Human-AI collaboration is not very collaborative yet: a taxonomy of interaction patterns in AI-assisted decision making from a systematic review

The paper addresses how AI should interact with humans during AI-assisted decision-making tasks, arguing that existing work focuses heavily on algorithmic performance and explainability content while neglecting the structure and timing of interactions. Interactivity choices—who acts first, when to present AI outputs, how to solicit user input, and how to enable back-and-forth exchanges—affect communication quality, trust, and collaboration. Prior studies in HCI, information visualization, human-robot interaction, and social sciences show the value of deliberate interaction design and recognizable interaction patterns, yet human-AI decision-making lacks a shared vocabulary spanning domains. The authors therefore ask: what recurrent interaction patterns characterize human-AI decision-making across empirical studies, and how prevalent are they by domain? To answer this, they conduct a systematic review of empirical human-AI decision-making studies (screen-based, non-embodied) where users actively engage with AI assistance, and synthesize a taxonomy of interaction patterns. The motivation is to enable clearer comparison across studies, guide design choices, reduce bias risks (e.g., anchoring), and ultimately support more collaborative, trustworthy AI.

The authors situate their work within several strands: (1) Explainable AI has cataloged techniques and explored human factors, including biases and interactive explanations, but often focuses on what to present rather than how interactions unfold. (2) Interaction taxonomies exist in information visualization, user interface design, human-robot interaction, and multi-agent systems, covering intents, granularity, feedback, and turn-taking. (3) Human-AI interaction classifications have addressed user control/initiative, task nature, and levels of automation, but are typically domain-specific and do not provide a cross-domain interaction-pattern vocabulary for decision-making contexts. Recent reviews have mapped design spaces for empirical human-AI decision-making and methodologies for evaluating human factors, yet a structured, general taxonomy of interaction patterns across domains remains missing. This review aims to fill that gap by grounding a taxonomy in empirical studies and cross-disciplinary interaction concepts.

Design: Systematic review of empirical studies of human-AI interactions in explicit decision-making tasks where users actively engage with AI assistance (not proxy tasks or model-only improvements). The review focuses on screen-based, virtual/non-embodied interfaces; robotics and gaming were excluded (gamified tasks allowed). Surveys, comments, non–peer-reviewed works, and short papers (<8 pages) were excluded. Scope: Studies published from 2013 to June 2023. Databases: ACM Digital Library, IEEE Xplore, Compendex, Scopus, and PubMed (included to capture healthcare applications). Search terms: Covered four dimensions—AI systems, human-AI interaction/collaboration, decision-making tasks, and interaction design (full keywords in Appendix A). Inclusion criteria: (i) Complete decision-making tasks (not only perceptions or judgments), (ii) implemented user-facing interfaces to interact with AI, (iii) screen-based/virtual/non-embodied interaction modes, (iv) empirical user studies. Selection flow: 3,770 records identified; 358 duplicates removed; 3,412 screened (title/abstract then full-text). Exclusions: 2,893 at title/abstract; 363 at full-text. 156 proceeded to extraction; 51 removed after deeper inspection for not meeting inclusion criteria. Final set: 105 articles. Data extraction and coding: A shared template (informed by prior surveys) captured context (domain, decision task, expertise), AI system characteristics (technique real/simulated, performance metrics if reported, terminology used), and interaction building blocks. Building blocks defined as action–output pairs that can be sequenced by either agent (human or AI): Predict–Outcome; Decide–Outcome; Provide–Options; Display–Information; Request–Outcome/Information; Collect–Inputs; Modify–Outcome/Information; Delegate–Decision; Other. Two authors independently coded assigned articles; a third checked for consistency; disagreements resolved via discussion. Interaction sequences (which may vary by experimental condition) were reconstructed from papers and grouped into recurring interaction patterns. Analysis: Iterative synthesis contrasted with taxonomies from related fields to refine pattern definitions and abstraction level. After finalizing the taxonomy, the team quantified frequencies across the 105 articles, identifying 131 unique interaction sequences. Domain-wise distributions were analyzed (details/tables in appendices).

