Can Generative AI accurately evaluate local communities, or does it hold hidden biases?

Описание: In this CIVICA public lecture, researchers compare GPT-4o's assessments of urban neighbourhoods, based on crowd-sourced street view images, with evaluations from both a nationally representative survey and a locally representative survey.

The findings reveal that while GPT-4o aligns closely with national averages, it deviates significantly from local perspectives, particularly on perceptions of safety, wealth and disorder. Furthermore, gender and racial biases in GPT-4o's assessments amplify these discrepancies, and attempts to mitigate this bias through tailored prompts were largely ineffective and sometimes counterproductive.

SPEAKERS

Ala Alrababah (Bocconi University): Assistant Professor in the Department of Social and Political Sciences. His research focuses on political violence, immigration and refugees, and authoritarian media, using fieldwork, experimental approaches, and computational methods.

Elias Dinas (European University Institute): Swiss Chair in Federalism, Democracy and International Governance. His research explores political socialisation, institutional interventions, and the ideological effects of authoritarian rule on democracies. His work is published in leading journals and has been featured in outlets such as The Economist and The New York Times.

Melissa Sands (London School of Economics): Researcher in the Department of Government. She studies the influence of context on political and civic behaviour, with her work appearing in journals such as American Political Science Review and Nature.