PaperlayerView exact revision on arXiv

Source-linked AI summary

Human-AI Collaboration: From Paradoxes to Patterns

Michael Weiss

arXiv:2609.36481v1cs.AIcs.HCcs.MA

TL;DR

Human-AI collaboration patterns are examined through the dimensions of autonomy and initiative. The paper maps tensions as paradoxes and derives pattern descriptions from those mappings, illustrating the approach with scenario examples.

  • Problem

    The paper addresses how to document human-AI collaboration patterns by identifying underlying problems and solutions and examining tensions within those problems.

    Source: §1 · p. 1

  • Method

    It uses a paradox perspective and a circular template to surface tensions, map opposing forces and their outcomes, and derive pattern descriptions.

    Source: §1 · p. 1, §2.1 · p. 2

  • Results

    The paper documents human-AI interaction patterns and illustrates paradox mappings for Instruction, Delegation, and Co-creation scenarios.

    Source: §4.2 · pp. 5–7, §6 · p. 13

  • Takeaways & Limitations

    The process aims to give paradoxes a more dynamic view and derive pattern descriptions from a model of the paradox.

    Source: §6 · p. 13

    The patterns assume users are domain experts who can assess AI outputs, so they do not directly apply to novices or learners.

    Source: §5 · p. 8

Abstract

from arXiv · show

Evidence shows that humans and AI systems perform better together, by collaborating, than alone. This paper examines two key design dimensions of human-AI collaboration (autonomy and initiative) and explores the collaboration patterns that they generate. Documenting these patterns starts with identifying the underlying problems and solutions, followed by examining the internal tensions within the problems. The paper uses a paradox perspective to analyze those tensions. It describes a process for surfacing the tensions and mapping the underlying paradoxes. It also illustrates how the pattern descriptions can be derived from mapping these paradoxes. Finally, the paper documents four human-AI collaboration patterns: Instruction, Delegation, Assistance, and Co-creation.

1 Introduction

The paper examines autonomy and initiative as dimensions of human-AI collaboration, then uses a paradox perspective to map tensions and derive collaboration-pattern descriptions.

1 Introduction

  • The paper examines autonomy and initiative as design dimensions that generate human-AI collaboration patterns, focusing on the human-controlled centaur model.It distinguishes this focus from the reverse centaur model.Source: §1 · p. 1
  • The paper analyzes tensions within problems through a paradox perspective, defining paradox as contradictory yet interdependent forces whose resolution recreates the conditions that caused it.The definition is attributed to Smith & Lewis (2011).Source: §1 · p. 1
  • The paper describes a process for surfacing tensions and mapping paradoxes, then shows how pattern descriptions can be derived from those mappings.Source: §1 · p. 1

2 Background and related work

The background frames persistent tensions as paradoxes to be navigated and situates human-AI collaboration among related work on AI agency, collaboration patterns, and creative tensions.

2.1 Reframing tensions as paradoxes

  • A both/and framing treats the forces in a tension as complementary, while the paper's circular template maps opposing forces, their outcomes, and strategies to alternate or integrate them.Source: §2.1 · p. 2
  • Paradoxical tensions involve contradictory yet interdependent forces that persist through attempted resolution; software design illustrates a recurring trade-off between integration and modularity.Source: §2.1 · p. 2

2.2 Human-AI collaboration

  • AI is described as shifting from a deterministic tool to an active collaborator, matching complex tasks' dynamic, nonlinear, and ambiguous character and redistributing agency and responsibility.Source: §2.2 · p. 2
  • Prior work documents ten human-AI collaboration patterns, while autonomy and initiative are identified as the two dimensions central to this paper's design space.Source: §2.2 · p. 3
  • The automation-augmentation paradox links automation of well-understood tasks to close human-AI collaboration for refining models and responding when conditions change.Source: §2.2 · p. 3
  • Creative collaboration research identifies five persistent tensions, including ambiguity versus precision, and connects them to design goals for supporting exploration while representing intent.Source: §2.2 · p. 3

3 Design dimensions for human-AI collaboration

The section frames human-AI collaboration through autonomy and initiative, which together define four scenarios whose appropriate allocation depends on task context.

3 Design dimensions for human-AI collaboration

  • Autonomy captures how independently AI operates and the scope of its assigned task, while initiative captures whether AI waits for human input or acts on its own.Source: §3 · p. 3
  • The two dimensions yield four scenarios, each requiring a different combination of autonomy and initiative.Figure 1 presents the design space based on these dimensions.Source: §3 · p. 3, Figure 1 · p. 4
  • How autonomy and initiative should be allocated depends on task context, including tacit knowledge, well-understood self-contained tasks, mixed cognitive and routine work, and tasks needing diverse perspectives.Table 1 cross-references collaboration scenarios with contexts.Source: §3 · p. 4, Table 1 · p. 4

4 From tensions to paradoxes and patterns

The section extends a static paradox template into a process that traces needs, responses, consequences, and underlying forces. It applies this mapping to four collaboration scenarios and outlines how to derive pattern descriptions from the resulting paradoxes.

