Which Contributions Deserve Credit? Perceptions of Attribution in Human-AI Co-Creation

Research Background and Issues

Identified Problems or Challenges

• With the widespread application of AI systems utilizing Large Language Models (LLMs) in creative workflows, determining attribution in human-AI collaborative creation has become a critical issue.
• Current attribution policies primarily adopt a binary approach (either no AI involvement or unified attribution), lacking nuanced treatment of the types and levels of AI contributions.
• Studies show that people's attribution judgments in collaborative creation with AI heavily depend on the type, quantity, and agency of contributions, yet these details remain underexplored.

Importance

• Reasonably defining attribution in human-AI collaborative creation is crucial, not only for recognizing creators but also for addressing significant ethical, legal, and technical transparency issues.
• The inadequacies in current publishing and legal frameworks may lead to unfair attribution judgments, hindering transparent applications of AI technologies.

Research Motivation and Related Work

• Building on related studies (e.g., how attribution and ownership are assigned), the authors aim to delve deeper, particularly focusing on finer-grained creative contexts.
• This research centers on knowledge workers' perceptions of AI contribution attribution, aiming to explore fair attribution distribution across diverse creative scenarios and design future AI attribution frameworks.

Solution

Methods and Solutions

• The study proposes a scenario-based survey (N=155) to systematically evaluate perceptions of attribution for humans and AI across different creative contexts, enabling a more granular analysis of attribution patterns.
• Quantitative and qualitative analyses were conducted across three dimensions (type of contribution, quantity of contribution, agency), examining human perceptions of the level of attribution AI contributions should receive.
• A flexible AI contribution attribution framework is proposed, not only clarifying AI involvement but also making AI contribution details transparent through attribution statements.

Innovations

• Compared to traditional "binary attribution" (whether AI is involved or not), the authors propose a more complex "spectrum-based" attribution framework based on contribution type and quality.
• The study systematically explores users' psychological mechanisms regarding attribution for the first time and provides improvement suggestions by integrating contribution and process factors.

Implementation Steps and Key Techniques

1. Survey Design:
   • Investigate attribution perceptions across six scenario combinations (AI or human collaborators, academic, technical, and professional writing contexts).
   • Quantify participants' attribution preferences across three dimensions (type of contribution, quantity of contribution, agency).
2. Data Collection:
   • Includes author-level attribution assessments in the questionnaire (7-point Likert scale) and open-ended questions.
3. Data Analysis:
   • Quantitative analysis of attribution scores using non-parametric tests (e.g., Wilcoxon rank-sum test).
   • Reflexive Thematic Analysis to extract additional factors influencing attribution and human attribution decision-making processes.
4. Innovative Design:
   • Explore user-input-based attribution statement templates that dynamically display AI's depth, type, and agency of involvement (inspired by the Creative Commons licensing framework).

Research Findings

Specific Findings

1. Key Discoveries:
   • Different types of contributions significantly influence the level of attribution AI receives: content-related contributions are more likely to receive higher attribution than formal contributions (e.g., spelling corrections).
   • When AI and humans contribute equally, AI is often assigned lower attribution weight (i.e., lower recognition for AI contributions).
2. Key Dimensions Influencing Attribution:
   • Type of Contribution: Content contributions (e.g., new ideas) are more important than formal contributions (e.g., grammar corrections).
   • Quantity of Contribution: Greater contributions lead to higher attribution.
   • Agency: In certain cases, AI-generated text initiated autonomously is more likely to receive attribution than text generated upon request.
3. Human Trade-offs in Attribution Decisions:
   • Quality and originality, technical transparency, established ethics, and existing rules significantly influence attribution decisions.
   • Humans often prioritize control over AI work and responsibility allocation.

Comparison with Existing Solutions and Advantages

• Compared to simple disclosure statements (e.g., "This work used AI assistance"), this study provides a more granular attribution model, clarifying AI's specific roles and contributions in the creative process.
• The survey reveals users' differing perceptions of AI as a "tool" versus a "collaborator," offering scientific evidence for future policy and attribution framework design.

Experimental or Evaluation Results

• Statistical analysis indicates that humans generally tend to assign higher attribution to "human collaborators," even when AI and human contributions are equal, revealing a bias against AI recognition.
• The proposed attribution design tools (e.g., AI contribution statement generator) show potential for integration into future policy tools to meet public demands for AI transparency.

Limitations and Future Directions

• Limitations:
  • Participants were primarily from technology companies, which may limit representativeness.
  • Coverage of specific writing formats and domains is relatively narrow, with multimodal content such as visual arts and music yet to be explored.
• Future Directions:
  • Expand to a globally diverse population, encompassing various cultural and professional backgrounds.
  • Investigate "dynamic attribution" decision mechanisms in actual AI usage scenarios.
  • Further evaluate user acceptance, usability, and operability of the fine-grained attribution design framework.

This study deeply uncovers the complexity of attribution distribution in human-AI collaborative creation and proposes practical solutions that can provide valuable references for academia, industry, and policy-making.