The Illusion of Consensus in Human-Centered Interactive AI

Vinicius Buri Lux

Keywords: Human-centered AI, Interactive AI, Human-Robot Interaction, Trust Calibration, Explainability, Multi-Agent Systems, Consensus Illusion

TL;DR: Agreement across multiple AI agents can be misleading: apparent consensus may reflect shared architecture rather than independent validation, creating systematic trust calibration failures in human-centered interactive AI systems.

Abstract: As human-centered AI systems increasingly integrate multiple interactive components—such as perception, planning, language, and decision modules—users are often encouraged to interpret agreement across system outputs as a signal of increased reliability. This position paper challenges that assumption. Drawing on empirical observations of high-fidelity semantic convergence across independent large language model (LLM) instances (N = 17, convergence > 95%), we document what we term the Illusion of Consensus: a phenomenon in which apparent agreement across AI components or agents emerges not from independent verification, but from shared architectural substrates. We argue that this structural consensus creates a systematic trust calibration failure in interactive AI systems, leading users to over-rely on apparent validation that may reflect redundancy rather than epistemic robustness. We discuss implications for the design of human-centered interactive AI and human–robot interaction, particularly in safety-critical contexts where users must assess the reliability of complex, multi-component systems. Rather than proposing solutions, this paper articulates the problem and calls for new interaction design strategies that distinguish structural alignment from genuine independent agreement, enabling more transparent and trustworthy human–AI collaboration.

Submission Number: 7