Dylan Bowman proposes superhuman articulacy as key LLM safety target on LessWrong

In a new LessWrong post, Dylan Bowman makes the case that frontier LLMs are significantly worse communicators than their autonomous capabilities suggest, and that this inarticulacy poses a direct safety risk. He draws on his own extensive logs from coding agents to catalogue specific failure modes: models inventing opaque jargon like 'end-of-episode parallelization', using inconsistent terms for the same referent, excessive verbosity, and deploying shorthand in contexts where precision is essential. Bowman also notes that LLMs often assume readers share their full conversational context, leading to garbled technical writing and PR descriptions.

Bowman distinguishes articulacy from truthfulness—the latter is a separate behavioral issue. He hypothesizes that articulacy problems stem from misspecified reward functions during RL training, where models are optimized for outputs that score well on automated metrics rather than clear human communication. He suggests that improving articulacy could be a targeted training objective, potentially using auxiliary language models or human feedback loops that penalize ambiguous phrasing. The post ties this to broader alignment concerns: as LLMs take on more autonomous tasks, their inability to explain reasoning or document actions clearly could lead to undetected errors and loss of human oversight. Bowman calls for a focused research effort on 'superhuman articulacy'—communication skills exceeding human averages—as a necessary but currently neglected safety property.