Bias

Large language models exhibit several types of bias, which, while unintentional, can compound each other and eventually lead to real-world harms.

The biased output in ChatGPT stems from a number of factors. The most significant is likely training data bias (essentially, the human-made material on which the model is trained contains human biases, which the model “absorbs”), but there are factors specific to the models’ architecture and function—as well as other processes—that can add or amplify even more (non-human) bias.