Finger Lakes Community College webinar lays out how AI bias emerges and how to mitigate it

Finger Lakes Community College hosted the inaugural FLX AI Hub monthly webinar on AI and bias, where computing science professor Dave Gadeau laid out why popular generative models reflect and amplify patterns in their training data and what practitioners and consumers can do to reduce harm.

Gadeau told attendees that "AI is nothing more than a probability machine," describing how models read vast corpora—books, articles, videos and more—then predict likely continuations. He used a knock‑knock example to illustrate how models surface common continuations and noted that conversational use of these models keeps context across turns, unlike a traditional web search.

Why it matters: biased training sets can produce distorted or discriminatory outputs. Gadeau gave multiple concrete examples: image generators that produced a white male for "doctor" and a white woman for "teacher"; hiring‑screening systems that prioritized historically common resumes (an Amazon example he said the company later scrapped); and risk‑assessment tools that produced divergent scores for defendants with different criminal records. He said such outcomes are not abstract: "If you don't have a ruler on your chest at that time, AI is not gonna identify it as a tumor," describing a documented medical‑imaging labeling error tied to spurious cues.

Gadeau also reviewed instances where social‑media training caused rapid harms—chatbots trained on Twitter data that adopted antisemitic content within hours—and where systems trained in one geography (autonomous vehicles) failed to account for different traffic patterns elsewhere. He highlighted both unintentional bias from skewed datasets and cases where data selection or platform choices can encode particular viewpoints.

On mitigation, Gadeau recommended a mix of practical steps. He named tools and approaches including inclusive LLMs (he cited a startup called Latimer), running models locally to control data, auditing sources with tools such as NotebookLM, and improving prompts to request diversity or neutrality. He left four succinct guidelines for users: co‑create, don’t let AI be the sole arbiter; don’t input personal identifiable information; remain alert to hidden bias; and verify AI outputs before sharing. "Don't put personal information in an AI. No PII in AI," he said.

During a short Q&A, Gadeau said responsibility for reducing bias is shared: users can practice responsible prompting and organizations building models must improve training and oversight. He acknowledged ongoing litigation against AI companies broadly but said he was "unaware of any" high‑profile lawsuit solely about bias off the cuff while noting many cases involving hiring tools and hallucinations.

The webinar concluded with hosts encouraging attendees to join FLCC's AI mailing list and upcoming events; a recording and slides will be shared with participants.

Next steps: FLCC plans further webinars in the series and a March session on NotebookLM; Gadeau encouraged ongoing community conversations about bias in model design and deployment.