Papers
Topics
Authors
Recent
Search
2000 character limit reached

Deconstructing Human-AI Collaboration: Agency, Interaction, and Adaptation

Published 18 Apr 2024 in cs.HC and cs.AI | (2404.12056v1)

Abstract: As full AI-based automation remains out of reach in most real-world applications, the focus has instead shifted to leveraging the strengths of both human and AI agents, creating effective collaborative systems. The rapid advances in this area have yielded increasingly more complex systems and frameworks, while the nuance of their characterization has gotten more vague. Similarly, the existing conceptual models no longer capture the elaborate processes of these systems nor describe the entire scope of their collaboration paradigms. In this paper, we propose a new unified set of dimensions through which to analyze and describe human-AI systems. Our conceptual model is centered around three high-level aspects - agency, interaction, and adaptation - and is developed through a multi-step process. Firstly, an initial design space is proposed by surveying the literature and consolidating existing definitions and conceptual frameworks. Secondly, this model is iteratively refined and validated by conducting semi-structured interviews with nine researchers in this field. Lastly, to illustrate the applicability of our design space, we utilize it to provide a structured description of selected human-AI systems.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (49)
  1. Effects of communication directionality and AI agent differences in human-AI interaction. In Proc. of the SIGCHI Conf. on Human Factors in Computing Systems (2021), ACM, pp. 1–15.
  2. Guidelines for human-AI interaction. In Proc. of the SIGCHI Conf. on Human Factors in Computing Systems (2019), ACM, pp. 1–13.
  3. Comparing visual-interactive labeling with active learning: An experimental study. IEEE Trans. on Visualization and Computer Graphics 24, 1 (2018), 298–308.
  4. Brehmer M., Munzner T.: A multi-level typology of abstract visualization tasks. IEEE Trans. on Visualization and Computer Graphics 19, 12 (2013), 2376–2385.
  5. A provenance task abstraction framework. IEEE Trans. on Visualization and Computer Graphics 39, 6 (2019), 46–60.
  6. Characterizing guidance in visual analytics. IEEE Trans. on Visualization and Computer Graphics 23, 1 (2017), 111–120.
  7. A review of guidance approaches in visual data analysis: A multifocal perspective. Computer Graphics Forum 38, 3 (2019), 861–879.
  8. What did my AI learn? How data scientists make sense of model behavior. In Proc. of the SIGCHI Conf. on Human Factors in Computing Systems (2023), vol. 30, ACM, pp. 1–27.
  9. Intellingo: An intelligible translation environment. In Proc. of the SIGCHI Conf. on Human Factors in Computing Systems (2018), ACM, pp. 1–13.
  10. “Hello AI”: Uncovering the onboarding needs of medical practitioners for human-AI collaborative decision-making. In Proc. of the SIGCHI Conf. on Human Factors in Computing Systems (2019), vol. 3, ACM, pp. 1–24.
  11. Hybrid intelligence. Business & Information Systems Engineering 61, 5 (2019), 637–643.
  12. Dudley J. J., Kristensson P. O.: A review of user interface design for interactive machine learning. ACM Trans. on Interactive Intelligent Systems 8, 2 (2018), 1–37.
  13. El-Assady M., Moruzzi C.: Which biases and reasoning pitfalls do explanations trigger? Decomposing communication processes in human-AI interaction. IEEE Trans. on Visualization and Computer Graphics 42, 6 (2022), 11–23.
  14. A grammar-based approach for applying visualization taxonomies to interaction logs. Computer Graphics Forum 41, 3 (2022), 489–500.
  15. Gotz D., Zhou M. X.: Characterizing users’ visual analytic activity for insight provenance. Proc. of IEEE Conf. on Visual Analytics Science and Technology (2008), 123–130.
  16. VATLD : A visual analytics system to assess, understand and improve traffic light detection. IEEE Trans. on Visualization and Computer Graphics 27, 2 (2021), 261–271.
  17. Gamut: A design probe to understand how data scientists understand machine learning models. In Proc. of the SIGCHI Conf. on Human Factors in Computing Systems (2019), ACM, pp. 1–13.
  18. Horvitz E.: Principles of mixed-initiative user interfaces. In Proc. of the SIGCHI Conf. on Human Factors in Computing Systems (1999), ACM, pp. 159–166.
  19. Artificial intelligence and machine learning. Electronic Markets 32, 4 (2022), 2235–2244.
