Guidelines for Human-AI Interaction
Citation. Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P. N., Inkpen, K., Teevan, J., Kikin-Gil, R., & Horvitz, E. (2019). Guidelines for Human-AI Interaction. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, 1–13. https://doi.org/10.1145/3290605.3300233
Summary. This CHI paper consolidates more than 150 prior academic and industry recommendations into 18 design guidelines for AI-infused systems, organized around what users need initially, during interaction, when systems are wrong, and over time. It iteratively assessed the guidelines through internal heuristic evaluation, a study of 49 HCI/design practitioners evaluating 20 popular AI-infused products, and an expert evaluation of the revised set.
Key claims
- The 18 guidelines specify interaction requirements spanning capability and error-rate communication; context, social norms, and bias; easy invocation, dismissal, correction, uncertainty scoping, and explanations; and memory, personalization, adaptation, feedback, global controls, and change notification.
- In the 49-participant heuristic-evaluation study, participants identified 785 adjusted examples across 20 products: 313 applications, 277 violations, 89 neutral ratings, and 106 “does not apply” ratings; every guideline had at least one application or violation in every product category tested.
- Guidelines can be broadly relevant without being equally observable or implemented: evaluators found some guidance difficult to assess in a single session or without access to underlying algorithms, and products involving voice, images, or sensor data produced more “does not apply” responses.
Evidence & limitations
- This is a peer-reviewed CHI proceedings paper that uses iterative design-guideline development and heuristic evaluation, not a field experiment testing whether implementation of the guidelines causally improves worker outcomes, system safety, or decision quality.
- The central user study recruited HCI/UX-experienced participants, primarily through internal mailing lists, and assessed 20 consumer-facing AI-infused products; its evidence supports relevance and clarity of the guidance rather than general effectiveness in contemporary agentic or workplace systems.
- The 2019 guidelines predate current LLM and agentic workflows; they remain a design baseline but need contextual adaptation and empirical testing for systems that plan, act across tools, or carry extended accountability consequences.