The first risks we consider are the risks that follow from the LMsabsorbing the hegemonic worldview from their training data. Whenhumans produce language, our utterances reflect our worldviews,including our biases [
It is important to remember that LMs are a reflection of human input, and therefore, human error. Our individual experiences create subconscious biases that make it impossible to deliver an unbiased LM system.