AI's integration into various sectors has raised significant concerns about bias. Tessa Charlesworth and William Brady, researchers from Kellogg, highlight that bias enters AI systems from multiple sources, not just the data used to train them. Annotators, often a homogenous group, may overlook biases that affect marginalized communities. Moreover, biases can amplify during the AI's optimization process, where decisions about what to prioritize can skew outputs, leading to negative societal impacts like polarization. Users often trust AI to be less biased than humans, but this assumption can result in misplaced confidence. Charlesworth points to several biases that are particularly problematic, such as the portrayal of women in training data as more passive than men, and biases linked to race, sexuality, and health. One less-discussed issue is the dominance of English in AI training data, which carries its own set of biases. This linguistic bias not only underrepresents non-English speakers but also perpetuates historical biases associated with dominant languages. The researchers emphasize that human involvement at every stage of AI development allows for the introduction of these biases, undermining the perception of objectivity in AI outputs.