The Collaboration Gap in Human-AI Work
Introduces a framework for understanding the fragility of human-AI collaboration, analyzing grounding conditions and repair burden.
Key Findings
Methodology
The study employs a constructivist grounded theory analysis based on 16 interviews with designers, developers, and applied AI practitioners. It introduces a framework to analyze grounding conditions and repair burden in human-AI collaboration by distinguishing three interaction structures: one-shot assistance, weak collaboration, and grounded collaboration.
Key Results
- The study finds that collaboration often breaks down when the appearance of partnership outpaces the grounding capacity of the interaction. Specifically, data from 16 interviews show that participants generally perceive current LLM systems as significantly lacking in supporting deep collaboration.
- In one-shot assistance, the user provides a prompt, the system produces an output, but shared understanding remains low, suitable for low-risk tasks.
- In grounded collaboration, the system helps surface assumptions and track context, making collaboration more stable.
Significance
This research offers a new perspective on understanding the fragility of human-AI collaboration, emphasizing the importance of grounding conditions and repair burden. The framework can assist academia in better designing and evaluating human-AI collaboration systems and provide guidance for industry to improve collaboration experiences in practical applications.
Technical Contribution
The technical contributions include proposing a new framework to analyze grounding and repair issues in human-AI collaboration, differing from existing methods by emphasizing the grounding capacity of interactions and the distribution of repair burden rather than focusing solely on model performance.
Novelty
This study is the first to systematically analyze grounding conditions and repair burden in human-AI collaboration, providing a new framework to understand and improve the fragility of such collaborations.
Limitations
- The study is limited by a small sample size of only 16 interviews, which may not fully represent all LLM application scenarios.
- The applicability of the framework needs to be validated in more practical applications.
Future Work
Future research could expand the sample size, validate the framework's applicability in different application scenarios, and explore how design improvements can enhance the grounding capacity of collaboration.
AI Executive Summary
In today's technological landscape, large language models (LLMs) are widely used in fields such as programming, design, writing, and analysis. Despite being considered potential collaborators, the actual collaboration experience often falls short of expectations. Users frequently need to diagnose misunderstandings, reconstruct missing assumptions, and repeatedly repair misaligned responses. This phenomenon reflects a broader issue of fragility in collaborative AI.
This paper introduces a new conceptual framework to understand this fragility, based on an analysis of 16 interviews with designers, developers, and applied AI practitioners. The study shows that stable collaboration depends not only on model capability but also on the interaction's grounding conditions. By distinguishing three structures of human-AI collaboration: one-shot assistance, weak collaboration, and grounded collaboration, the research reveals the reasons for collaboration failures.
In one-shot assistance, the interaction structure remains close to a request-and-response pattern, where the user provides a prompt, and the system produces an output, but shared understanding remains low. This structure works for low-risk tasks such as summarization or boilerplate generation but does not support deeper collaboration.
Weak collaboration emerges when interaction becomes iterative. Users refine prompts, correct outputs, add context, or ask for revisions. These interactions may appear collaborative, but the burden of repair remains largely human. The user must infer what has gone wrong, reconstruct missing assumptions, and guide the system back toward the task.
In grounded collaboration, the interaction begins to support explicit clarification, signaling, and mutual repair. The system helps surface assumptions, track context, and make misalignment more visible. Final authority may still remain with the human, but the interaction itself becomes more balanced because repair no longer depends entirely on human improvisation.
This research provides a new perspective on the design of human-AI collaboration, emphasizing the importance of grounding conditions and repair burden. By reframing the collaboration gap as a grounding and repair problem, this work offers a conceptual lens for rethinking how human-AI collaboration is designed.
Deep Analysis
Background
With the rapid development of artificial intelligence technology, large language models (LLMs) have been widely applied in professional workflows such as programming, design, writing, and analysis. These models are not only seen as tools but also as potential collaborators. However, the actual collaboration experience often falls short of expectations, with users frequently needing to diagnose misunderstandings, reconstruct missing assumptions, and repeatedly repair misaligned responses. This phenomenon reflects a broader issue of fragility in collaborative AI. Studies have shown that while LLMs perform well in isolated environments, their performance may degrade when required to collaborate, a phenomenon known as the 'collaboration gap.'
Core Problem
The core problem lies in the fragility of human-AI collaboration. Although LLMs perform well in many tasks, in collaborative environments, users often need to diagnose misunderstandings, reconstruct missing assumptions, and repeatedly repair misaligned responses. This collaboration failure is not merely about whether outputs are correct but about participants' inability to reliably establish shared assumptions, interpret task state, or efficiently repair misunderstandings.
Innovation
The core innovation of this paper is the introduction of a new framework to analyze grounding conditions and repair burden in human-AI collaboration. By distinguishing three interaction structures: one-shot assistance, weak collaboration, and grounded collaboration, the study reveals the reasons for collaboration failures. This framework emphasizes the grounding capacity of interactions and the distribution of repair burden, rather than focusing solely on model performance.