  20. Towards a science of human-AI decision making: An overview of design space in empirical human-subject studies. Proc. of the ACM Conf. on Fairness, Accountability, and Transparency (2023), 1369–1385.
  21. The state-of-the-art in predictive visual analytics. Computer Graphics Forum 36, 3 (2017), 539–562.
  22. A human-AI collaborative approach for clinical decision making on rehabilitation assessment. In Proc. of the SIGCHI Conf. on Human Factors in Computing Systems (2021), ACM, pp. 1–14.
  23. Bridging from goals to tasks with design study analysis reports. IEEE Trans. on Visualization and Computer Graphics 24 (2018), 435–445.
  24. Where are we so far? Understanding data storytelling tools from the perspective of human-AI collaboration. In Proc. of the SIGCHI Conf. on Human Factors in Computing Systems (2024), ACM.
  25. A review on human–AI interaction in machine learning and insights for medical applications. Int. Journal of Environmental Research and Public Health 18, 4 (2021), 2121.
  26. Human-computer collaboration for visual analytics: An agent-based framework. Computer Graphics Forum (2023).
  27. RLHF-blender: A configurable interactive interface for learning from diverse human feedback. Interactive Learning with Implicit Human Feedback Workshop at ICML (2023).
  28. RuleMatrix: Visualizing and understanding classifiers with rules. IEEE Trans. on Visualization and Computer Graphics 25, 1 (2019), 342–352.
  29. Mou Y., Xu K.: The media inequality: Comparing the initial human-human and human-AI social interactions. Computers in Human Behavior 72 (2017), 432–440.
  30. ProtoSteer: Steering deep sequence model with prototypes. IEEE Trans. on Visualization and Computer Graphics 26, 1 (2020), 238–248.
  31. Believe it or not: Designing a human-AI partnership for mixed-initiative fact-checking. In Proc. of the Annual ACM Symposium on User Interface Software and Technology (2018), ACM, pp. 189–199.
  32. A typology of guidance tasks in mixed-initiative visual analytics environments. Computer Graphics Forum 41, 3 (2022), 465–476.
  33. A methodology for task-driven guidance design. EuroVis Workshop on Visual Analytics (2023), 6 pages.
  34. Experimental evidence of effective human–AI collaboration in medical decision-making. Scientific Reports 12, 1 (2022), 14952.
  35. Towards effective human-AI decision-making: The role of human learning in appropriate reliance on AI advice. Int. Conf. on Information Systems (2023).
  36. Lotse: A practical framework for guidance in visual analytics. IEEE Trans. on Visualization and Computer Graphics (2022), 1124–1134.
  37. A survey of human-centered evaluations in human-centered machine learning. Computer Graphics Forum 40, 3 (2021), 543–568.
  38. Learning and teaching in co-adaptive guidance for mixed-initiative visual analytics. In EuroVis Workshop on Visual Analytics (Goslar, 2020), Turkay C., Vrotsou K., (Eds.), Eurographics Association, pp. 61–65.
  39. Co-adaptive visual data analysis and guidance processes. Computers & Graphics 100 (2021), 93–105.
  40. Cobots in knowledge work. Journal of Business Research 125 (2021), 135–142.
  41. Learning contextualized user preferences for co-adaptive guidance in mixed-initiative topic model refinement. Computer Graphics Forum 40, 3 (2021), 215–226.
  42. Knowledge generation model for visual analytics. IEEE Trans. on Visualization and Computer Graphics 20, 12 (2014), 1604–1613.
  43. Sundar S. S.: Rise of machine agency: A framework for studying the psychology of human-AI interaction (haii). Journal of Computer-Mediated Communication 25, 1 (2020), 74–88.
  44. From human-human collaboration to human-AI collaboration: Designing AI systems that can work together with people. In Extended Abstracts in the Proc. of the SIGCHI Conf. on Human Factors in Computing Systems (2020), ACM, pp. 1–6.
  45. A survey on ML4VIS: Applying machine learning advances to data visualization. IEEE Trans. on Visualization and Computer Graphics 28, 12 (2022), 5134–5153.
  46. Podium: Ranking data using mixed-initiative visual analytics. IEEE Trans. on Visualization and Computer Graphics 24, 1 (2018), 288–297.
  47. The What-If tool: Interactive probing of machine learning models. IEEE Trans. on Visualization and Computer Graphics (2019), 1–1.
  48. Xu W., Gao Z.: Applying human-centered AI in developing effective human-AI teaming: A perspective of human-AI joint cognitive systems. Interactions 31, 1 (2024).
  49. Re-examining whether, why, and how human-AI interaction is uniquely difficult to design. In Proc. of the SIGCHI Conf. on Human Factors in Computing Systems (2020), ACM, pp. 1–13.

Summary

  • The paper introduces a unified framework that deconstructs human-AI collaboration into agency, interaction, and adaptation dimensions.
  • It employs a systematic literature review and expert interviews to refine model dimensions and inform design space analysis.
  • Case studies demonstrate the framework's utility in comparative analysis, design reasoning, and gap identification for future research.

Deconstructing Human-AI Collaboration: Agency, Interaction, and Adaptation

Introduction

The paper "Deconstructing Human-AI Collaboration: Agency, Interaction, and Adaptation" (2404.12056) addresses the increasing complexity and ambiguity in the characterization of human-AI collaborative systems. As full automation by AI remains unattainable in most real-world scenarios, the focus has shifted to hybrid systems that leverage the complementary strengths of humans and AI agents. The authors identify a lack of systematic frameworks for comparing and classifying such systems, noting that existing models and terminology are fragmented and insufficiently expressive. To address this, the paper introduces a unified conceptual model structured around three high-level aspects: agency, interaction, and adaptation. This model is developed through a systematic literature review and iterative refinement via expert interviews, culminating in a design space that enables comprehensive analysis and comparison of human-AI collaborative systems.

Methodology

The conceptual model is constructed through a two-phase process:

  1. Systematic Literature Review: The authors survey recent literature using targeted keywords and manual curation from major conferences (CHI, IUI, EuroVis, VIS), assembling a corpus of works relevant to human-AI collaboration. From this corpus, they extract and aggregate system properties, collapsing redundancies and focusing on dimensions that describe system characteristics rather than user motivations.
  2. Interview Study: Semi-structured interviews with nine domain experts (PhDs and postdocs in visual analytics, explainable AI, NLP, RL, GNNs, ML) are conducted to refine the initial set of dimensions. The interview process is staged to elicit unbiased ideation, encourage thinking in terms of dimensions, and compare participant suggestions with the proposed model. Iterative feedback leads to convergence on a compact set of dimensions grouped into three categories.

The methodology emphasizes balancing expressiveness and generalizability, ensuring the model is sufficiently granular to distinguish between systems while remaining broadly applicable.

Conceptual Model: Design Space

The proposed design space is organized into three high-level categories, each decomposed into specific dimensions:

Agency

  • Agency Distribution: Specifies which agents (human, AI, mixed) possess decision-making control.
  • Agency Allocation: Describes whether agency is pre-determined (static assignment) or negotiated (dynamic, opportunistic assignment).

The model recognizes the trade-off between control and automation, highlighting the need for further research into unbiased negotiation mechanisms and temporal dynamics of agency.

Interaction

  • Interaction Intent: Captures agent motives (receiving guidance, requesting information, exploration, providing feedback).
  • Guidance Degree: Quantifies the extent of guidance (orienting, directing, prescribing).
  • Guidance Focus: Specifies the elements targeted by guidance (system, interface, interaction, data, task-specific).
  • Feedback Type: Differentiates between explicit, implicit, or both forms of feedback.

The interaction category is modeled symmetrically for human and AI agents, avoiding anthropocentric bias and enabling per-agent analysis. The dimensions are informed by prior work in visual analytics, communication models, and guidance frameworks.

Adaptation

  • Adapting Agents: Identifies which agents (human, AI, both) are capable of learning during collaboration.
  • Adaptation Method: Distinguishes between adaptation for task performance improvement and communication enhancement.
  • Information Learned: Details the types of information acquired (domain, data, task, agent goals, agent preferences).

The adaptation category acknowledges the challenge of modeling human learning and the intertwined nature of information acquisition in collaborative settings.

Case Studies

Three representative systems are analyzed using the proposed design space:

  1. Co-Adaptive Guidance for Topic Model Refinement: A mixed-initiative system where both human and multiple AI agents iteratively refine topic models. Agency is negotiated, interaction involves directing guidance and explicit/implicit feedback, and both agents adapt by learning preferences and improving communication.
  2. Podium: A visual analytics tool for ranking multivariate data. Agency is mixed but pre-determined, interaction includes directing/prescribing guidance and explicit feedback, and adaptation occurs for both agents, primarily through learning data properties and user preferences.
  3. ProtoSteer: An interface for steering deep sequence models via prototype modification. Agency is human-centered, interaction involves directing guidance and explicit feedback, and adaptation is present for both agents, with the AI learning domain/task information and the human improving communication.

These case studies demonstrate the model's capacity to encode diverse collaborative systems and highlight the multidimensional nature of real-world human-AI collaboration.

Implications and Future Directions

Practical Implications

The design space provides a systematic framework for:

  • Comparative Analysis: Enables structured comparison of collaborative systems across domains.
  • Design Reasoning: Assists practitioners in making informed design choices by clarifying trade-offs in agency, interaction, and adaptation.
  • Gap Identification: Reveals latent dimensions and underexplored areas, guiding future research and system development.

Theoretical Implications

  • Taxonomy of Agency: The model exposes the need for a more nuanced taxonomy of agency, including mechanisms for negotiation and dynamic allocation.
  • Temporal Dynamics: Current static assignment of properties limits analysis of systems with evolving collaboration modes; modeling temporal aspects is a critical research opportunity.
  • Expressivity vs. Generalizability: The categorical simplification of continuous dimensions facilitates systematic analysis but may obscure fine-grained distinctions; future work could explore more expressive frameworks with rigorous definitions.

Limitations

  • Human Learning Modeling: The symmetric treatment of human and AI learning is an acknowledged oversimplification; cognitive modeling of human adaptation remains an open challenge.
  • Granularity: The choice of categorical values for dimensions may limit expressivity; more granular options (e.g., human-in-the-loop, AI-in-the-loop) could enhance descriptive power.
  • Corpus Coverage: The current model is validated on three systems; comprehensive encoding of a larger corpus is necessary for broader generalization.

Speculation on Future Developments

  • Complex System Engineering: Platforms that abstract underlying complexity could facilitate experimentation with variations in design space dimensions.
  • Dynamic Collaboration Models: Integration of temporal and context-dependent properties will enable richer modeling of adaptive, evolving human-AI teams.
  • Interdisciplinary Research: Advances in cognitive science, HCI, and machine learning will inform more sophisticated models of human learning and agent interaction.

Conclusion

The paper presents a unified conceptual model for human-AI collaboration, structured around agency, interaction, and adaptation. Developed through literature synthesis and expert feedback, the design space enables systematic analysis and comparison of collaborative systems. Application to representative case studies demonstrates its utility and highlights areas for future research, including agency taxonomy, temporal modeling, and cognitive adaptation. The framework advances the capacity of the research community to reason about, design, and analyze human-AI collaborative systems in a rigorous and comprehensive manner.

File Document Download Save Streamline Icon: https://streamlinehq.com Markdown

Paper to Video (Beta)

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Ai Sparkles Streamline Icon: https://streamlinehq.com Sign Up to Generate

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Ai Sparkles Streamline Icon: https://streamlinehq.com Generate Now

Collections

Sign up for free to add this paper to one or more collections.

User Circle Single Streamline Icon: https://streamlinehq.com Sign Up

Tweets

Sign up for free to view the 2 tweets with 11 likes about this paper.

User Circle Single Streamline Icon: https://streamlinehq.com Sign Up for Free