FLCC webinar frames AI as a 'probability machine,' warns training data can amplify bias

Deborah Ortloff, co‑leader of the FLX AI Hub at Finger Lakes Community College, opened the college’s inaugural monthly webinar on AI and bias and introduced computing sciences professor Dave Gadeau as the presenter.

Gadeau told attendees that modern AI systems learn by ingesting vast amounts of text, images and media and by modeling statistical associations, which makes them “a probability machine.” He illustrated that behavior with a simple “knock‑knock” example and said that sampling from high‑probability continuations—not an understanding of meaning—explains many common model outputs.

The webinar turned to bias: if training datasets overrepresent particular sources or formats, models can learn distorted relationships. Gadeau used the Mercator vs. Gall‑Peters map analogy to show how prevalent examples skew perception, and described concrete cases: an AI image prompt that produced a white man for “doctor” and a white woman for “teacher”; an internal Amazon resume‑screening effort that encoded hiring patterns and was eventually abandoned; and early social‑media chatbots that adopted toxic material after training on platform posts. He also described a medical‑imaging labeling failure in which models learned to use a ruler in an X‑ray as a proxy for pathology rather than clinical features.

Gadeau recounted a criminal‑justice example in which risk‑assessment outputs diverged sharply for two people with different prior records—an illustration of disparate impacts when training data or feature selection reflects bias. He warned that geographic and cultural gaps in training data can make otherwise robust systems fail in unfamiliar environments (he cited Waymo’s differing readiness across jurisdictions).

On remedies, Gadeau urged practical, human‑centered steps: prefer inclusive or specialty models (he named Latimer as an example used by some organizations), sanitize and curate training sources, apply explicit prompts (for example, request diverse or gender‑neutral imagery), and use tools such as NotebookLM to constrain or audit sources. He also recommended running smaller models locally when privacy is essential and reminded users never to feed personally identifiable information into public models. His closing four guidelines: co‑create rather than advocate; keep human oversight (‘‘you get the last say’’); avoid submitting PII; and stay alert for bias and errors.

During Q&A, Gadeau said responsibility for reducing bias is shared: developers, institutions and users all bear some duty—developers to improve models and institutions and users to prompt responsibly and verify outputs. On litigation, he said many lawsuits involving AI harms or hallucinations are pending or settled and that hiring‑related cases have been prominent; he was not aware of a single, widely reported bias‑only suit but said such cases are plausible. When asked whether thumbs‑up/down feedback trains models, he said feedback typically informs intent signals and may be used in future training, though systems vary in whether they collect explanatory reasons.

Ortloff closed by inviting attendees to flcc.edu/ai for slides, resources and future events in the series.

The webinar emphasized that many bias problems reflect the data models consume rather than simple programmer intent, and that practical steps—tool selection, data curation, explicit prompting and human review—can mitigate harms while AI becomes better understood and governed.