Methodology
- �� Employs constructivist grounded theory analysis based on 16 interviews with designers, developers, and applied AI practitioners.
- �� Iterative coding of practitioner accounts to identify common structures and issues in collaboration.
- �� Distinguishes three interaction structures: one-shot assistance, weak collaboration, and grounded collaboration.
- �� Analyzes the grounding capacity and repair burden distribution of each structure.
Experiments
The experimental design is based on 16 semi-structured interviews with designers, developers, and applied AI practitioners. The interviews cover the use of LLMs in workflows involving drafting, ideation, coding, evaluation, and decision support. Participants described when collaboration with the model felt productive, when it became fragile, and how they responded when outputs diverged from task requirements or expectations.
Results
The study finds that collaboration often breaks down when the appearance of partnership outpaces the grounding capacity of the interaction. Specifically, data from 16 interviews show that participants generally perceive current LLM systems as significantly lacking in supporting deep collaboration. In one-shot assistance, the user provides a prompt, the system produces an output, but shared understanding remains low, suitable for low-risk tasks. In grounded collaboration, the system helps surface assumptions and track context, making collaboration more stable.
Applications
The framework can be used to improve the collaboration experience of LLMs in fields such as programming, design, writing, and analysis. By enhancing the grounding capacity of interactions and the distribution of repair burden, systems can better support deep collaboration, reducing the user's burden in diagnosing misunderstandings and repairing misaligned responses.
Limitations & Outlook
The study is limited by a small sample size of only 16 interviews, which may not fully represent all LLM application scenarios. The applicability of the framework needs to be validated in more practical applications. Additionally, the research primarily focuses on the perspectives of designers and developers, potentially overlooking the needs and challenges of other user groups.
Plain Language Accessible to non-experts
Imagine you're in a kitchen cooking a meal. You have an assistant, a large language model, who can help you chop vegetables, season dishes, and even suggest new recipes. However, sometimes the assistant misunderstands your instructions, like using salt instead of sugar or baking when you wanted to boil. This is similar to the problems in human-AI collaboration: while the assistant is smart, it doesn't always understand your intentions.
To make the collaboration smoother, you need to constantly communicate your ideas to the assistant, check its work, and correct it when it makes mistakes. This is what the paper refers to as the 'repair burden.' If you always have to spend a lot of time correcting the assistant's errors, the collaboration becomes tiring.
To improve this situation, the study suggests some methods, like having the assistant repeat your instructions or ask questions when it's unsure. It's like in the kitchen, where you have the assistant confirm each step to ensure it understands correctly.
Through these methods, human-AI collaboration can become more seamless, just like in the kitchen, where you and the assistant can work together to create delicious dishes.
ELI14 Explained like you're 14
Hey there! Imagine you're playing a super cool online multiplayer game. You have an AI teammate who can help you fight monsters, level up, and collect gear. Sounds awesome, right? But sometimes, this AI teammate messes up, like defending when you need it to attack or running left when you want it to go right.
This is like the problems in human-AI collaboration: the AI is smart, but sometimes it doesn't get what you're thinking. To make the game go smoothly, you need to keep telling the AI your plans, check its actions, and correct it when it makes mistakes.
Researchers found that if the AI can better understand your instructions or ask you when it's unsure, the collaboration would be smoother. It's like in the game, where you have the AI confirm each step to make sure it understands correctly.
With these methods, the gaming experience can be more fun, just like you and your AI teammate can work together to defeat all the enemies!
Glossary
Large Language Model (LLM)
A type of AI model trained on vast amounts of text data, capable of generating and understanding natural language.
In this paper, LLMs are used as collaborators in fields like programming and design.
Human-AI Collaboration
The process where humans and AI systems work together to complete tasks, emphasizing interaction and cooperation.
The paper explores the fragility and reasons for failure in human-AI collaboration.
Common Ground
Refers to shared beliefs, assumptions, and goals among participants, essential for effective communication and collaboration.
The study emphasizes the importance of common ground in stable collaboration.
Repair Burden
The responsibility for identifying and correcting misunderstandings in collaboration.
The paper analyzes the distribution of repair burden in different collaboration structures.
One-shot Assistance
A simple request-and-response pattern where the user provides a prompt, and the system produces an output.
Suitable for low-risk tasks but does not support deep collaboration.
Weak Collaboration
An interaction that appears collaborative, where users need to constantly adjust prompts and correct outputs.
The study points out that the repair burden in weak collaboration is primarily human.
Grounded Collaboration
An interaction that supports explicit clarification and mutual repair, with the system helping to surface assumptions and track context.
Grounded collaboration makes human-AI collaboration more stable.
Constructivist Grounded Theory
A method of generating theory through data analysis, emphasizing the abstraction of concepts from practice.
The paper uses this method to analyze interview data.
Signaling
The process of enhancing shared understanding through explicit feedback and confirmation.
