Finger Lakes Community College webinar explains how AI learns bias and how users can reduce it

Deborah Ortloff, co‑leader of the AI initiatives and FLX AI Hub at Finger Lakes Community College, opened the college’s inaugural monthly webinar and introduced a session on “AI and bias.” The presentation was led by Dave Gadeau, a computing science professor at the college, who framed modern large language and image models as statistical systems that “read everything there is to know” and then produce outputs by sampling high‑probability continuations; as he put it, “AI is nothing more than a probability machine.”

Gadeau walked attendees through a range of real‑world examples of how biased or incomplete training data produces flawed or discriminatory results. He contrasted map projections to show dataset imbalance, noting that if a model mainly sees Mercator maps rather than size‑accurate projections, its internal associations are skewed. In images, he said decades of cultural habits — for instance, more baby photos of girls in pink and boys in blue — can train vision models to associate color with gender rather than facial features, producing stereotyped outputs.

The talk included technical failure modes. Gadeau described a published medical‑imaging case in which models learned to treat the presence of a ruler in X‑ray images as a proxy for pathology; when images lacked the ruler, the system failed to detect disease. He also recounted hiring and moderation mistakes: an Amazon resume‑screening tool that favored male resumes and early chatbots trained on unfiltered social media that quickly produced hateful content.

Beyond harms, Gadeau emphasized that many problems are fixable or mitigable. He described tools and techniques for practitioners and community users: adversarial tools like Nightshade (used by some artists to perturb images so automated ingesters misclassify them), inclusive LLMs such as the startup Latimer, and source‑auditing tools like NotebookLM that let users control the set of documents a model can consult. He advised concrete user practices: check AI outputs with human judgment, avoid placing personally identifiable information into chat tools (“No PII in AI”), use careful prompting to request diverse or specific representations, and consider running smaller models locally when privacy is a priority.

During Q&A, attendees asked where responsibility should lie for correcting bias. Gadeau said responsibility is shared: users can practice responsible prompting and model selection, but model developers and platform operators also must improve training practices and auditing. When asked about litigation, he said there are many lawsuits involving AI (for hallucinations, harms and hiring practices), though he could not cite a single landmark bias case off the top of his head. On feedback mechanisms, he said thumbs‑up/down and similar signals likely help models fine‑tune future responses but warned that those mechanisms are not a substitute for systematic auditing.

The webinar concluded with a reminder to join FLCC’s AI mailing list at flcc.edu/ai for slides and recordings and an encouragement to treat AI as a thought partner rather than an arbiter: “co‑create, don’t abdicate,” Ortloff said. The organizers said slides, citations and additional resources from the presentation would be shared with registrants.