Core Projects - Trust, Perception, and Human-AI Interaction

In digital life, first impressions increasingly happen through profile pictures. From online marketplaces to public services, from political communication to social media, we often decide whether to trust, cooperate, or engage with someone based on a face on a screen. At the same time, recent advances in artificial intelligence make it easy to subtly edit facial images so that a person appears more trustworthy, warmer, or more competent — without changing their identity. This raises a crucial question: do AI-edited faces actually change how people behave toward others in economically meaningful situations? And if they do, can transparency protect us?

This project investigates whether small, AI-driven changes in facial appearance influence beliefs and decisions in situations involving trust, cooperation, conflict, and fairness. Rather than focusing on extreme manipulation or deception, we study subtle edits that are already realistic and widely accessible. The core idea is simple but powerful: if AI-edited faces systematically bias expectations about others’ behavior, they may distort decisions even when people have clear economic incentives to act rationally. Such effects would represent a new, low-cost channel of influence in digital environments.

The relevance of the project is both societal and policy-oriented. If AI-edited faces affect trust and cooperation, this has implications for online markets, political communication, influencer marketing, and fraud prevention. At the same time, if simple disclosure proves effective, it would point to a low-cost, scalable safeguard that platforms and regulators could adopt without restricting innovation. By providing causal evidence on when appearance manipulation matters — and when transparency helps — the project contributes to building more resilient, trustworthy digital interactions in an AI-mediated world.

Much of cognitive neuroscience is driven by a simple but fascinating question: how does neural activity support brain function? To address this question, the neuroimaging community has increasingly turned to big data, combining datasets from large numbers of participants. This population-based approach has proven valuable for revealing how neural systems support cognition and behavior on average. However, it tells us far less about why brain activity is unique to each of us. For example, which neural features account for differences from one person to another?

To address this challenge, our project aims to develop a “digital twin brain” - a detailed and informative virtual model of the brain. Rather than averaging patterns of brain activity across people, our approach focuses on creating a personalized model capable of accurately reconstructing the brain activity of a single individual. The key advantage of our method lies in its ability to map how different brain regions activate and interact by using data from two complementary neuroimaging techniques: functional magnetic resonance imaging (fMRI) and electroencephalography (EEG).

In perspective, our digital twin brain is explicitly meant to address many long-standing questions in the field of cognitive neuroscience and neuroimaging. For example, how do functional interactions between regions emerge from reciprocal structural connections? How can our virtual model inform us on how non-invasive techniques, such as transcranial magnetic stimulation (TMS), can be used to effectively modulate brain activity?

Importantly, we envisage that our digital twin brain can be used to study diverse populations, including individuals with developmental conditions or with sensory deprivation. Ultimately, our personalized approach will improve our understanding of what makes each of us unique and support the discovery of more targeted and effective treatments.

Some individuals behave prosocially across repeated decisions, while others reduce prosociality after a good deed. We argue that these patterns reflect the curvature of moral satisfaction, the hedonic utility from acting in line with one’s moral values. Increasing marginal moral satisfaction generates moral consistency, whereas diminishing marginal moral satisfaction generates moral licensing. Thus, we ask whether individuals experience increasing or decreasing marginal moral satisfaction from good deeds, and if this curvature varies across people. We conduct an online two-stage experiment that holds material incentives constant while varying moral framing. Combining behavioral choices with linguistic measures of moral engagement allows us to recover substantial heterogeneity in the shape of moral satisfaction. We find that participants with higher moral engagement exhibit diminishing marginal moral satisfaction, while less engaged participants display increasing marginal moral satisfaction. Finally, we find that moral satisfaction extends across domains: moral actions in different contexts appear to serve as substitutable sources of hedonic utility, helping explain when good deeds crowd in versus crowd out future prosocial behavior. These findings clarify when moral incentives reinforce prosocial behavior and when they instead induce self-licensing, with implications for the design of moral nudges.