Project links for Persona-non Graphica
AI Bias and Stereotypes
AI systems learn from human-generated data, and that data inevitably reflects the assumptions, stereotypes, and inequalities present in society. At the same time, AI systems do not simply reproduce these patterns: once deployed, they can reinforce and amplify them, influencing how people perceive, categorize, and interact with others.
Our research investigates this two-way relationship between people, data, and AI. We examine how biases enter datasets and computational models, how system design can make some groups and perspectives less visible, and how AI-generated representations and behaviours can feed existing stereotypes back into human decision-making.
A particular concern is what gets left out. Data collection and modeling practices often favor majority groups, average across diverse experiences, or remove observations that do not fit dominant patterns. These choices can make differences and minority experiences statistically invisible. What appears to be a neutral or representative model may therefore encode a particular view of what is “normal,” while treating other experiences as exceptions or noise.
We also study how AI agents themselves become social actors. Their appearance, language, roles, behaviours, and personas can trigger existing stereotypes and create new expectations about what different kinds of agents—or the people they represent—should be like. These expectations can then influence how people interact with and evaluate AI.
Our research examines
Bias in Human-Generated Data
How social assumptions, unequal representation, and historical biases become embedded in datasets and models.
Missing & Marginalized Data
How underrepresented groups, atypical experiences, and minority perspectives can disappear through data collection, aggregation, or model optimization.
Stereotypes in AI
How models and artificial agents reproduce, amplify, or potentially challenge existing social stereotypes.
AI as a Feedback Loop
How AI-generated representations influence people's perceptions, expectations, and subsequent behaviour, potentially reinforcing the patterns from which the systems learned.
Plurality & Difference
How AI can represent variation between people rather than averaging diverse experiences into a single “typical” user or behaviour.
Contextual Interpretation
Recognizing that human behaviour and social meaning cannot always be reduced to one correct label or prediction.
Ultimately, we ask not only “Is the AI biased?”, but also:
“What assumptions about people are embedded in the data and design of AI— and what happens when those assumptions become part of how people understand one another?”
Our goal is to develop AI systems that better preserve human diversity, uncertainty, and plurality, rather than making differences invisible in the pursuit of a single generalized model.
