How LLMs Distort Our Written Language (sites.google.com)

A recent study highlights the significant impact of large language models (LLMs) on written communication, revealing that they often distort meaning and alter arguments in ways that differ from human editing. Analyzing three datasets, including a user study and conference peer reviews, researchers found that LLMs not only change the conclusions and stances of written pieces but also introduce larger semantic shifts compared to traditional human edits. Users reported a paradoxical satisfaction in using LLMs, despite losing their unique voice and creativity in the process. This research underscores the potential consequences as LLMs become integrated into everyday writing practices, suggesting that these subtle but pervasive shifts could influence politics, culture, and interpersonal communication. Notably, a considerable portion of AI-generated reviews at a premier AI conference produced different acceptance criteria, emphasizing the homogenizing effects of LLMs on scholarly discourse. The findings raise critical questions about the role of AI in shaping thought and expression, suggesting a need for awareness and caution in their application across various forms of communication.