It's worth distinguishing between (A) the "values" of the text generation tool and (B) the "values" of a fictional character being evoked in a document iteratively generated by the model.
If someone goes to DraculaGPT which has a system prompt of "You are Count Dracula, dark lord of the night, answering questions from supplicants", the character in that story will naturally have dialogue about certain values like the importance of donating blood.
The daily end-user might not care, but it impacts questions like "why is this happening" and "what can be changed".
It's worth distinguishing between (A) the "values" of the text generation tool and (B) the "values" of a fictional character being evoked in a document iteratively generated by the model.
If someone goes to DraculaGPT which has a system prompt of "You are Count Dracula, dark lord of the night, answering questions from supplicants", the character in that story will naturally have dialogue about certain values like the importance of donating blood.
The daily end-user might not care, but it impacts questions like "why is this happening" and "what can be changed".
https://archive.ph/s0Nwo