FLCC webinar: How biased training data, design choices and prompts shape AI outputs — and what to do about it

Deborah Ortloff, co-leader of the AI initiatives and FLX AI Hub at Finger Lakes Community College, opened the inaugural monthly webinar and introduced computing science professor Dave Gadeau, who led a 90-minute primer on how artificial intelligence systems develop and reproduce bias.

Gadeau said large language models and image generators are trained by ingesting vast collections of text, video and images and then computing statistical associations to predict likely next words or pixels. "AI is nothing more than a probability machine," he said, illustrating the point with a "knock knock" example to show how models select among high-probability continuations.

Why that matters, Gadeau said, is that models reflect the composition and limitations of their training corpora. He used the Mercator versus Gall–Peters map example to show how dominant sources can skew a model's 'view' of the world, and described how historical photo trends led image models to infer gender from clothing color rather than facial features. "When I asked for a photorealistic image of a doctor, I got a white male," he said, then showed a prompt that produced a white woman for "teacher," using both examples to demonstrate stereotyped outputs.

The presentation highlighted domain-specific failures. In medical imaging, Gadeau explained, models once learned to associate the presence of a ruler in X-rays with pathology because many labeled images included rulers; when images lacked rulers, the system missed findings. He also recounted corporate examples: an Amazon resume-screening experiment that reproduced gender bias and was later abandoned, and chatbots trained primarily on social-media streams that adopted hateful content quickly.

Gadeau discussed geographic and cultural limits for models, noting that an autonomous driving system trained on U.S. roads may not recognize objects common elsewhere (he cited rickshaws as an example). He reviewed IBM Watson's earlier attempts in medicine and attributed part of Watson's problems to differences in clinical practice, unstructured records and privacy constraints that limited data access.

On intent and mitigation, Gadeau distinguished unintentional bias from deliberate tuning and described both defensive and corrective approaches. He named tools used by creatives (Nightshade) to reduce scraping of artwork and recommended inclusive or audited LLMs (he cited a startup called Latimer), local model runs for privacy, curated-source tools such as NotebookLM, and careful prompting (for example, specifying demographic diversity when requesting images). He advised users to avoid submitting PII to public models and to treat AI as a "thought partner" that must be reviewed and edited.

During audience Q&A, an attendee identified as Tom Petros asked whether responsibility lies with model developers or prompt authors. Gadeau said it is both: users should practice responsible prompting and verification, while model developers must also improve datasets and safeguards. On legal exposure, he acknowledged there are numerous AI-related lawsuits (most commonly about hallucinations and harms), and said hiring-related cases over automated screening have occurred and in some cases settled.

Ortloff closed by directing attendees to the FLCC AI page for slides and follow-up resources, inviting participants to sign up for future webinars and thanking them for joining the inaugural session.

The webinar combined technical explanation with practical guidance: Gadeau emphasized that bias appears in many forms, that mitigation requires both technical fixes and human oversight, and that community input and source auditing are central to producing fairer AI outputs.