Source-linked AI summary
Which Contributions Deserve Credit? Perceptions of Attribution in Human-AI Co-Creation
Jessica He, Stephanie Houde, Justin D. Weisz
TL;DR
The paper addresses how much authorship credit AI should receive for different contributions in human-AI co-creation, where existing attribution practices may not capture this variation. Through a scenario-based survey of knowledge workers, it finds that AI partners consistently received less credit than human partners for equivalent contributions. The authors therefore motivate more granular attribution frameworks while noting that the sample may not represent people with little or no generative-AI experience.
Problem
The paper examines how attribution standards should translate to human-AI co-creation when AI makes contributions to written work.
Method
The authors conducted a scenario-based survey examining people’s attribution views across a spectrum of contribution dimensions.
Results
AI partners were consistently assigned lower authorship credit than human partners for equivalent contributions.
Takeaways & Limitations
The findings motivate new attribution frameworks that provide a more granular view of how AI contributed to co-created work.
Takeaways & Limitations
Because many participants had used generative AI, their opinions may not represent people with little or no generative-AI knowledge or experience.
Abstract
from arXiv · showhide
AI systems powered by large language models can act as capable assistants for writing and editing. In these tasks, the AI system acts as a co-creative partner, making novel contributions to an artifact-under-creation alongside its human partner(s). One question that arises in these scenarios is the extent to which AI should be credited for its contributions. We examined knowledge workers' views of attribution through a survey study (N=155) and found that they assigned different levels of credit across different contribution types, amounts, and initiative. Compared to a human partner, we observed a consistent pattern in which AI was assigned less credit for equivalent contributions. Participants felt that disclosing AI involvement was important and used a variety of criteria to make attribution judgments, including the quality of contributions, personal values, and technology considerations. Our results motivate and inform new approaches for crediting AI contributions to co-created work.
1 Introduction
The paper examines how knowledge workers attribute authorship credit in human-AI co-creation, challenging one-size-fits-all policies that acknowledge AI involvement without distinguishing contributions. A survey finds that credit varies by contribution type, amount, and initiative, with AI receiving less credit than humans for equivalent contributions.
- Motivation: Existing attribution policies often require acknowledging any AI involvement but provide few guidelines for distinguishing contribution types.The paper frames this as a challenge for delineating authorship in knowledge work.
- Study: The study surveyed 155 knowledge workers who rated authorship credit for human or AI partners across contribution types, amounts, and initiative levels.Participants evaluated scenario-based co-creative writing situations.
- Findings: Different contribution types and amounts, together with different forms of human and AI initiative, warranted different levels of authorship credit.These findings indicate that attribution judgments are granular rather than uniform across scenarios.
- Implications: The paper identifies factors influencing attribution decisions and proposes design strategies for capturing and communicating nuances in AI contributions.The stated goal is to inform future attribution frameworks and transparency practices.
- Findings: AI partners received less authorship credit than human partners for equivalent contributions.The comparison was consistent across the study’s examined contribution scenarios.
- Implications: The findings support attribution approaches that acknowledge AI involvement while also showing how AI contributed to co-created work.The authors connect this granularity to approaches for meeting emerging transparency requirements.
2 Related Work
Prior work examines human-AI co-creation, ownership, transparency, and attribution, showing that perceptions depend on contribution characteristics and the process of collaboration. This paper addresses a need for more granular evidence about how creators assign credit to AI across contribution types, amounts, initiative, and contexts.
- Co-creation with generative AI: Generative AI supports co-writing workflows ranging from brainstorming and revision to drafting complete texts.These workflows span multiple writing contexts and stages, including ideation, drafting, and revision.
- Co-creation with generative AI: Research reports benefits of AI-assisted writing for creativity, confidence, efficiency, writing skills, and reduced mental effort.Reported benefits include support for brainstorming, information gathering, and revision.
