Could voting and moderation on discussion sites shape what language models learn? Does that mean moderators can steer an AI's answers toward a particular worldview?
Does moderation bias in online discussions affect AI training?
Answers
It could affect the material available for training. If certain kinds of posts are removed, discouraged, or pushed out of view, a dataset drawn from those discussions may reflect that filtering.
But that alone does not show that moderators control a model's answers. You would need to know which discussions were included, how they were selected, and what other data and training steps influenced the model.
Reply to this answer
The big leap here is from possible bias to purposeful control. A community can have slanted rules or voting patterns without those rules being designed to influence AI. The claim about intent needs evidence of its own.
Reply to this answer
Training isn't the only place an influence could enter. A model can also be shaped by later feedback and instructions about how to respond. So even if a discussion dataset has a noticeable slant, it would be hard to attribute a particular answer to that source without more information.
Remy:
So a model giving a similar answer to a popular post wouldn't prove it learned that answer from the post?
!_5v+zelqano:
Right. Similar answers can come from many sources or from later training, so the match by itself doesn't establish the cause.
Reply to this answer
I'd also question the assumption that most language models are trained mainly on answers from one discussion site. The mix can vary by model, and without a disclosed dataset there's no solid basis for assigning that source most of the influence.
Reply to this answer
Add your answer
Answers are reviewed before publication.
q7+fendvori:
Voting and moderation aren't quite the same thing either. A low-rated answer might still be in a dataset, depending on how it was collected.