4.1 Process of mapping a paradox

  • A paradox map identifies a scenario-specific need, action and reaction, their consequences, the circular response, and the tension between underlying forces.The listed elements organize the mapping beyond the forces and pull shown in the earlier template.Source: §4.1 · pp. 4–5
  • Mapping is iterative; in the instruction scenario, prompting and observing AI responses lets the user refine prompts while balancing sensitivity to behavior against prompt structure.The probe–reflect cycle captures how each action’s consequences shape the next response.Source: §4.1 · p. 5

4.2 Paradoxes in human-AI collaboration scenarios

  • In instruction, the need to codify tacit knowledge leads to probing and reflection, whose consequences form a cycle around the tacit–explicit tension.Prompting brings high sensitivity to AI behavior but loose prompt structure; reflection improves the prompt while risking sensitivity to what its new structure did not anticipate.Source: §4.2 · p. 5, Figure 2 · p. 6
  • Delegation offloads well-understood, clearly scoped tasks to free human capacity for higher-priority work, but creates a centralized–decentralized control tension and risks lost oversight.Reviewing at least a sample of AI outputs is presented as a way to reclaim control.Source: §4.2 · p. 6, Figure 3 · p. 6
  • Assistance is organized around a whole–part tension, with the part denoting what the AI assists with.Source: §4.2 · p. 7, Figure 4 · p. 7
  • Co-creation centers on a self–other tension, with the AI playing the role of the other.Source: §4.2 · p. 7, Figure 5 · p. 7

4.3 From paradoxes to patterns

  • The first four pattern-mapping steps identify the scenario and context, name the interdependent forces, and frame the problem as a both/and tension that cannot be resolved.Source: §4.3 · pp. 7–8
  • The remaining steps explain the forces, describe the action/reaction solution, trace the resulting context, and identify nested paradoxes; Table 2 applies the steps to instruction.Source: §4.3 · p. 8, Table 2 · p. 8

5 Patterns

The patterns organize human-AI collaboration around recurring problems, solutions, and consequences, assuming users have already chosen AI and can assess its output. Four patterns address different task contexts, from tacit knowledge to work requiring multiple perspectives.

5 Patterns

  • Each pattern is a recurring solution to a common problem in context; the patterns assume AI use is already chosen and the user is a domain expert able to assess its output.They do not directly apply to novices or learners.Source: §5 · p. 8

5.1 Instruction

  • Instruction addresses tacit knowledge that is difficult to externalize: probing and reflecting improve alignment, but structuring prompts sharpens intent while making misalignments and surprises easier to miss.Judgment remains difficult to codify, and instructions cannot fully capture tacit knowledge.Source: §5.1 · pp. 9–10
  • Use structure sooner for repetitive, well-defined tasks, but preserve openness for exploratory work; once a task is codified, it can transition to Delegation, Assistance, or Co-Creation.Co-Creation leaves room for iterative exploration while assigning the interaction structure to a team of agents.Source: §5.1 · pp. 10–11

5.2 Delegation

  • Delegation offloads self-contained, codified tasks to agents, freeing attention for higher-level work but risking disengagement; verify outputs and reclaim authority when they miss expectations.In the example workflow, the human specifies the process and checks results if needed, while agents execute it.Source: §5.2 · p. 11

5.3 Assistance

  • Assistance offloads a routine part of mixed-level work, improving that part’s efficiency but risking poor fit with the whole; review and integrate its output in context.The example embeds agents in an editor to give feedback on claims in the user’s latest changes.Source: §5.3 · pp. 11–12

5.4 Co-Creation

  • Co-Creation brings AI perspectives into work requiring diverse viewpoints, broadening inputs while risking diluted ownership; reassert human judgment, voice, and intent as the output evolves.In the thematic-analysis workflow, the user can revise agent-identified themes until satisfied.Source: §5.4 · p. 12

6 Conclusion

The paper contributes human-AI collaboration patterns and a process for mapping tensions to paradoxes and deriving pattern descriptions from those mappings.

6 Conclusion

  • The paper documents human-AI collaboration patterns as one of its two contributions.Source: §6 · p. 13
  • Its process maps tensions to paradoxes and derives pattern descriptions from those models, aiming to give paradoxes a more dynamic view.Source: §6 · p. 13
  • Testing the process on a larger collection of patterns remains future work.Source: §6 · p. 13
Loading 2609.36481v1…