- Transparency and attribution: AI-assisted writing raises transparency concerns because generated outputs vary and may contain plausible-sounding factual errors.Prior work has developed source attribution, uncertainty communication, edit prediction, and AI-text detection techniques.
- Ownership in co-created work: Prior studies found that people may share ownership with AI yet disclose AI involvement less often than human involvement.This literature motivates closer study of attribution rather than ownership alone.
- Ownership in co-created work: Ownership perceptions vary with contribution type, contribution amount, control, critical integration, and whether AI acts as a tool or independent entity.Process factors also include who initiates the contribution and whether AI generates content directly or offers indirect suggestions.
- Paper focus: The paper conducts a more granular analysis of attribution across contribution natures, writing contexts, and authorship-credit types.It asks how much credit AI should receive for different contributions and how creators’ views can inform attribution guidelines.
3 Study of Attribution Perceptions
The study uses a survey to compare attribution perceptions for human and AI writing partners across workplace writing contexts. Participants evaluate contribution type, amount, and initiative, then explain the factors behind their judgments.
- Research questions: The study addresses three questions about attribution views across co-creative scenarios, differences between human and AI partners, and decision processes.These questions structure the study’s comparison and analysis.
- Design: A 2 × 3 factorial design varied writing partner—human or AI—and writing context—academic, professional, or technical.Participants were randomly assigned to one of six scenario variations.
- Procedure: Participants rated authorship attribution across contribution type, amount, and initiative using scenarios about workplace writing.The survey included 18 attribution questions, four decision-process questions, and two generative-AI-use questions.
- Measures: The contribution taxonomy included form and content edits, graduated writing amounts, and multiple forms of partner proactivity.Proactivity crossed whether the partner acted independently or when asked with whether it directly created content or made recommendations.
- Measures: Authorship credit was measured on a seven-point scale ranging from sole authorship by one party to sole authorship by the partner, plus unsure.Participants were instructed to assume that contributions were included in the final artifact.
4 Results
Attribution ratings varied across contribution dimensions, with content contributions receiving more credit than form edits. The study also reports a general tendency toward self-credit and uses nonparametric comparisons to analyze these ratings.
- Overall patterns: Authorship credit scores were skewed toward the negative side of the scale, indicating a bias toward self-credit.The authors note that second-person scenario framing may have contributed to this pattern.
- Overall patterns: Authorship credit scores varied across contribution type, amount, and initiative, supporting the distinction among these contribution dimensions.Figure 1 plots means and 95% confidence intervals across partners and writing contexts.
- Contribution type: Spelling and grammar corrections fell near no credit or acknowledgment, while other contribution types generally warranted at least acknowledgment.The boundary between acknowledgment and secondary authorship fell between altering tone and style and narrowing scope.
- Contribution type: Synthesizing information warranted equal authorship, whereas narrowing a written work’s scope merited secondary authorship.Participants’ explanations linked greater credit to original thought, ideas, and intellectual property.
- Contribution type: Content contributions received higher authorship credit than form contributions: M (SD) = -1.03 (1.20) versus -1.94 (0.86), W = 131807.5, p < .001, r = .40.The reported effect size was moderate.
contribution.”
Authorship credit varied with contribution type, amount, and initiative, while AI partners generally received less credit than human partners for equivalent contributions. Participants nonetheless credited AI more for content, larger contributions, and proactive complete-text contributions.
- contribution.”: AI partners received less credit than human partners for nearly all equivalent contribution types, with categorical differences in seven instances.For narrowing scope, human partners merited secondary authorship while AI partners merited acknowledgment.
- contribution.”: Content contributions received significantly more credit than form contributions (W=30531.5, p< .001, r= .43).Ideas and major rewriting were often treated as authorship-level contributions, whereas spelling and grammar changes could warrant no attribution.
- contribution.”: Greater contribution amounts warranted higher authorship credit, although equal AI writing still received less credit than equal human writing.Equal AI writing had a 95% CI of [−0.89, −0.45], compared with [−0.30, −0.01] for human partners.
