The research team set out to answer a deceptively simple question: how do humans perceive and evaluate creativity when it is attributed to artificial intelligence? As generative AI systems produce paintings, poems, and ideas that are increasingly difficult to distinguish from human output, the psychological question of evaluation becomes urgent. If identical works are judged differently depending on whether people believe a human or an algorithm produced them, then much of what passes for creativity assessment may be contaminated by attitudes toward the creator rather than qualities of the creation.

The investigators addressed this question across three experiments, each probing a different corner of creative production. The first experiment focused on artistic creation, using paintings and poetry as stimuli. Participants evaluated creative outputs, and the researchers manipulated only the source label: some participants were told a work was created by a human, while others were told the identical work came from an AI system. The results were unambiguous. Outputs labeled as human-created were rated as more creative than the same outputs bearing an AI label. In the realm of art, at least, a persistent pro-human bias remains firmly intact.

This finding aligns with a long tradition in creativity research suggesting that artistic creativity is judged not only on formal properties such as novelty and usefulness but also on beliefs about intentionality, emotion, and lived experience. A painting is not merely an arrangement of pigments; it is presumed to encode a mind that chose, suffered, and expressed. When that presumption is removed by an AI attribution, evaluators appear to discount the work, even though the physical stimulus before their eyes has not changed by a single brushstroke or syllable.

The second experiment added a layer of nuance by examining co-created works, in which the attribution of responsibility between human and machine was manipulated. Here the picture became more textured. When poetry was labeled as involving human ideation with AI assistance, perceived originality was enhanced relative to other attributions. In other words, for poetry, a collaborative framing actually boosted judgments of originality, suggesting that participants saw the combination of human conceptual direction and machine execution as a virtue rather than a contaminant. Painting, however, told a different story: works attributed primarily to humans remained dominant in evaluators’ judgments. The medium itself, it seems, shapes how comfortable people are with machine involvement in the creative process.

Why would collaboration elevate poetry but not painting? The authors’ findings point toward a domain-specific pattern in attitudes toward AI creativity. Poetry, as a verbal art, may be perceived as closer to the kind of combinatorial, idea-driven processing at which large language models excel, so human-AI partnership reads as a natural extension of the writer’s toolkit. Painting, by contrast, may be more tightly bound to embodied skill, personal vision, and the aura of the individual artist, making any dilution of human authorship costly in evaluators’ eyes. The study’s summary of these results captures the tension precisely: viewing collaboration positively yet maintaining a persistent bias in valuing AI’s creative contributions.

The third experiment delivered the study’s most striking reversal. Instead of asking participants to judge finished artworks, the researchers used the Alternative Uses Test, a classic measure of divergent thinking in which respondents generate unusual uses for everyday objects. Divergent thinking is a cornerstone construct in creativity research, operationalizing the capacity to produce varied and original responses in open-ended tasks. When the same output was presented under different source labels, participants gave higher creativity ratings to responses bearing an AI label. The pro-human bias that dominated artistic judgment flipped entirely when the task was framed as idea generation rather than artistic creation.

This reversal is more than a curiosity; it carries real implications for how organizations, educators, and researchers measure creative potential. If evaluators unconsciously reward ideas attributed to AI in brainstorming and ideation contexts, then human contributors may be systematically undervalued in precisely the settings, such as innovation workshops, design sprints, and research labs, where divergent thinking is most prized. Conversely, in galleries, literary prizes, and art markets, AI-labeled work may face a penalty that no improvement in algorithmic technique can overcome. The bias is not a single attitude but a pair of domain-specific expectations about where minds, biological or artificial, belong.

Methodologically, the study relied on repeated-measures multivariate analysis of variance, known by the abbreviation RM MANOVA, a statistical framework well suited to comparing ratings across multiple dependent measures when the same evaluators encounter multiple conditions. By holding the creative products constant and varying only the attribution, the design isolates attribution itself as the causal factor, ruling out the alternative explanation that AI outputs simply differ in quality. The research was approved by the Ethics Committee of the Jing Hengyi School of Education at Hangzhou Normal University under approval number 2023019, and all procedures followed the ethical standards of the Declaration of Helsinki, with informed consent obtained electronically before data collection. The work was supported by the National Natural Science Foundation of China, the Natural Science Foundation of Zhejiang Province, and the Organized Research Program of Hangzhou Normal University.

There is also an irony worth savoring: the authors acknowledge that during the preparation of the work they used ChatGPT, DALL-E, and other AI tools to make experimental materials and correct the manuscript. A study revealing that humans discount AI-labeled art while favoring AI-labeled ideas was itself produced with substantial machine assistance, a fact that would likely have boosted its perceived originality had it been framed as a poetry project, and diminished it had it been framed as a painting. As generative models weave themselves ever deeper into creative work, this research suggests that the hardest problem may not be teaching machines to create, but teaching ourselves to judge what they create on its merits. Until then, the label on the canvas, or on the idea, will keep quietly rewriting the verdict before anyone has truly looked.

