Does moderate AI involvement feel more creative?
Across two Chinese online experiments with 600 mostly student participants, a workflow preserving meaningful human choices produced higher self-rated creativity than low- or high-involvement alternatives. The tasks were short, the workflows bundled several features and the result is not evidence that the finished art was objectively better.
The short version
- What happened
- The paper reports two between-subject experiments with 300 valid participants each, recruited in China through Credamo; most participants were students and had low-to-moderate art and AI experience.
- Who it affects
- Creative workers may benefit from tools that preserve revision and final-selection authority instead of automating every available step.
- What remains unproven
- The mostly student Chinese sample had limited art and AI experience and cannot represent professional creative practice.
Keep in mind
The mostly student Chinese sample had limited art and AI experience and cannot represent professional creative practice.
The tasks lasted roughly 10 to 12 minutes and used one digital-art theme, so sustained collaboration and other media remain untested.
Study 1 did not hold the form of the final artefact and required effort constant across conditions.
Study 2 bundled several interaction permissions, preventing attribution to a single degree or feature of AI involvement.
The central outcome was participants' own perceived creativity, not validated long-term skill or independently judged artistic value.
Source analysis · 1 linked source. Primary analysed source date: 9 October 2026.
Read the full report · 7 min The evidence, methods, implications and limitations
The direct answer: keeping meaningful choices mattered in this task
In these short online art experiments, participants rated their work as more creative when they shared control with the image generator rather than surrendering the process to it. The clearest result came from the second experiment: the workflow labelled moderate AI involvement produced a mean perceived-creativity score of 5.155 on the study's seven-point scale, compared with 3.985 for the low-involvement workflow and 3.647 for the high-involvement workflow. The condition effect was large within this experiment, F(2,297)=57.204, with partial eta-squared of 0.278.
The result is about a specific interaction design, not a universal optimum. Moderate involvement preserved more opportunities to shape prompts, regenerate, edit and choose. High involvement removed more of those choices. Because the conditions changed several permissions at once, the experiment cannot say whether prompting, editing, final selection or another element produced the difference. It also measured how creative participants felt, not whether professional critics, audiences or markets judged the work to be better.[1]
Two experiments separated creation mode from workflow design
Study 1 compared three creation modes: human-only work, AI-only generation and human-AI co-creation. Study 2 compared low-, moderate- and high-involvement AI workflows. Each experiment retained 300 valid participants, giving 100 people per condition. The average participant was about 24 years old. In Study 1, 57.7% were undergraduates and another 19% were junior-college students; Study 2 was similar, with 60.3% undergraduates. Participants reported relatively limited art or design experience and low-to-moderate generative-AI experience.
Recruitment used China's Credamo platform, and the tasks and questionnaires were delivered in Simplified Chinese. Participants spent about 10 to 12 minutes on the creation task and survey and received RMB 10 after passing attention checks. The researchers powered each experiment for a medium-sized one-factor, three-group difference and increased the target above the minimum of roughly 252 to support indirect-effect and adjusted analyses. This is adequate for the experimental contrasts, but the sample is not representative of professional artists or the wider Chinese public.[1]
Ownership, agency and flow moved with the workflow
The researchers measured psychological ownership—the sense that the work was 'mine'—alongside perceived agency, creative flow and perceived creativity. In Study 2, the moderate workflow also produced the highest scores for ownership, agency and flow. The high-involvement condition produced the lowest mean psychological ownership at 3.875, agency at 3.425 and flow at 3.692. Aesthetic evaluation and emotional resonance also differed across conditions, again favouring the moderate workflow.
Study 1 fitted an indirect-effects model in which co-creation was associated with creativity judgements through ownership, agency and flow. The full serial pathway had an estimated effect of 0.0418 with a bootstrapped 95% confidence interval from 0.0122 to 0.0834. The direct effect was not statistically reliable after the proposed mediators were included. That pattern is consistent with the authors' theory, but mediation fitted to post-task measures does not prove a psychological sequence unfolded exactly in that order.[1]
Why the study does not rank AI art quality
Perceived creativity was the participant's judgement of their own work. The paper distinguishes that from condition-blind external ratings, but the headline findings concern self-perception. A tool can make someone feel more capable or invested without producing a more original, technically accomplished or culturally valuable artefact. Conversely, a demanding workflow might yield a strong outcome while feeling less fluent. Creative organisations need both process and output measures rather than treating satisfaction as quality.
Study 1 also did not hold the form of production fully constant. The human-only condition submitted an artwork concept, while AI-enabled conditions involved viewing or generating images. Differences in concreteness, effort and task form could therefore influence ratings. The experiment used one art theme and one short encounter with a system. It does not establish effects for writing, music, film, long-term collaboration, commissioned work or the pressures of professional deadlines.[1]
What designers, teachers and creative workers can use
The practical lesson is to make automation adjustable and preserve consequential human decisions. A creative tool can propose alternatives while leaving concept, style, revision and final selection with the person. Product teams should evaluate whether users understand which choices remain theirs, whether they can reverse an automated step and whether the interface rewards reflection rather than one-click output. Completion time and output count alone would miss the ownership and agency effects examined here.
For teachers, prompt histories and revision rationales can make the human contribution visible. Students can explain why they rejected options, what they changed and how their own assessment compares with independent feedback. For creative workers, the same evidence supports resisting workflows that remove judgement while leaving the worker accountable for the result. It does not justify forcing AI into a role, grading students on AI use or defining creativity through a single self-report scale.[1]
Funding, disclosure and what would change the assessment
The Shaanxi Provincial Art Science Planning Project funded the work under grant SYH2025006. The authors declared no commercial or financial conflict. They disclosed using Stable Diffusion XL 1.0 to generate experimental stimuli under the reported parameters and generative tools for language editing and formatting. Electronic consent was used for the anonymous minimal-risk online experiment under the authors' institutional process.
Confidence would increase if researchers reproduced the result with equivalent final artefacts, professional creators, longer projects and several models or interfaces. Factorial designs should vary prompt authority, regeneration, editing and selection separately. Independent judges and behavioural outcomes should sit beside self-ratings, and follow-up should test whether agency, skill and creative confidence persist after repeated use. Until then, the defensible conclusion is narrow: this tested middle-control workflow felt more creative than the tested extremes for a young, mostly student sample.[1]