How should we categorize different types of human-AI interaction?

Proposing principles for interactive AI evaluation, the paper identifies "five prototypical modes of humanmodel interaction" drawn from "observed use cases and literature reviews": Collaborative ("human and model work in tandem towards completing joint goal-oriented tasks (e.g., human and model iteratively write and refine a report together)"), Directive ("human delegates specific tasks to the model for independent completion (e.g., human instructs model to generate a marketing campaign)"), Assistive ("model provides supporting input while human maintains primary agency (e.g., model suggests edits while human writes a document)"), Cooperative ("human and model make distinct contributions toward a shared goal without directly working together (e.g., model generates data visualizations while human writes analysis)"), and Explorative ("human engages in open-ended interaction without specific task goals (e.g., casual conversation or creative brainstorming)").

The taxonomy is built to serve a specific purpose, not to classify interaction generally: the paper's first evaluation principle is that "user goals or objectives have been shown to shape how users engage with systems and thus influence the outcomes of these interactions," so scenario design for interactive evaluation needs both a categorization of harmful scenarios by user objective and a categorization of interaction modes, because "a single generative model can support multiple interaction modes." The five modes vary along two implicit axes — who initiates and retains agency (human vs. model), and whether the human and model work toward the goal jointly or separately — which the paper treats as the dimensions an ecologically valid evaluation scenario must reproduce rather than abstract away.

This sits alongside How should AI systems be designed across different interaction phases? as a second attempt to impose structure on human-AI interaction design, but the two cut the problem differently: Amershi and colleagues organize guidance by when in an interaction a concern arises, while this taxonomy organizes by what kind of engagement is happening regardless of phase. It also gives the scenario-design vocabulary that Do static AI evaluations miss harms that build over time? calls for but does not itself supply — that paper argues evaluation scenarios must be ecologically valid without specifying how to construct them; this taxonomy is the mechanism. The Collaborative mode also connects to Can AI teammates assess collaboration without losing naturalness?, which addresses how to evaluate exactly that mode of joint, goal-oriented work without sacrificing experimental control.

The excerpt presents the taxonomy as a design tool for evaluation scenario-building, not as a validated classification: it cites no study testing whether the five categories are exhaustive, mutually exclusive, or reliably distinguishable by raters, and it does not report how real interaction logs distribute across them. The modes should be read as a proposed organizing scheme for evaluation design, not as an established empirical finding about how people actually use generative AI systems.