The study suggests using signaling to improve the grounding capacity of collaboration.
Design Mechanisms
Specific methods and strategies designed to enhance system functionality and user experience.
The paper proposes three design mechanisms to improve human-AI collaboration.
Open Questions Unanswered questions from this research
- 1 Current LLM systems are significantly lacking in supporting deep collaboration, particularly in establishing and maintaining common ground. Research needs to explore how to enhance the system's grounding capacity to reduce the user's burden in repairing misunderstandings.
- 2 Although the paper proposes a new framework to analyze grounding and repair issues in human-AI collaboration, its applicability needs to be validated in more practical applications. Future research should expand the sample size and test the framework's effectiveness in different application scenarios.
- 3 The study primarily focuses on the perspectives of designers and developers, potentially overlooking the needs and challenges of other user groups. Future work should include more diverse user groups to fully understand the challenges of human-AI collaboration.
- 4 The current research mainly relies on interview data, lacking quantitative experimental support. Future studies could validate the framework's hypotheses through experiments and quantify the effects of different collaboration structures.
- 5 The distribution of repair burden may vary across different application scenarios, and research needs to further explore how to optimize the distribution of repair burden in different scenarios.
Applications
Immediate Applications
Programming Assistant
By enhancing the grounding capacity of LLMs, develop smarter programming assistants to help developers debug and optimize code more effectively.
Design Tools
Integrate LLMs into design tools to provide smarter design suggestions and automation features, enhancing designers' productivity.
Writing Assistant
Develop smarter writing assistants to help users quickly generate high-quality text content and provide real-time feedback and correction suggestions.
Long-term Vision
Intelligent Collaboration Platform
Create an intelligent collaboration platform integrating multiple AI technologies to support cross-domain team collaboration and innovation.
Personalized Learning System
Develop a personalized learning system based on LLMs, providing customized learning content and feedback according to users' learning styles and needs.
Abstract
LLMs are increasingly presented as collaborators in programming, design, writing, and analysis. Yet the practical experience of working with them often falls short of this promise. In many settings, users must diagnose misunderstandings, reconstruct missing assumptions, and repeatedly repair misaligned responses. This poster introduces a conceptual framework for understanding why such collaboration remains fragile. Drawing on a constructivist grounded theory analysis of 16 interviews with designers, developers, and applied AI practitioners working on LLM-enabled systems, and informed by literature on human-AI collaboration, we argue that stable collaboration depends not only on model capability but on the interaction's grounding conditions. We distinguish three recurrent structures of human-AI work: one-shot assistance, weak collaboration with asymmetric repair, and grounded collaboration. We propose that collaboration breaks down when the appearance of partnership outpaces the grounding capacity of the interaction and contribute a framework for discussing grounding, repair, and interaction structure in LLM-enabled work.
References (19)
The Construction of Shared Knowledge in Collaborative Problem Solving
J. Roschelle, Stephanie D. Teasley
Interpreting Interpretability: Understanding Data Scientists' Use of Interpretability Tools for Machine Learning
Harmanpreet Kaur, H. Nori, Samuel Jenkins et al.
Constructing Grounded Theory
Kathy Charmaz
"To LLM, or Not to LLM?": How Designers and Developers Navigate LLMs as Tools or Teammates
Varad V. Vishwarupe, Ivan Flechais, Nigel Shadbolt et al.
Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance
Gagan Bansal, Tongshuang Sherry Wu, Joyce Zhou et al.
Human-Centered Artificial Intelligence: Reliable, Safe & Trustworthy
B. Shneiderman
On Clark and Schaefer’s Contribution Model and its applicability to Human-Computer Collaboration
D. Traum
Coordination of knowledge in communication: effects of speakers' assumptions about what others know.
Susan R. Fussell, R. Krauss
The use of visual information in shared visual spaces: informing the development of virtual co-presence
R. Kraut, Darren Gergle, Susan R. Fussell
Using Visual Information for Grounding and Awareness in Collaborative Tasks
Darren Gergle, R. Kraut, Susan R. Fussell
Bringing Transparency Design into Practice
Malin Eiband, H. Schneider, Mark Bilandzic et al.
Guidelines for Human-AI Interaction
Saleema Amershi, Daniel S. Weld, Mihaela Vorvoreanu et al.
Grounding Gaps in Language Model Generations
Omar Shaikh, Kristina Gligori'c, Ashna Khetan et al.
Questioning the AI: Informing Design Practices for Explainable AI User Experiences
Q. Liao, D. Gruen, Sarah Miller
Grounding in communication
H. Clark, S. Brennan
A "speech acts" approach to grounding in conversation
D. Traum, James F. Allen
Referring as a Collaborative Process
Philip R. Cohen, J. Morgan, M. Pollack
On Using Language
C. K. Grant
The Collaboration Gap
Tim R. Davidson, Adam Fourney, Saleema Amershi et al.