Data Sciences Institute - Data Speaker Series

We’re Talking About the Wrong Error: Why Variance Matters More than Bias in AI

Bias gets all the attention when it comes to AI. And for good reason: in social systems, bias determines whose résumés are seen, who gets access to care, whose voices are amplified or erased, and more. For social scientists, the fact that computational models encode bias, or have algorithmic fidelity to the social and cultural associations embedded in text and images, is precisely what makes them analytically valuable. The focus on bias from both the practical and analytical sides, however, is rooted in a relatively older technology: static word embeddings. With large language models, I argue, the bias challenge/opportunity has been flipped on its head. LLMs are no longer faithful encodings of specific biases, they are amalgams and it is the amalgam, or, more precisely, their lack of variance, that is the core challenge with LLMs. In this talk, I argue for shifting the focus among computational social scientists and data scientists from bias, to variance. I show how low-variance LLMs can be incredibly powerful for certain uses, such as some forms of text classification and annotation, and also why they pose challenges for other uses, especially social simulation and comparative analysis. I close by reflecting on what exactly we are measuring when we use large language models, and why working with them requires a fundamental rethinking of what computational methods are for.

Biography:

Prof. Nelson uses computational methods—principally text analysis, natural language processing, machine learning, and network analysis—to study social movements, culture, gender, and organizations and institutions. Her research examines processes around the formation of collective identities, social movement strategies in feminist and environmental movements, the role of place in shaping activism, intersectionality in women’s movements and 19th-century U.S. South experiences, gender inequality in startups and STEM fields, and how academic ideas are translated into practice through programs like NSF’s ADVANCE. Methodologically, she has proposed frameworks combining computational and qualitative approaches, including the grounded theory framework. She develops and teaches courses on computational methods, Python, R, data science, and sociological theory. Her current projects include researching intersectionality in women’s movements, media coverage of social movements over time, historical memory, and gender inequality across different sectors.

This talk is co-sponsored by the Data Sciences Institute and the Department of Sociology, University of Toronto Scarborough.

For more information, please visit https://datasciences.utoronto.ca/dsi-home/data-sciences-speaker-series/.

Also check out other Contests in Toronto.