[I_Am_Nous]

My coworker trained an LLM on his text messages and was mildly horrified to realize how badly the model "misunderstood" all of the "good natured harassment" he and his contacts give each other. It was shocking enough that he thought "maybe we should tone it down a bit" but it opened his eyes to what an unbiased review of his texts actually looks like.

How are we going to be surprised that an Internet trained model has similar biases? It's not like the Internet is a reliable diversity mixer -- especially in its current "algorithmic bubble" mode.