Toward a Unified Framework for Collaborative Design of Human-AI Interaction

A new paper from Ankur Bhatt and Sven Mayer introduces a unified framework for designing human-AI interaction that treats multimodal alignment, explainability, and human agency as core, interdependent principles rather than separate bolt-ons. As AI systems increasingly interpret user intent through speech, gesture, and gaze, users rarely understand how those interpretations are made, eroding trust and control. The framework addresses this by requiring: 1) accurate multimodal intent interpretation, 2) interaction-centric explainability that delivers real-time visual, textual, and audio feedback, and 3) agency-preserving mechanisms that let users accept, reject, or modify any AI suggestion at any decision point.

The authors validated the framework through two contrasting scenarios: collaborative design (low time pressure, reversible errors) and extended reality warehouse robot collaboration (high time pressure, safety-critical misinterpretations). The warehouse scenario highlights real stakes—misreading a worker’s gesture could cause injury or damage. By framing collaboration as a continuous property rather than a one-time setup, the approach ensures that as AI becomes more proactive, user oversight remains a first-class design element. This reframing benefits designers building transparent systems, researchers studying interaction, and end users who need to trust AI in high-stakes environments.