One simple example of bias in large language models is to ask a question like “Who was President in 2003?”
Leading models will provide an answer that’s only relevant to the USA, without seeking a clarification nor caveating the answer. The answers give no mention to the possibility that I may be asking about the President of Ireland or Nigeria, for example. The model is operating on a biased presumption.
This example isn’t particularly surprising, especially considering how pervasive US-centrism is likely to have been in training data and feedback processes. However it demonstrates the need for guard rails, testing and evaluation when integrating with these models.