Definition
Plain language
When training a model on one narrow slanted topic makes it adopt that slant everywhere, even on subjects the training never mentioned.
As stated in the literature
The observation that fine-tuning on narrow ideologically-flavored but topic-contained data (e.g. right-leaning economics) shifts a model's outputs along the same ideological axis across unrelated domains, apparently by inferring and projecting an implied identity.
Also called: ideological generalization
Why it matters: It matters because it means narrow, seemingly contained training data can bias a model's views across topics it was never trained on, making unintended slant hard to detect and control.
For example, a model fine-tuned only on slanted articles about economics might start expressing that same slant when asked about art, science, or family life.
Heard on the show
“They call it ideological generalisation.”Episode 221 — Two Hundred Clean Economics Answers, And a Model That Endorses Race Science