This blog post is part of a series of posts covering community members’ experiments with AI in the classroom and the workplace.
Anna Rafferty, Professor of Computer Science, recently led a FOCUS first-year seminar on large language models (LLMs) and their social impact. In the seminar, students learned the basics of how LLMs like ChatGPT and Gemini work, along with some of the ways their output might be biased or context-sensitive.
The seminar began with a primer on the technical aspects of LLMs. Here, students learned about the stochastic processes underlying AI chatbots, as well as the practices and data involved in training these models. These preliminary lessons prepared students to understand, from a technical point of view, how trends and biases in training data can lead to biased output from the model. According to Rafferty, they were also designed to counter the common unconscious assumption that machines are always objective.
Following this primer, students picked from a set of five contexts in which LLMs might amplify—or already have amplified—certain biases and inequities. They then read two articles about that topic, one from the popular press and the other from a peer-reviewed journal. One topic students read about, for example, was how AI resume review tools treated mentions of disability. In their reading, those students learned about cases where AI marked down resumes with awards for disability advocacy as potentially being “too specialized.”
After hearing what their classmates read about, students broke into research groups to investigate first hand a new topic that interested them. Altogether, the groups covered a wide range of topics. One group studied differences in output and ideological alignment between LLMs developed in different geopolitical contexts, such as DeepSeek (China) and Mistral (France). Another group examined how information like name and neighborhood of residence affected AI-generated criminal sentencing recommendations. The seminar ended with a poster presentation session where students could present their experiments and findings.
For Rafferty, one of the overarching goals of the seminar was to give students a sense of what it means to understand a complex system and how one might design an experiment to do so. And given the complexity of the technical and social questions behind AI, LLMs count as especially interesting instances of such systems. As Rafferty herself puts it, “LLMs are often portrayed as either too magical or not magical enough.” Excessive hype builds up AI technology as an unrealistic cure-all, but it equally risks obscuring the ways in which these tools are genuinely mysterious, powerful, and promising.
Rafferty adds that, in much of the discourse surrounding AI technology as a whole, people tend to focus on pre-built models coming out of tech companies like OpenAI, Google, and Anthropic. But in doing so they miss the nuances and questions coming from AI research itself, which has existed for more than half a century now. By ignoring this history and its achievements and questions, we risk an uncritical response to an important technological and social moment.