Who Owns AI-Generated Art? How Prompt Revisions Shape Ownership
Paper: Who owns AI-generated artwork? Revisiting the work of generative AI based on human-AI co-creation
Authors: Yohan Hwang, Dongkwang Shin, and Jang Ho Lee (2025)
Journal: Telematics and Informatics, 98, 102266
Central Argument
The more actively people participate in generating and refining AI-created images, the more likely they are to perceive those images as their own work.
Rather than determining who legally owns AI-generated artwork, this study investigates how people develop a sense of authorship and ownership through their interactions with generative AI. Its central focus is the relationship between iterative prompt refinement and perceived creative ownership.
1. The Research Problem: Can Humans Claim Authorship of Images Generated by AI?
Traditionally, the relationship between artistic creation and its human creator has been relatively straightforward. Generative AI complicates this relationship by separating creative intention from image production.
Humans develop ideas, formulate prompts, evaluate results, and request modifications. However, the images themselves are generated by AI systems.
This raises several questions. Can someone who has not physically drawn an image be considered its creator? Does repeatedly refining prompts constitute creative authorship? What degree of ownership can users claim over AI-generated images?
These questions are particularly relevant in education, where students increasingly use generative AI to produce assignments and creative projects.
The study therefore begins with a fundamental question: How should we understand human authorship when the creative process is distributed between humans and AI?
2. The Research Gap: Beyond Legal Ownership
Previous research on generative AI and copyright has largely focused on two issues.
The first concerns the use of copyrighted materials in training AI systems. Researchers have examined whether such materials can be used without permission and how the rights of original creators should be protected.
The second concerns the legal ownership of AI-generated outputs. These debates ask whether copyright should belong to AI developers, users who provide prompts, or creators whose works contributed to the training datasets.
While these discussions establish the legal and ethical context, the authors identify a need for empirical research on how actual users perceive ownership after participating in AI-assisted creative processes.
In particular, relatively little empirical attention has been given to the relationship between repeated prompt refinement and perceived ownership.
The study shifts the focus from determining who should legally own AI-generated artwork to examining how users come to regard such artwork as their own.
3. The Core Idea: Creative Involvement Shapes Ownership Perception
The authors understand AI-assisted image generation as an iterative process of human-AI interaction.
Users begin with an intended image, translate their ideas into prompts, evaluate the generated results, and modify their instructions when those results differ from their intentions.
Through this process, users actively select, interpret, and direct the development of the images.
The authors propose that greater participation in this iterative process is associated with a stronger sense of creative agency and ownership.
This suggests that perceived ownership is not necessarily a binary distinction between human-created and AI-created works. Instead, it may vary according to the degree of human involvement.
Importantly, stronger feelings of ownership do not automatically establish legal copyright.
4. How the Argument Develops
The study develops its argument through six stages.
Stage 1: Questioning conventional authorship. Generative AI separates the technical production of an image from the human decisions that shape its creation.
Stage 2: Shifting from outputs to processes. Rather than examining the final image alone, the authors investigate the human activities involved in producing it.
Stage 3: Introducing prompt refinement as a measurable variable. The number of prompt inputs and revisions serves as an indicator of user engagement.
Stage 4: Comparing ownership perceptions. Participants are divided into groups according to prompt frequency, and their ownership and copyright attitudes are statistically compared.
Stage 5: Interpreting the findings through co-creation. The authors draw on theories of creative participation and reader response to explain the user's active role in interpreting and directing AI outputs.
Stage 6: Extending the implications to education and copyright. The findings lead to recommendations for AI literacy, ethical awareness, and educational approaches emphasizing collaborative creative decision-making.
5. Research Methods
The study involved 30 first-year university students majoring in English language and literature in South Korea. Participants were aged 20–22 and included 26 Korean students and four international students from Vietnam. None had prior experience using generative AI tools.
The participants completed an image-generation project in a course titled Digital Literacy and Humanities.
They selected English words and created images representing their historical origins and changing meanings. After preparing initial visual interpretations, students used ChatGPT to develop prompts and Bing Image Creator to generate images.
They revised their prompts until the generated images more closely reflected their intended concepts.
One participant, for example, attempted to represent the meaning of "persona" using masks surrounding a human silhouette. When the AI interpreted the word "mask" as a medical mask, the student refined the prompt to specify the intended symbolic imagery and composition.
After the project, the researchers collected questionnaires and reflection papers.
The survey included eight seven-point Likert-scale questions addressing ownership and copyright perceptions.
Students were divided into three groups according to prompt frequency: fewer than 10 prompts (12 participants), 10–49 prompts (10 participants), and 50 or more prompts (eight participants).
The researchers used descriptive statistics, the Kruskal-Wallis test, and post-hoc comparisons to examine differences between the groups.
They also analyzed students' written reflections using the VADER and SentiArt sentiment-analysis models, together with qualitative content analysis.
6. Key Findings
Finding 1: Students with more prompt interactions reported stronger feelings of ownership.
