Superhuman Articulacy as an LLM Safety Target

TL;DR: Current LLMs are bad communicators relative to their agentic capabilities. I claim that articulacy is useful (and perhaps necessary) for AI safety and suggest a path for improving articulacy. Briefly: a theory for articulacy Frequently, LLM agents miscommunicate with their human operators, such as when they write documentation or respond to queries about their activity during a coding session. Any given communication failure can be ascribed to either or both of these two factors: Articulacy Is the model capable of communicating in a precise and human-readable way? Truthfulness Does the model have the propensity to accurately report what it sees, or does it overclaim etc.? Does the model have the propensity to attempt to retrieve more information so it can produce a more accurate output? Does the model have the propensity to inaccurately report what it sees so that it can accomplish some downstream objective? In this document I’ll discuss the first item: articulacy. Truthfulness is its own issue and belongs with the behavioral cloud Ryan Greenblatt describes in “Current AIs seem pretty misaligned to me” . Current LLMs are inarticulate Human operators of coding agents constantly complain about LLM technical writing, in both documentation (e.g., PR descriptions) and in direct communication between the LLM and user. In absence of some coherent theory for this, here’s a list of phenomena mined from my own coding agent history: LLMs will make up jargon for abstractions they…