Microsoft: Human-AI Interaction and HAX

Evidence class: research-derived guidelines with practitioner evaluation, plus maintained application guidance. Read: 2026-09-13. Website transfer is limited to interfaces that expose AI-generated suggestions or delegated work.

What was evaluated

Amershi and colleagues' CHI 2019 publication proposes 18 guidelines. Its author-hosted abstract reports evaluation involving 49 design practitioners and 20 AI-infused products. This pass verified the publication record and abstract, not the full paper's methods. That evaluation is not a benchmark proving current agent products safe or usable. Source: Guidelines for Human-AI Interaction, read 2026-09-13.

What the maintained library holds

The HAX (Human-AI eXperience) Design Library presents the same 18 guidelines, G1 to G18, each with design patterns and examples. They cover what the system can do and how well (G1, G2), timing, relevance, social norms and social biases (G3 to G6), invocation, dismissal and correction (G7 to G9), scoping when in doubt and explaining behavior (G10, G11), and memory, learning, cautious updates, feedback, consequences, global controls and change notices (G12 to G18). Source: HAX Design Library, read 2026-10-29. The guideline list was checked against the library; the individual design patterns were not verified.

Our website application

Distinguish suggestions from authorized work and completed effects. Let the user inspect important changes and correct an unsuitable result. Keep access, recipients, and irreversible consequences understandable. Offer cancellation or reversal only where the actual system supports it. These state and review choices are proposed applications, not additional findings of the 2019 study.

Tensions and scope

Review can support control while creating interruption. More explanation is not necessarily better checking, and an explanation is not evidence that an answer is correct. Test whether people detect consequential errors, not only whether they report trust. Ordinary static pages do not need an AI control panel. Chat versus structured review remains a task-specific design hypothesis.