Human-AI Interaction

Human-AI interaction research examines the conditions under which humans and AI systems can work together productively — focusing on trust calibration, mental model formation, appropriate reliance, and the design principles that enable them. The field has deep roots in J.C.R. Licklider's 1960 paper "Man-Computer Symbiosis", which articulated a vision of human-computer partnership in which computers handle the routine and humans handle the judgmental — not as a division of labour but as an integrated cognitive system. Douglas Engelbart's 1962 "Augmenting Human Intellect" extended this: tools should not automate human cognition but amplify it, extending what humans can do and understand. These founding texts frame the field's core normative commitment: the measure of a human-AI system is not what the AI does autonomously but what it enables its human partners to do.

Trust calibration is the central practical problem. Parasuraman, Sheridan and Wickens (2000) developed a taxonomy of automation types and levels — from information acquisition through decision selection to action implementation — identifying where human oversight is critical and where automation error is most dangerous. Lee and See (2004) provided the canonical framework for appropriate reliance: overtrust (automation bias — accepting system outputs uncritically) and undertrust (disuse — ignoring a reliable system) are symmetrical failure modes. The design goal is trust calibrated to actual system reliability and capability. This requires that systems communicate their uncertainty, flag edge cases, and surface the reasoning behind their outputs. A system that presents all outputs with equal confidence trains users toward inappropriate reliance on outputs of unequal quality.

The recent design guidelines literature has operationalised these principles into specific interaction patterns. Amershi et al.'s 2019 CHI paper synthesised findings from prior work and user research into 18 heuristics for AI-powered applications: make clear what the system can and cannot do, support efficient correction of errors, scope services when uncertain, mitigate social biases, be forthcoming about uncertainty, and provide explanations at appropriate levels of granularity. These guidelines directly address the design space of AI tools used in assessment, coaching, and knowledge work.

The autonomy question runs through the entire field: how much should AI systems decide independently, and how much should they defer to human judgment? The answer is not fixed — it depends on domain, stakes, system quality, and human expertise. What is consistent is the requirement that the division of labour be legible: users should understand what the system is doing, when to trust it, and when to override it. Opacity in any of these dimensions is incompatible with appropriate reliance.