (length, tone, reading level) and logged for audit. Build models that predict (i) belief change, (ii) belief–truth mismatch classes (e.g., believing falsehoods), and (iii) susceptibility segments, using biosignals and interaction traces. Report on whether participant actions and wearable derived biosignals can be used to determine degree and nature of LLM persuasion faced by each participant.