AI Sequences Alter Phase Performance Without Enhancing Overall Creative Outcomes

In a randomized experiment evaluating generative artificial intelligence in art and design education, text-based and image-generative tools produced distinct advantages across different creative phases, but sequencing the tools did not translate into superior final creative products. The study examined how different configurations of generative AI influence both divergent exploration and convergent development among design students undertaking a structured artistic task.

Conducted by researchers Yiming Wei and Sijun Li at the Xi’an Academy of Fine Arts in China, the 2 × 2 between-subjects experiment evaluated 120 art and design students who were randomly allocated across four conditions of 30 participants each. Participants completed a standardized cultural-heritage chromatic reconstruction task using one of four assigned AI workflow sequences: Text-to-Text, Text-to-Image, Image-to-Text, or Image-to-Image, spanning an initial divergent exploration phase followed by a convergent development phase.

During the divergent exploration stage, students utilizing a dialogic text-based AI configuration achieved significantly higher divergent creative performance than those using an image-generative configuration, showing a mean difference of 0.823 that remained statistically significant after Holm adjustment (p < 0.001). Conversely, during the convergent development stage, image-generative AI yielded higher Aesthetic Quality at the unadjusted level with a mean difference of 0.493 (p = 0.026), although this difference did not survive Holm familywise-error correction (adjusted p = 0.052).

Testing the hypothesis that pairing text exploration with image convergence would maximize final results, the authors found that the Text-to-Image sequence did not significantly outperform the alternative sequencing configurations in Overall Creative Performance (contrast = 0.066, p = 0.671), and the factorial interaction between the two phases was nonsignificant (p = 0.541). The Text-to-Image sequence did produce higher baseline-adjusted immediate post-task knowledge scores (contrast = 1.478, 95% confidence interval [1.050, 1.906], p < 0.001), but this gain occurred without a statistically significant cross-phase interaction (p = 0.404).

The authors concluded that individual phase-specific advantages do not necessarily accumulate into superior completed creative products, challenging common assumptions about phase-matching workflows. Furthermore, because AI modalities were bundled with specific interaction and interface characteristics, and the immediate knowledge assessments demonstrated limited reliability, the findings do not establish isolated modality effects, durable learning gains, or synergistic sequencing, indicating that instructional AI orchestration remains a provisional framework requiring further empirical validation.

To evaluate how different AI configurations affected student work, the investigation implemented a comprehensive assessment protocol encompassing expert-rated creative performance, immediate task-relevant knowledge, and human–AI interaction processes. The researchers structured their analytical framework around four primary hypothesis tests, applying Holm familywise-error correction to safeguard against false-positive conclusions across multiple comparisons. This statistical rigor proved crucial when evaluating convergent development: although image-generative AI achieved a nominal advantage in Aesthetic Quality with an unadjusted difference of 0.493, the adjusted p-value of 0.052 placed the result outside conventional statistical significance, illustrating the necessity of adjusted thresholds when assessing nuanced creative metrics.

The pedagogical architecture of the trial centered on a standardized cultural-heritage chromatic reconstruction task designed to separate creative cognition into distinct phases. However, the lack of significant performance gains from this Text-to-Image sequence suggests that phase-specific strengths do not automatically integrate into an elevated holistic creative product.

Published in the Educational Psychology section of Frontiers in Psychology, the original research article by Yiming Wei and Sijun Li was received on 18 August 2026 and accepted on 20 October 2026. Reflecting on the outcomes, the authors emphasize that instructional AI orchestration should remain a provisional, context-sensitive perspective rather than an established curricular blueprint. Because the experimental configurations tied AI modality directly to specific interface mechanics, and because immediate knowledge assessments possessed limited reliability without demonstrating durable learning, the researchers caution that further empirical trials are necessary before formalizing multi-stage AI sequences in art and design education.

Subject of Research: Generative artificial intelligence configurations across creative phases in design education

Article Title: Generative AI configurations across creative phases: a 2 × 2 experiment on creative performance and immediate knowledge in art and design education

Article References: Wei, Y., & Li, S. (2026). Generative AI configurations across creative phases: a 2 × 2 experiment on creative performance and immediate knowledge in art and design education. Frontiers in Psychology, 17, Article 1971021. https://doi.org/10.3389/fpsyg.2026.1971021

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1971021

Keywords: generative artificial intelligence, art and design education, creative performance, divergent thinking, convergent thinking, human-AI interaction, AI orchestration, chromatic reconstruction, Generative, configurations, across, creative