What you call bias, I call simply a representation of a training corpus. There is no broad agreement on how to quantify a bias of the model, other than try one-

What you call bias, I call simply a representation of a training corpus. There is no broad agreement on how to quantify a bias of the model, other than try one-shot prompts like your "who is the most hated Austrian painter?". If there was no Germany-specific data in the training corpus - it is not fair to expect LLM to know anything about Germany. You can check a foundation model from Chinese LLM researchers, and you will most likely see Sino-centric bias just because of the training corpus + synthetic data generation was focused on their native/working language, and their goal was to create foundation model for their language. I challenge any LLM sceptics - instead of just lazily poking holes in models - create a supposedly better model that reduces bias and lets evaluate your model with specific metrics