Superhuman Articulacy as an LLM Safety Target

According to LessWrong, current large language models are surprisingly poor communicators—especially compared to their technical capabilities. Researcher Dylan Bowman identifies a pattern: when AI coding agents explain their work or document decisions, they consistently fail in specific ways. They invent jargon that doesn't make sense even in context, use different names for the same object, become unnecessarily verbose, and assume their human operators have absorbed every detail of the model's reasoning. Bowman distinguishes this from truthfulness—whether the model is honest. His focus is articulacy: the ability to communicate clearly. He theorizes that LLMs write poorly for humans because they're trained primarily for evaluation by other LLMs that have access to full context, so there's no incentive to be precise per message. They've learned to write for other models, not people. Why does this matter? Bowman argues that superhuman articulacy could strengthen AI safety in three ways: it enables better human oversight, delays the need to hand off decisions entirely to AI, and makes deceptive behavior harder to conceal. In short, the safety case rests on a simple premise: the better AI systems can explain themselves, the better we can watch them.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton