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

1 INTRODUCTION

• Human-AI interaction design must consider how and when AI communicates, alongside what information it provides, to align algorithmic suggestions with human expectations.The paper links deliberate interaction choices with clear communication, trustworthiness, and meaningful collaboration.Source: §1 · p. 2
• Interaction choices should be deliberate: adding interactivity is not inherently beneficial unless it serves the intended goals.The paper connects appropriate interaction balance with clear communication, trustworthiness, and meaningful collaboration.Source: §1 · p. 2
• Empirical studies span domains, but their varied interaction configurations lack a common language, limiting synthesis and exploration of new paradigms.Existing AI interaction classifications tend to focus on particular domains or contexts rather than providing a general framework.Source: §1 · pp. 2–3
• The authors reviewed 105 articles and used the observed interaction trends to develop a taxonomy of seven human-AI interaction patterns.The review also considered task context, AI systems, and evaluation methods.Source: §1 · p. 3

2.1 Search strategy and selection criteria

• The search covered five databases and studies published from 2013 through June 2023, using terms spanning AI, human-AI interaction, decision-making, and interaction design.The databases were ACM Digital Library, IEEE Xplore, Compendex, Scopus, and PubMed.Source: §2.1 · p. 3
• Studies had to evaluate complete decision-making tasks, present an implemented interface, and use empirical user studies.The selection excluded studies limited to perceptions of AI advice rather than completing a decision task.Source: §2.1 · p. 4
• Eligible interactions used screen-based interfaces, virtual agents, or non-embodied setups; robotics and gaming studies were excluded from the search.Gamified decision-making tasks were retained, while studies involving gaming scenarios were excluded as more complex and less constrained for studying interaction.Source: §2.1 · p. 4

2.2 Study selection

• 105 articles remained after screening 3,412 records, including full-text review and a detailed check that candidate studies supported actual AI-assisted decision-making tasks.The initial search returned 3,770 papers; 358 duplicates were removed, and 51 of 156 extraction-stage papers were excluded after detailed reading.Source: §2.2 · p. 4

2.3.1 Analysis process.

• Two authors independently coded assigned articles, checked extraction consistency, and discussed discrepancies before iteratively grouping interaction sequences into design patterns.The authors refined each pattern’s component interactions through discussion.Source: §2.3.1 · p. 4

2.3.2 Context.

• For each study, the review recorded the decision-making domain, task, and expertise required of the human decision-maker.The domain categories were adapted from a prior survey of AI-assisted decision-making.Source: §2.3.2 · p. 4

2.3.3 AI system.

• The review characterized each AI system’s purpose, recommendation technique, reported performance, and terminology used to introduce it to participants.It did not focus on details such as data type and source, which had been covered in prior work.Source: §2.3.3 · p. 5

2.3.4 Evaluation.

• The reviewers recorded study constructs and classified dependent variables as objective or subjective according to how they were assessed.The extraction focused on how people perceived AI assistance and the outcomes of the interaction.Source: §2.3.4 · p. 5

2.3.5 Interaction building blocks.

• Interaction patterns were built from action-output pairs, treated as granular components that could be combined into sequences and triggered by one another.The sequence and responsible agent were identified from the studies rather than preset in the extraction template; plausible sequences could differ with experimental manipulations.Source: §2.3.5 · pp. 5–6
• The coded building blocks include predicting or deciding an outcome, providing options, displaying information, and requesting information or solutions.Prediction may be performed by either agent; providing options and displaying supplementary evidence are typically AI actions, while deciding the final outcome is typically the user’s action.Source: §2.3.5 · p. 5
• Other building blocks capture collecting task inputs, modifying a counterpart’s solution, and delegating responsibility, with delegation ranging from surrendering agency to supervision.Collecting inputs is typically done by the user, while modification is typically optional and user-executed; either agent may delegate.Source: §2.3.5 · pp. 5–6

3.1 Taxonomy of interaction patterns for AI-assisted decision making

• AI-first presents the AI outcome with the decision problem, whereas AI-follow elicits a preliminary human judgment before presenting AI advice for comparison and reassessment.These patterns are also described as the concurrent and sequential paradigms, respectively.Source: §3.1 · pp. 6–7
• Secondary assistance supplies information users must interpret rather than a direct solution; request-driven assistance lets users choose when to seek AI information or solutions.Secondary support may include a risk estimate relevant to a separate decision, while request-driven support can allow more time for deliberation.Source: §3.1 · pp. 7–8
• Dialogic engagement iterates AI instructions and user-provided information, while user-guided adjustments let people change the AI’s outcome space or provide feedback.Adjustments may update interface displays or provide feedback intended to improve underlying models.Source: §3.1 · p. 8
• Delegation assigns task leadership according to agents’ relative capabilities, while “Others” captures combinations and more complex interactions; the categories are not mutually exclusive.The taxonomy was formed through automated searches for action-output pairs followed by manual inspection and verification.Source: §3.1 · pp. 6, 8

3.2 Identification of interaction patterns in AI-assisted decision making studies

• 67 sequences used AI-first assistance, compared with 28 using AI-follow assistance, making them the most common patterns among 131 analyzed sequences.AI solutions appeared with additional information in 81% of AI-first and 68% of AI-follow interactions.Source: §3.2 · p. 8
• Secondary assistance appeared in 16 sequences, 11 of them in expertise-requiring tasks; “Others” included multiple outcomes, continuous tasks, and third agents.Other cases also averaged independent human and AI solutions as a final verdict.Source: §3.2 · pp. 8–10
• Less common patterns included 25 request-driven sequences, six dialogic interactions, nine user-guided adjustment sequences, and nine delegation sequences.Request-driven interactions could involve requests for direct solutions or supporting information.Source: §3.2 · p. 9

3.3 Landscape over domains evaluated in AI-advised human decision making

• Most studies addressed real-world applications, fewer than 15% used artificial tasks, and healthcare was the largest domain with 26 of 108 tasks.Generic tasks were the second most common domain, with 20 of 108; non-expert users participated in 60 of 108 tasks.Source: §3.3 · p. 10

4 DISCUSSION

• Some reviewed papers did not describe interactions well enough for complete recreation, requiring the authors to interpret and encode them.Source: §4 · p. 13
• AI-first assistance can invite unreflective acceptance or dismissal and anchoring, while Request-driven assistance offers agency but can still invite confirmation or anchoring bias.Only 10 identified AI-first occurrences let users request AI inferences.Source: §4 · p. 13
• AI-follow assistance gives users an initial independent decision, but users may retain provisional judgments, and re-evaluation can increase cognitive costs.One cited study found participants rarely revised provisional diagnoses when AI inferences differed.Source: §4 · p. 14
• Studies used diverse interaction patterns only sparsely across domains; which pattern suits non-experts remains unclear, while domain expertise may make Secondary assistance useful.Ethical and legal considerations also shape design choices in high-stakes domains.Source: §4 · p. 14
• Most common patterns limit humans to supervising predictions, while less frequent designs add dialogue, user adjustments, or shared decision-making and delegation.The paper describes collaboration ranging from parallel work to coordinated turn-taking.Source: §4 · p. 14
• The review identifies gaps in continuous interaction, multi-agent teamwork, coordination of delegated tasks, and evaluation in realistic settings with decision consequences.Many studies use artificial tasks and end after decisions, so users do not encounter the full consequences of their choices.Source: §4 · p. 15
• Interaction design should account for user psychology, agency, and the time and effort associated with cognitive costs.Source: §4 · p. 15

4.1 Limitations

• The taxonomy is preliminary, and its review is limited to published empirical studies of screen-based AI assistance, with search and publication bias also possible.The authors note that embodied AI may support additional interactions and that the framework needs further expansion and refinement.Source: §4 · p. 13, §4.1 · p. 15

5 CONCLUSION

• A systematic review of human-AI interactions in AI-advised decision tasks informed the formulation of an interaction-pattern taxonomy.Source: §5 · p. 15
• Current practices often lean toward AI-driven or human-led decisions, with limited emphasis on interactive functionality throughout the interaction.Source: §5 · p. 15
• The taxonomy is presented as a resource for designing AI decision-support systems and more productive, engaging, user-centered human-AI collaboration.Source: §5 · p. 15

A SEARCH STRATEGY

• The query was used to screen titles, abstracts, and keywords in all consulted databases.Source: A SEARCH STRATEGY · p. 20
• The search string combined terms for AI, human-AI collaboration, decision-making tasks, and interaction design.Source: A SEARCH STRATEGY · p. 20
• Articles with specified review-related or robot, gaming, and vehicle terms in their title and abstract were excluded.Source: A SEARCH STRATEGY · p. 20

• Metrics are grouped by domain, the construct measured, and whether the measure is objective or subjective.Source: Table 2 · p. 20
• Evaluation measures cover efficacy, including decision accuracy and performance, and efficiency, including task duration and decision time.Source: B SURVEY FINDINGS · p. 21
• Trust and reliance measures include agreement with AI, delegation, advice weighting, compliance, overtrust, and undertrust.Source: B SURVEY FINDINGS · p. 21
• Other measured constructs include understanding, fairness, and decision risk perception.Source: B SURVEY FINDINGS · pp. 21, 23
• System usability measures include usefulness, satisfaction, acceptance, and perceived adoption or implementation, while decision satisfaction and mental-demand measures include confidence and task difficulty.Source: B SURVEY FINDINGS · pp. 21–22