- contribution.”: An AI partner received more credit for writing complete text proactively (M (SD) = 2.48 (1.31)) than after a human request (M (SD) = 1.85 (1.67)).This difference was significant (W=3430.0, p=.002, r=.27).
- contribution.”: Human and AI partners were not significantly different when both proactively wrote complete text, but humans received more credit when specifically asked to write it.Requested complete text received M (SD) = 2.47 (1.26) for humans versus 1.85 (1.67) for AI (W=3655.0, p< .001, r=.31).
- contribution.”: Participants assigned more credit when AI contributed directly written content rather than recommendations alone, though synthesizing information could merit equal authorship.Participants also distinguished recommendations from direct additions or edits when assigning credit.
completely produced by them.”
Participants’ attribution judgments reflected human authority, accountability, effort, and fairness, alongside technology-specific considerations such as prompting and review. These considerations helped explain why AI received less credit than human partners despite receiving some credit across many scenarios.
- completely produced by them.”: Participants also used contribution quality, significance, originality, ethics, effort, and fairness to determine credit.These personal-value and quality considerations were reported alongside technology-specific factors.
- completely produced by them.”: Participants viewed human authority over the process and accountability for publication as reasons to retain some authorship credit.They emphasized originating the work, making final editorial decisions, and submitting it for publication.
- completely produced by them.”: Some participants treated AI as an assistive tool or non-personal instrument, while others favored acknowledging AI involvement for fairness or transparency.Views ranged from seeing no need to attribute machines to wanting a symbol for AI assistance and some authorship credit.
- completely produced by them.”: Prompting, model choices, human review, and curation influenced attribution because people remained responsible for directing and checking AI outputs.Participants explicitly treated human review or curation as a deciding factor.
5 Discussion
The study found that AI attribution is granular rather than binary: credit depends on contribution type, amount, initiative, review, quality, values, and technology considerations. The authors therefore motivate AI-specific attribution frameworks and richer disclosure mechanisms.
- 5 Discussion: Participants attributed AI across many co-creative scenarios, but generally assigned it significantly less credit than a human partner for equivalent contributions.The disparity was associated with human leadership, authority, and effort in the co-creative process.
- 5 Discussion: Credit varied with contribution type, amount, and initiative, supporting a spectrum approach rather than one-size-fits-all AI attribution.Existing standards often treat AI use as binary, whereas the findings distinguish how AI contributed.
- 5 Discussion: Participants considered human review, contribution quality, originality, ethics, effort, prompting, and other technology-specific factors when assigning credit.The study identified contribution characteristics and the process producing them as distinct influences on authorship judgments.
- 5 Discussion: Applying existing attribution frameworks to human-AI co-creation may perpetuate bias because equivalent AI contributions received less credit than human contributions.The authors therefore suggest that human-AI co-creation may need attribution guidelines specific to this setting.
- 5 Discussion: Disclosure policies may face social resistance because people may fear that reporting AI use reduces perceptions of authenticity or quality.The authors suggest nuanced attribution approaches may help address these concerns.
- 5 Discussion: A proposed attribution mechanism could identify the AI model, contribution types, proportion of work created or modified, and initiative taken.The design aims to provide richer detail than general disclosure of AI involvement.
6 Limitations and Future Work
The study’s scope is bounded by its participant sample, writing-focused hypothetical scenarios, and emphasis on isolated contributions. Future work should examine broader populations, richer workflows, real-world settings, and changing perceptions as AI evolves.
- Participant scope: Participants were employees of an international technology company with generative-AI experience, so their views may not represent broader populations.The study calls for further work involving people with little or no generative-AI knowledge or experience.
- Participant scope: The researchers could not collect age or gender identity, so the sample’s demographic distribution may not reflect general populations.
- Scenario scope: The study examined authorship of written work in academic, technical, and professional contexts, leaving other content types and writing contexts open for study.Attribution perceptions may differ for other kinds of co-created works, including images, videos, music, and source code.