Participants who used fewer than 10 prompts scored an average of 5.08 out of 7 on the statement that an AI-generated image was their own artwork.
The average rose to 6.50 among those using 10–49 prompts and remained 6.50 among those using 50 or more prompts.
The differences across the groups were statistically significant (χ² = 12.013, p = 0.002).
However, there was no significant difference between the two higher-frequency groups.
The findings therefore support an association between greater prompt engagement and stronger ownership perceptions, rather than demonstrating that ownership feelings increase indefinitely with every additional revision.
Finding 2: Prompt frequency was also associated with perceptions of copyright attribution.
Agreement with the statement that copyright should belong to the person entering the prompt differed significantly across the three groups (χ² = 13.122, p = 0.001).
The mean scores were 5.17 for the lowest-frequency group, 6.60 for the middle group, and 6.75 for the highest-frequency group.
Students who engaged in more extensive prompting were therefore more likely to support attributing copyright to prompt creators.
However, the study measured participants' opinions about copyright, not the legal validity of those claims.
Finding 3: Participants recognized multiple potential owners.
Across the entire sample, agreement that copyright should belong to the prompt creator averaged 6.07 out of 7.
At the same time, agreement that copyright should belong to the AI developer averaged 6.03.
This reveals an important tension. Students could acknowledge the contribution of human users while also recognizing the role of AI developers.
Reflection papers revealed different positions, including individual ownership, developer ownership, shared ownership, and the rejection of ownership claims.
Thus, participants did not converge on a single model of copyright attribution.
Finding 4: Creative satisfaction coexisted with ethical concerns.
Sentiment analysis revealed generally favorable emotional responses to AI-generated images, particularly feelings of happiness and surprise.
Students were pleased when repeated refinements brought the generated images closer to their intended results.
However, some also expressed concerns about dependence on AI, biases in training data, the possible displacement of human artists, and copyright infringement.
These findings suggest that positive experiences of human-AI collaboration can coexist with concerns about its ethical and social consequences.
7. The Most Distinctive Claim: From AI as Producer to AI as Creative Partner
One of the study's most distinctive arguments concerns how the identity of the creator may change in AI-assisted art.
The authors refer to Andy Warhol's artistic practice, in which existing images and photographs were selected, transformed, and incorporated into new artworks.
They suggest that AI-assisted creation may involve a comparable reconsideration of creative contribution, particularly when human users actively direct and transform the generated outputs.
The paper concludes by suggesting a possible shift in how people describe AI-generated artwork:
"AI drew this one for me."
becomes
"I drew this with AI."
This change in language reflects a possible shift in the recognition of human creative agency.
The authors argue that creative education should therefore extend beyond teaching students how to operate generative AI tools. It should also develop their ability to evaluate outputs, refine ideas, make creative decisions, and address ethical questions.
Nevertheless, the recognition of creative participation must be distinguished from the legal recognition of copyright.
8. New Questions Raised by the Study
Questions explicitly proposed by the authors
Would professional artists perceive ownership in the same way?
The participants were students without professional artistic training. Further research is needed to determine whether artists with established creative practices would respond similarly.
How do ownership perceptions change over time?
Longitudinal research could explore whether initial enthusiasm and feelings of ownership persist as users gain experience with AI tools.
How should the rights of artists whose works appear in training datasets be addressed?
The authors suggest examining copyright frameworks that consider both prompt creators and the original sources used to train AI systems.
Additional questions emerging from the findings
Does the number of prompt revisions accurately measure creative contribution?
Fifty repetitive prompts may involve less creative judgment than five carefully considered revisions. Future studies could distinguish the quantity of interaction from the quality of creative decisions.
Why does a stronger sense of ownership not necessarily correspond to a stronger perception of personal creativity?
The study found significant group differences in ownership perception, but not in the belief that the generated image represented the participant's creative abilities (p = 0.400).
This suggests that feeling ownership over an artwork and recognizing one's own creativity within it may be distinct psychological experiences.
What happens when AI agents perform most of the iterative work autonomously?
If a human establishes the artistic direction while an AI agent independently generates and refines hundreds of variations, how would that affect the human's perception of authorship?
This question becomes increasingly relevant as generative AI systems move toward autonomous creative workflows.
Critical Assessment
The study offers empirical evidence that participation in AI-assisted image creation is associated with how users perceive authorship and ownership.
However, several limitations qualify its conclusions.
The sample consisted of only 30 university students in a specific educational setting. The findings therefore cannot be generalized directly to professional artists or broader populations.
Furthermore, participants were not randomly assigned to different levels of prompting. The observed differences establish an association, but do not prove that additional prompting causes stronger ownership perceptions.
The study also investigates subjective beliefs about copyright rather than determining the legal ownership of AI-generated works.
Its most defensible conclusion is that students who engaged more extensively in AI image generation tended to identify more strongly with the resulting artworks and were more likely to recognize the prompt creator's claim to copyright.
The broader implication is that human-AI co-creation challenges traditional assumptions about creative agency, making the process of creation increasingly important to how authorship is understood.
