FLCC webinar explains how AI training produces biased outputs and what to do about it

Dave Gadeau, computing science professor at Finger Lakes Community College, told attendees that large language and image models learn from "everything there is to know" in their training sets and that those sources shape their outputs. He used everyday examples — map projections, baby photographs and hiring histories — to show how common cultural artifacts and historical practices can produce systematic distortions in AI results.

Gadeau said that the production process for modern models is fundamentally probabilistic: "AI is nothing more than a probability machine" and it ranks and samples among high‑probability continuations. That means biases in the underlying corpus will surface repeatedly unless explicitly corrected. The webinar noted high‑profile failures — resume‑screening prototypes and social‑media–trained chatbots — to argue for human oversight, data hygiene and better vendor practices.

Why this matters: biased outputs can affect who is hired, how medical images are triaged and what images or narratives are presented to the public. Presenters recommended concrete mitigation steps including auditing training sources, instructing models with careful prompts, considering private LLM deployments for sensitive uses, and using tools that provide verifiable source lists. Organizers said slides and a recording will be shared and invited signups for the FLCC AI mailing list.