Dunning-Kruger and the Communication Tax (danieltan.weblog.lol)

A recent exploration into the Dunning-Kruger effect highlights the complexities of communication between humans and large language models (LLMs). As humans seek to interact with LLMs, they often underestimate the models’ domain knowledge, leading to a communication “tax” where LLMs may overly simplify information for users lacking expertise. Conversely, when experts engage with LLMs, the burden falls on humans to navigate the model’s responses. This asymmetry points to a significant challenge in achieving effective human-LLM interaction, especially as current calibration mechanisms—like user preference prompts—struggle with mismatched expectations and can lead to a breakdown in communication. The framework presented analyzes communication costs quantitatively, focusing on capability gaps between sender and receiver, and proposes ways to operationalize these costs. By introducing variables such as specification tax, re-prompting rates, and quality gaps, the study suggests that optimizing for mutual understanding in human-LLM interactions requires a shift in targets toward minimizing mismatches in expectations over time, rather than merely averaging user experiences. This approach not only underscores the limitations of existing calibration strategies but also sets the stage for more nuanced designs that prioritize effective communication while navigating the inherent asymmetries in knowledge and understanding.