Artificial intelligence is often described as objective, but it is nothing of the sort. AI systems learn from human-created data, and they inherit the patterns, gaps, and prejudices of that data. My research on AI bias and cultural representation examines how this plays out — particularly in AI-generated imagery.
Where bias comes from
Bias in AI models has identifiable origins: skewed training data, the choices of the people who build the systems, and feedback loops that reinforce dominant patterns. When a dataset over-represents some groups and under-represents others, the model treats the majority as the default and the minority as the exception.
The case of AI-generated imagery
Image-generation models make bias visible. Ask a model to depict a profession, a place, or a culture, and it returns the statistical average of its training data — which often flattens or distorts cultural identity. In my content analysis of how AI systems represent Sri Lankan cultural identity, the gap between authentic representation and the model's output was striking.
Why representation matters
Representation is not a cosmetic concern. The images and text AI produces increasingly shape how people see the world and themselves. When a technology consistently misrepresents or erases a culture, it does quiet, cumulative harm — especially to children forming their sense of identity.
The copyright and ownership dimension
Cultural bias intersects with questions of ownership. When AI mimics a distinctive artistic style or cultural motif, who owns the result? My work on the legality of AI-generated art and viral "style" imagery explores how these systems raise unresolved questions about authorship, consent, and cultural appropriation.
What can be done
Mitigation is possible but not automatic. It requires more representative data, diverse teams, transparency about how models are trained, and ongoing auditing. None of this happens by default; it happens when builders treat fairness as a design requirement rather than a public-relations exercise.
Final thought
AI bias is not a glitch to be patched once; it is a structural property of how these systems learn. Addressing it means taking representation seriously at every stage. You can read the full studies on my research page.