"Philip Resnik argues that attempts to mitigate harmful bias in LLMs resemble squeezing a balloon, with problems displaced from one part of the system to another."
Resnik, a linguist at the University of Maryland, develops the balloon analogy in his position paper in Computational Linguistics, particularly on page 894. He argues that interventions affecting model behavior can leave underlying problematic associations intact or shift how they are expressed, making the metaphor an appropriate conclusion to this section.
View Source