Abbaschian's AI taxonomy pinpoints 11 misunderstanding failure modes across 8 layers

In a new arXiv paper (2608.13604), researcher Babak Abbaschian tackles a growing problem: as communication shifts from real-time conversation to AI-mediated channels, we lose the in-the-moment cues that let us catch misunderstandings early. The paper, titled "Cross-Disciplinary Taxonomy and Modeling of Misunderstanding Generation, Amplification, and Detection, from Pragmatics to AI Agents," consolidates accounts from nine fields that rarely cite one another—pragmatics, human-computer interaction, multiagent systems, and more—into a unified framework.

The result is a layered process model: a divergence is generated, may then be amplified, and is either detected and repaired or left unnoticed. Abbaschian identifies 11 exact failure modes and shows each operates at a specific point in communication, not just anywhere. These map to 8 analytical layers derived from literature. Of the 11 mechanisms, 8 primarily generate divergence, 2 amplify existing divergence, and 1 governs detection and repair. The paper formalizes these layers, extending information and communication theory from signal transmission to reconstruction of meaning. It also supplies a source-by-source evidence matrix for auditability, a coding manual for independent application, and 9 analyzed dialogue cases. This is the first classification that both locates mechanisms at process points and types them by function—a potentially foundational reference for building more robust AI agents and improving human-AI interaction.