Linguistics Colloquium Series: AI and the Compression of Human Communication: Bias, Big Data, and Language Education

The Linguistics Department of Ohio University presents Dr. Elliott J. Casal discussing: "AI and the Compression of Human Communication: Bias, Big Data, and Language Education" in Alden 319 and via Teams Join the meeting here Abstract: Generative AI and other AI-powered systems are often framed as neutrally reflecting human language, emotion, and communication at unprecedented scale. The scale is certainly there, but these systems also operationalize particular distributions, training histories, and assumptions about what counts as representative communication. In doing so, they tend to privilege prototypicality, central tendencies, and highly frequent forms while effectively compressing variation and marginalizing the broader "long tail" of linguistic distributions in human expression. In this talk, I examine these tensions across multimodal and linguistic AI research contexts, highlighting two different but related forms of biases in these systems. First I draw on collaborative research using United Nations General Assembly data (Windsor et al., under review) to discuss how commercial AI-powered emotion-recognition software struggled to detect and classify faces with darker skin pigmentation in our ongoing project. This raises important questions about representation and bias in training and development of AI-mediated research tools. I then turn to corpus-based analyses comparing human and AI-generated language across thousands of texts (Casal et al., 2025; Windsor e