The impact of human-AI co-creation on perceived creativity in generative AI art: a serial mediation mechanism of psychological ownership, perceived agency, and creative flow

The rapid integration of generative artificial intelligence into visual arts is fundamentally transforming how digital works are conceived, developed, and evaluated. As text-to-image models and multimodal generation systems evolve from passive tools into active participants, researchers and creators face a paradigm shift. Rather than focusing solely on whether machines can independently produce artistic outputs, behavioral scientists and human-computer interaction researchers are increasingly turning their attention to the collaborative space between human intent and machine generation.

A comprehensive study published in Frontiers in Psychology sheds new light on this dynamic. Led by researchers examining the psychological dimensions of digital art creation, the project investigates how different creation modes and workflow structures influence a user’s perception of creativity. By analyzing how human-AI interaction alters cognitive and emotional states during the artistic process, the findings offer critical insights for designers, educators, and technology developers navigating the AI-assisted creative landscape.

Background Context and Evolution of Artistic Tools

Historically, digital art tools served primarily supportive or administrative functions, allowing users to draw, edit, layout, or post-process media through direct manual manipulation. The advent of modern generative architectures—such as Stable Diffusion XL and analogous large-scale visual models—has entirely restructured this relationship. Users now participate in an iterative dialogue with algorithms, continuously shaping outcomes through prompt engineering, stylistic selection, multi-round modifications, and result curation.

This shift has elevated computational systems from simple instruments to collaborative partners. However, prior research into AI-generated content has largely concentrated on static aesthetic evaluations, source attribution labels, algorithmic bias, and public acceptance. Observers have frequently noted that labeling a work as AI-generated can trigger social-cognitive biases, causing audiences to undervalue its originality or emotional depth. Despite these contributions, a significant gap persisted in the literature: most studies positioned users merely as passive observers evaluating completed works, leaving the internal psychological processes of active human-AI co-creators largely unexplored.

Methodological Framework and Experimental Design

To address these theoretical and empirical gaps, the research team designed two rigorous, between-subject experiments involving a total of 600 valid participants. Conducted in early 2026, the studies utilized standardized digital art tasks centered on a unified theme: "The Symbiosis of Future Cities and Nature." This thematic constraint was implemented to minimize the confounding effects of topical variance on perceived creativity, aesthetic quality, and emotional resonance.

The backend generation processes across all AI-enabled conditions utilized the Stable Diffusion XL 1.0 framework, ensuring uniform technical parameters such as a 1,024-by-1,024 resolution, 30 sampling steps, and a constant Classifier-Free Guidance scale.

In Study 1 (involving 300 participants, average age 24.39 years), the researchers manipulated the creation mode across three distinct conditions:

• Human-only creation, where participants independently conceived and finalized artwork proposals without algorithmic generation.
• AI-only generation, where participants viewed and selected from system-provided candidate images without editing prompts or regenerating outputs.
• Human-AI co-creation, where participants actively inputted prompts, selected styles, executed multi-round modifications, and curated the final visual output.

Study 2 (involving 300 participants, average age 24.28 years) focused specifically on varying degrees of AI involvement across three prespecified workflows:

• Low AI involvement, characterized by weak algorithmic support and high user control.
• Moderate AI involvement, featuring balanced multi-round interaction where users engaged in continuous dialogue with the AI while retaining ultimate decision-making authority over prompts, styles, and curation.
• High AI involvement, where the system dominated the generative process and participants were restricted to selecting or confirming the final output.

Following the completion of the creative tasks, participants completed validated 7-point Likert scales measuring perceived creativity, psychological ownership, perceived agency, and creative flow. Additionally, the system logged objective behavioral metrics—such as creation duration, prompt revisions, and edit counts—while independent domain experts evaluated the final artworks using the Consensual Assessment Technique (CAT).

Key Findings and Psychological Mechanisms

The empirical results revealed pronounced differences across creation modes and workflow designs, substantiating the authors’ theoretical model regarding subjective creativity evaluations.

In Study 1, human-AI co-creation produced significantly higher perceived-creativity ratings (M = 4.645) compared to AI-only generation (M = 3.315). Furthermore, the data supported a serial mediation pathway. When psychological ownership, perceived agency, and creative flow were incorporated into the model, the direct effect of co-creation became non-significant, while the total indirect effect was robust (Effect = 1.1077, 95% CI [0.8226, 1.4392]). Specifically, deeper collaboration fostered a heightened sense of psychological ownership over the artwork; this ownership positively predicted perceived agency (the user’s belief in their ability to influence the outcome); perceived agency subsequently facilitated creative flow (focused immersion and process enjoyment); and creative flow culminated in elevated evaluations of the work’s overall creativity.

Study 2 provided crucial nuance regarding the optimal distribution of labor between humans and machines. The moderate-involvement workflow yielded significantly higher perceived-creativity ratings (M = 5.155) than either the low-involvement workflow (M = 3.985) or the high-involvement workflow (M = 3.647). Similar non-monotonic patterns emerged across psychological ownership, perceived agency, creative flow, aesthetic evaluation, and emotional resonance.

The data suggest that when AI involvement is too low, users lack generative expansion and dynamic feedback. Conversely, when AI involvement is excessively high—restricting prompt modifications and local edits—users experience a loss of personal contribution and control, mirroring the lower evaluations seen in purely automated generation. Balanced workflows that combine generative assistance with consequential user choice maximize subjective creative engagement.

Implications for Industry, Design, and Education

The implications of these findings extend across multiple sectors, offering actionable guidance for software developers, creative professionals, and educators.

For interface designers and platform developers, the research underscores the necessity of building collaborative systems that preserve user control. Features such as prompt history tracking, side-by-side variant comparisons, local editing tools, and reversible selection mechanisms allow users to visibly trace their influence on the final product. By intentionally designing workflows that prevent the system from completely overtaking decision-making, developers can foster stronger psychological ownership and agency among users.

In the realm of design education, the study advocates for a re-evaluation of how generative tools are integrated into curricula. Rather than discouraging AI use or rewarding passive, one-click output generation, educators can structure assignments around documented co-creative processes. Students can be encouraged to submit prompt histories, articulate iterative decision-making rationales, and critically reflect on how algorithmic suggestions interacted with their personal artistic intent.

Broader Context and Methodological Limitations

While the findings offer robust empirical support within the tested parameters, the researchers noted several limitations that invite further investigation. The participant pool primarily consisted of young adults and students with moderate familiarity with digital tools, meaning the behavioral patterns observed may not entirely mirror the practices of seasoned professional artists. Additionally, because psychological variables were measured concurrently after task completion, the serial mediation pathways should be interpreted as theoretically consistent associations rather than definitive temporal-causal chains.

Future research will need to examine whether these psychological patterns hold across diverse professional domains—such as professional graphic design, commercial copywriting, and musical composition—and explore how advanced AI literacy modulates user perception over extended periods of sustained collaboration.

Conclusion

The research demonstrates that the perceived creativity of generative AI art is not merely a byproduct of algorithmic sophistication or source labeling. Instead, it is deeply contingent upon how users experience contribution, control, and cognitive immersion during the act of creation. As generative technologies continue to mature, frameworks that thoughtfully balance computational capability with meaningful human agency will remain essential for fostering genuine innovation and creative fulfillment.