include associating certain professions with a particular gender or race when generating text, or presenting historical narratives from only one group’s viewpoint. Bias in AI systems can arise at various stages of system development—from social stereotypes embedded in training data to imbalances in data collection—and such bias may be reproduced during the system’s use. To address this, technical and ethical approaches such as diversifying training data, improving algorithms, and conducting continuous monitoring are commonly explored. However, just as the bias of an individual neuron is a fixed constant determined during training, the bias of an AI system can also be understood as a type of positionality inherent to how the system is built. In that sense, creating an AI system completely free of bias is an unattainable goal. Nevertheless, it is important to design technical systems in ways that prevent harmful forms of bias—such as the exclusion of marginalized groups—from occurring. Temperature Temperature is a hyperparameter that controls the diversity (randomness) of outputs produced by a generative AI model. In a text generation model, when predicting the next word, a higher temperature increases the likelihood of selecting less probable words, resulting in more creative or varied outputs. Conversely, a lower temperature favors the most likely words, producing outputs that are more predictable. 24 25