Toan Quoc Nguyen
Germany
PhD Student
TU Dortmund University
Neuroscience/Cognitive Science, Machine Learning/Technical AI
Germany PhD Student TU Dortmund University Neuroscience/Cognitive Science, Machine Learning/Technical AI
Project
The Narrative Prior: A Computational Model of Personhood for Advanced AI
Project Overview
Attributing experience to another mind is always an interpretation of ambiguous evidence under some narrative frame. For AI systems, this is unavoidable: no ground truth exists against which to check such attributions. This project uses a case where ground truth does exist, human neural signals paired with expert emotion annotations, to calibrate how far narrative framing pulls interpretation away from truth. We call the risk that a persistent narrative about a person, for example framing them as "recovering" or "in crisis", distorts interpretation of new evidence protagonist bias. This project makes three contributions. First, we characterise the extent to which human brain activity, recorded via functional Magnetic Resonance Imaging (fMRI) from 20 participants during naturalistic story listening, reflects a character's emotional arc over time, rather than only the immediate scene. Second, we develop and validate a classifier that decodes emotion directly from this fMRI signal, quantifying how much emotion-relevant information it carries. Third, we evaluate whether a large language model's (LLM's) interpretation of this brain-derived emotional signal is measurably biased by a protagonist-bias narrative prior, compared to a matched non-narrative control. Together, these contributions ground the study of narrative bias in AI emotion interpretation in human neural data.
Mentors
About the Scholar
Toan Nguyen is a PhD researcher in the Human-AI Interaction Group at TU Dortmund University, Germany. His research focuses on Human-AI Interaction, Brain–Computer Interfaces (BCIs), and multimodal machine learning. He is particularly interested in developing adaptive and user-centred intelligent systems that leverage physiological signals and behavioural data to better understand and support human interaction with AI technologies.
Links