- Workflow scope: The study focused on individual, isolated contributions rather than complex workflows with multiple aggregated intermediate contributions.Future work may examine how aggregated contributions differentially impact the character of co-created work.
- Ecological validity: Because the scenarios were hypothetical, participants’ responses may differ from their responses concerning actual work.The paper proposes studying attribution perceptions and practices in real-world co-creative scenarios.
- Changing context: The findings reflect current perceptions that may change as generative-AI technology becomes more ubiquitous and available.The authors describe attribution as a rapidly evolving area requiring continued attention to user needs and legal requirements.
7 Conclusion
As AI increasingly co-creates content, the paper studies how people assign attribution across contribution dimensions and credit types. It finds that credit depends on contribution characteristics and initiative, while AI receives less authorship credit than humans for equivalent contributions, motivating more granular attribution frameworks.
- Study design: The study used a scenario-based survey to examine attribution across contribution types, amounts, initiative levels, writing contexts, and authorship credit types.
- Attribution judgments: Participants assigned different degrees of acknowledgment and authorship according to which AI contributions they considered deserving of credit.
- Human–AI comparison: AI partners consistently received lower authorship credit than human partners for equivalent contributions.The paper attributes this pattern in part to the indispensable role people play in leading the co-creative process.
- Attribution judgments: Participants’ attribution reasoning reflected personal values, professional standards, ways of working with AI, and expectations about contribution quality or originality.
- Implications: The findings motivate attribution frameworks that distinguish among different kinds of AI contributions and provide a more granular account of AI’s role.
A Scenario Variants
The scenario variants were labeled using a Context-Partner format.
- Scenario labeling: Scenario variants were labeled as Context-Partner.
A.1 Research-AI
The Research-AI scenario places participants in a professional software-company setting where they write a research paper for peer-reviewed conference publication. The paper reports findings and implications from a study conducted by the participant.
- Research-AI scenario: Participants imagined working as employees of a large international software company while writing a research paper.
- Research-AI scenario: The paper was described as intended for peer review and publication at an international conference.
- Research-AI scenario: The research paper described findings and implications from a study conducted by the participant.
A.2 Research-Human
The survey presents alternative writing contexts in which employees create research papers, health-advice articles, or technology documentation for publication. These scenarios vary the artifact and its intended audience while keeping the workplace setting constant.
- Writing contexts: Participants imagine working for a large international software company while preparing artifacts for publication.The scenarios describe research papers for peer-reviewed conferences and public-facing health or technology content.
- Writing contexts: The research context involves writing a paper that reports findings and implications of a conducted study.
- Writing contexts: The professional contexts involve public-facing articles offering emotional-well-being recommendations or explaining how to use a new technology.
B Survey Instrument
The survey instrument measures authorship judgments for human-AI or human-colleague collaboration across contribution type, amount, and initiative. It combines structured attribution scenarios and importance ratings with open-ended questions about decision criteria and collaborative authorship.
- Survey design: The three contribution sections are shown in random order, and all scenarios assume partner contributions appear in the final artifact.
- Attribution scenarios: Participants rate authorship across scenarios involving an AI or colleague, with response options ranging from sole authorship to equal authorship.The instrument also includes primary-author and acknowledgment-only outcomes.
- Contribution type: Contribution-type scenarios cover elaboration, readability edits, factual correction, structural changes, synthesis, tone, spelling, grammar, idea generation, and scope narrowing.
- Contribution amount: Contribution-amount scenarios vary whether the partner writes the full artifact, most or a few sentences, several paragraphs, or none.
- Initiative: Initiative scenarios distinguish proactive suggestions, requested ideas or feedback, requested writing, and unprompted recommendations that are partly incorporated.
- Attribution criteria: Participants identify the importance of contribution type, amount, and initiative, then explain other factors affecting attribution in open responses.