LLMS AND THE GENERATION OF MODERATE SPEECH

Abstract

For the past year, using large language models (LLMs) for content moderation appears to have solved some of the perennial issues of online speech governance. Developers have promised 'revolutionary' improvements (Weng, Goel and Vallone, 2023), with large language models considered capable of bypassing some of the semantic and ideological ambiguities of human language that hinder moderation at scale (Wang et al., 2023). For this purpose, LLMs are trained to generate “moderate speech” – that is, not to utter offensive language; to provide neutral, balanced and reliable prompt outputs; and to demonstrate an appreciation for complexity and relativity when asked about controversial topics. But the search for optimal content moderation obscures broader questions about what kind of speech is being generated. How does generative AI speak “moderately”? That is: under what norms, training data and larger institutions does it produce “moderate speech”? By examining the regulatory frameworks AI labs, comparing responses to moderation prompts across three LLMs and scrutinising their training datasets, this paper seeks to shed light on the norms, techniques and regulatory cultures around the generation of “moderate speech” in LLM chat completion models.