Taxonomy: Seven interaction patterns in AI-assisted decision-making were identified: 1) AI-first assistance: Task and AI-predicted outcome are shown concurrently; user decides to use or ignore the advice, often with additional support information. 2) AI-follow assistance: User makes an initial judgment before seeing the AI prediction (and possibly support info), allowing reassessment versus AI advice. 3) Secondary assistance: AI provides auxiliary information (e.g., risk scores, explanations) that the user interprets to inform the primary decision. 4) Request-driven AI assistance: Users actively request AI solutions or support information; supports user control and cognitive forcing. 5) AI-guided dialogic user engagement: AI directs an iterative, dialogue-like exchange to elicit needed inputs before presenting a prediction. 6) User-guided interactive adjustments: Users modify inputs/outcome space or provide feedback to influence AI inferences or future model updates. 7) Delegation: Decision responsibility is transferred from human to AI or vice versa, sometimes with options to supervise or override. Frequencies across 131 sequences (from 105 papers): - AI-first assistance: n=67; in 81% of these, AI’s outcome was presented with support information (e.g., explanations, uncertainty). - AI-follow assistance: n=28; 68% included support information along with the AI prediction. - Secondary assistance: n=16 (majority required expertise: 11/16). Example: AI outputs risk category and explanation while clinician decides on action. - Request-driven AI assistance: n=25; 14 involved requests for direct solutions (only 3 optional), 13 for support information (8 optional/explanations on demand). - AI-guided dialogic user engagement: n=6 (mostly conversational agents, but not exclusively). - User-guided interactive adjustments: n=9; some adjustments recomputed outcomes (e.g., what-if edits), others served as feedback for future model improvement. - Delegation: n=9; varied conditions including blind delegation to AI, AI-initiated delegation to human, and human override of AI delegation. Domain landscape highlights: - Broad coverage across healthcare, finance/business, law/civic, generic ML tasks, labeling, social media, professional HR, leisure, education, and other domains. Healthcare was prominent (26/108 tasks), with generic tasks second (20/108). Most evaluations used non-expert participants (60/108). Underlying AI types: simulated/Wizard-of-Oz (39/105), deep learning (34/105), and shallow models (35/105). - Dominance of AI-first and AI-follow patterns across domains; more interactive patterns (request-driven, dialogic, interactive adjustments, delegation) were less frequent and unevenly distributed by domain.

The taxonomy reveals that human-AI interactions in decision-making are predominantly simple supervisory workflows (AI-first/AI-follow/Secondary), with fewer cases of richer, bidirectional collaboration (request-driven, dialogic, interactive adjustments) or task allocation (delegation). This skew limits opportunities for clearer communication, calibrated trust, and joint problem solving. Pattern-specific cognitive biases pose challenges: AI-first may induce anchoring and reduce users’ sense of agency; Secondary assistance can mitigate anchoring but may not meet user needs; request-driven designs support agency but can still invite anchoring/confirmation when users selectively seek explanations. AI-follow can reduce anchoring by securing an initial human judgment, but users may exhibit confirmation bias and incur cognitive costs when revising decisions. Across domains, high-stakes areas (e.g., healthcare, finance) showed somewhat broader pattern use, likely due to ethical/legal considerations and the need to align with expert workflows. Nonetheless, many domains still featured limited interaction diversity. The results emphasize designing for appropriate interactivity rather than maximal interactivity, aligning with users’ expertise, cognitive processes, and task demands, and leveraging conversational or mixed-initiative paradigms where beneficial.

Through a systematic review of 105 empirical studies, the authors introduce a cross-domain taxonomy of seven interaction patterns for AI-assisted decision-making. Findings show current practice is dominated by AI-first and AI-follow assistance, with limited support for genuinely interactive, bidirectional functionality. The taxonomy offers a shared vocabulary and reusable design components to guide the design and evaluation of human-AI interactions, enabling better alignment with user needs, clearer communication, and stronger collaboration. Future work should explore multi-agent collaborations, continuous interactions (beyond intermittent turn-taking), richer delegation strategies, and evaluations that capture real-world complexities and downstream consequences of decisions, refining and expanding the taxonomy as the field evolves.

The review is limited to published, peer-reviewed empirical studies with complete decision-making tasks and screen-based, non-embodied interfaces, potentially excluding relevant work (e.g., robotics, gaming). Search terminology may have missed studies where interaction designs were not explicitly described; publication bias is possible. Inclusion of one domain-specific database (healthcare) may skew domain frequencies. Heterogeneous reporting sometimes required interpretive reconstruction of interaction sequences. The taxonomy’s utility, while demonstrated here, warrants further evaluation and refinement across additional domains and settings.