How should AI systems be designed across different interaction phases?

Amershi, Weld, Vorvoreanu, Fourney, Nushi and colleagues at Microsoft Research (CHI 2019) synthesize "over 150 AI-related design recommendations collected from academic and industry sources into a set of 18 generally applicable design guidelines for human-AI interaction." The 18 guidelines are grouped into four categories "based on when during the user's interaction they applied": "Initially" (Guideline 1 and 2), "During interaction" (Guidelines 3 through 6), "When wrong" (Guidelines 7 through 11), and "Over time" (Guidelines 12 through 18). The set was "validated through multiple rounds of evaluation including a user study with 49 design practitioners who tested the guidelines against 20 popular AI-infused products," with guidelines applied or reviewed across "10 product categories" in total.

The authors describe a three-source, multi-phase consolidation. In Phase 1 they gathered candidate guidelines from an internal/external industry review, public articles and editorials, and scholarly HCI papers, arriving at "168 potential AI design guidelines." Three team members affinity-diagrammed these into 35 concepts, then filtered out concepts judged "too vague to design for directly (e.g., 'build trust'), too specific to a particular AI scenario (e.g., 'establish that the bot is not human'), or not AI specific (e.g., 'display output effectively')," leaving 20 guidelines. In Phase 2, a modified heuristic evaluation had 11 team members inspect 13 AI-infused products (email filtering, navigation, e-commerce recommendations, photo organization, a design-assistance feature, a research-assistance feature, a news feed, web and image search) over one-hour sessions looking for applications and violations of each guideline. That round surfaced ambiguity between Guideline 9 ("allow efficient correction") and Guideline 17 ("allow coarse controls") — several evaluators mistook instance-level corrections for global settings — so the authors reworded Guideline 17 to explicitly say "global."

Against the newer notes nearest this one, this paper is the generic, interface-level ancestor they build on or react to. How should AI interfaces handle the shift from doing to supervising? extends the same project — guidelines for designing the interaction surface between a person and an AI system — into agentic territory where plans are negotiated rather than just corrected. What UX principles do workplace users want in AI agents? is a narrower, 2026 descendant of the same consolidation method (recommendations distilled into a numbered principle set, then validated with practitioners) aimed specifically at workplace agents rather than AI-infused systems in general. And Does granting agents more autonomy undermine human oversight? reads, from this paper's vantage point, like a report on exactly the failure modes several of these guidelines were meant to forestall — efficient correction, conveying the consequences of automated actions, and controllable levels of autonomy.

The excerpt available here covers the paper's abstract, introduction, related work, method, discussion, and limitations — it does not include Table 1's actual text for the 18 guidelines, so their individual wording is not established in this note. The authors themselves flag that the guidelines may not transfer well to systems without a graphical interface (voice assistants, activity trackers) or to high-risk, highly regulated domains (semi-autonomous vehicles, robot-assisted surgery, financial systems), and they note an open, unresolved tension between guidelines — conveying the consequences of a system's actions (Guideline 16) can work against letting the system actively learn from user behavior (Guideline 13) — that they say needs further research rather than claiming to have solved. The implication is that this 2019 guideline set should be read as a generic, GUI-oriented baseline that predates today's agentic LLM systems; later agent-specific frameworks are better read as filling gaps this paper itself named than as replacing it.