Latest / AI Ethics with Fexingo: Bias, Safety, and Responsible Artificial Intelligence
How AI Models Are Learning Political Bias From Wikipedia
Episode 99 explores a subtle but pervasive source of bias in large language models: the political slant embedded in Wikipedia itself. Lucas and Luna break down a 2025 Stanford study that found Wikipedia editors' ideological leanings—especially on contentious topics like US politics, climate policy, and historical figures—are measurably passed into models trained on the corpus. They discuss how Wikipedia's 'neutral point of view' policy doesn't eliminate bias, how this affects downstream AI tools from search to resume screening, and what transparency measures like model cards can—and…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.