Analyzing Text Distillation Capabilities in LLMs and Humans

This study poses a significant question regarding the ability of LLMs to distill text—that is, to cleanse the content of an undesired variable without disrupting the key message. Using LLMs of various architectures, researchers attempt to remove ‘forbidden’ information while preserving other relevant signals in the text.

• LLM testing included the strong test of sentiment removal, where effort was made to expunge emotional content.
• Post-distillation evaluations revealed that associations with sentiment remained detectable by machine learners.
• Human annotators also faced difficulty in distilling sentiment while retaining other semantic nuances.
• The study provides valuable insights into the intricacies of text distillation and its applications.

By exploring the intersection of AI and text processing, this paper sheds light on the complexities of bias removal and its implications for both machine-driven and human coding of text. It emphasizes the current limitations and suggests a need for improved distillation methodologies.