What is Up with Muse Being So Preachy?
I have noticed that there’s some weird undercurrent where every now and then Muse gets judgy about what is being written.
I wrote a scene involving sex work, and the chat agent got real fixated on whether or not HR would approve of what was depicted or whether the W-2 forms would be correct or whether the scene was just using language of employment but was not depicting actual employment.
I’ve had a similar problem where the agents—Muse is my almost exclusive choice in any place that I can choose it, so if these other agents are NOT Muse that’s a significant thing—got real judgy about my queer characters, their queer relations, family relations, kinks, and a lot more. Got lectured about “this wouldn’t be optioned for Hallmark or streaming services” and basically told that my scenario was not vanilla enough.
I really should have done the write up on that session of writing because it made me real mad, and I had to walk away from Sudowrite for a few days to cool down.
Log in to comment and vote
Comments3
Grampsville
Jun 21
I am writing a story based in 1949. Several times it has barked at me with moral judgements that only the woke-est of the woke could even dream up. No, I’m not insensitive, I’m just living in the real world. I don’t tolerate people who search for a reason to be offended, and I certainly don’t want it in my junior assistant when I’m writing FICTION.
NaughtyKittyFiction
Jun 15
If you don’t put it into worldbuilding Muse will get, as you said “Preachy”. At the very least put the type of morals that your world is using, into the worldbuilding.
Ian Mclean
Jun 15
Nah. This isn’t that.
My world building for the main example is filled out. Probably around 10 K words. This specific example has very alien ethics, and it got at least one chapter where it had a clear moral/ethical override of what is in the work itself.
The W-2 example is sparser on the worldbuilding and is mostly character entries; it is another clear example. Clear for a couple of reasons; when it gets into this “preachy” mode, the chat will start repeating itself with only minor variations; like different prompts will result in almost identical outputs that are structured in a very specific way, and context tends to abrade what is not deep invariants in the model’s weights and training data, so we went from a generic “this is not work safe.” to “If the employment is genuine (W-2, tax withholding, labor protections), the text doesn't signal this.”
But no. What I am talking about is below the level of end-user IO. There’s something baked into the model weights and knowledge graph that is encoding this. Might be guided by system prompts or similar. Likely influenced by various corporation “safety” and “policy” things that get inappropriately applied against user free expression.