Speakers highlight real‑world harms from biased training data

Gadeau reviewed several documented examples of biased outcomes from model training. He pointed to the Amazon resume‑screening project that favored men because the company's historical hiring patterns informed the model, and to high‑profile chatbot failures where social media training caused rapid adoption of toxic or antisemitic language.

He also discussed judicial risk‑assessment tools and medical imaging failures where dataset artifacts produced incorrect results. "This is how biased datasets can give you biased results," Gadeau said, citing cases where models produced high risk scores for some defendants while downplaying risk for others with different criminal histories.