Subject of Research: Domain-specific human bias in evaluating creativity attributed to AI versus humans

Article Title: The domain-specific bias in creativity evaluation: human evaluators favor human-created art, but prefer AI outputs for divergent thinking

Article References: Hou, Y., Zhang, Y., Wang, B., Zhu, Y., & Wu, X. (2026). The domain-specific bias in creativity evaluation: human evaluators favor human-created art, but prefer AI outputs for divergent thinking. BMC Psychology. https://doi.org/10.1186/s40359-026-05704-x

Image Credits: AI Generated

DOI: 10.1186/s40359-026-05704-x

Keywords: artificial intelligence, creativity, creativity evaluation, divergent thinking, Alternative Uses Test, human-AI collaboration, artistic creation, poetry, painting, attribution bias, BMC Psychology, psychology

The research team, led by Shuiling Lu of Wuzhou University together with colleagues at UCSI University and Zhengzhou Normal University, followed PRISMA 2020 guidelines to sift the literature published between January 2015 and May 2025. Starting from 1,964 database records drawn from Web of Science, Scopus, Sage, Springer Link and JSTOR, and applying a search string that combined problem-based learning terms with creativity and higher-education keywords, the team ultimately retained fifteen peer-reviewed studies for systematic review. Only six of those studies, all quasi-experimental designs with pretest-posttest control groups, provided the complete statistical data needed to enter the meta-analysis.

The pooled result from those six studies was eye-catching. Compared with traditional instruction, problem-based learning was associated with a large improvement in creativity outcomes, yielding a pooled Hedges’ g of 1.64 with a 95 percent confidence interval running from 0.96 to 2.32. The overall effect was statistically significant, with a Z-statistic of 4.72 and a p-value below 0.001. In every one of the six studies, the problem-based learning group outperformed the control group on creativity measures, though the magnitude of the advantage varied considerably from one context to another.

That variation is where the caution begins. The heterogeneity statistic, I-squared, came in at 88.5 percent, meaning that the overwhelming majority of the variability between studies reflects real differences in contexts, interventions and measurements rather than random noise. The Q-test confirmed this, returning a value of 43.29 with five degrees of freedom and a p-value below 0.001. With only six studies available, the authors could not run formal subgroup analyses or meta-regression to statistically disentangle those sources of variance. They describe their pooled estimate as an upper-bound figure, potentially inflated by methodological weaknesses, and frame the primary contribution of their work not as a single number but as a demonstration that positive effects appeared consistently across diverse settings.

Methodological quality was a persistent concern. Using the ROBINS-I tool for non-randomized studies, two independent reviewers rated all fifteen included studies as carrying serious or critical risk of bias. The main culprits were inherent to educational research: the absence of randomization, which leaves the confounding domain at serious risk, and the impossibility of blinding students and instructors to the teaching method they were receiving. Missing outcome data and selective reporting, by contrast, were consistently rated low risk. A funnel plot showed slight asymmetry, hinting that small negative studies may be under-represented in the literature, though with only six points the plot has limited reliability.

The individual study effects tell a story of their own. Problem-based learning in engineering education produced the largest single effect, a Hedges’ g of 2.83, though that figure was converted from a reported Cohen’s d because raw standard deviations were unavailable, introducing uncertainty. Writing education yielded a g of 2.52, while PBL combined with digital mind-mapping in mathematics produced a moderate-to-large g of 1.40. Two engineering studies reported comparable effects of 0.99 and 1.41, and the smallest effect, 0.75, came from a dance-based intervention, suggesting that topic relevance, facilitator expertise and implementation fidelity all moderate outcomes. The authors stress that these subgroup comparisons are purely descriptive and hypothesis-generating, since no statistical comparisons between variants were possible.

One of the review’s most striking findings concerns how creativity itself is measured, and how differently that happens across disciplines. In arts and humanities contexts such as fashion design, collage making, dance and writing, assessors foregrounded aesthetic originality, narrative coherence, emotional connection and self-expression. In science and engineering settings, by contrast, creativity was judged through solution feasibility, technical viability, collaborative problem-solving and inventiveness in tackling authentic design problems. This pattern supports the domain-specificity view of creativity, treating it not as a single unified construct but as something that manifests differently depending on disciplinary expectations, a distinction educators must respect when designing both activities and assessment rubrics.

The measurement problem runs deeper still. The included studies operationalized creativity through markedly different instruments, from the Torrance Tests of Creative Thinking to self-report Likert scales and researcher-developed rubrics, each capturing distinct facets of fluency, flexibility, originality and usefulness. Two of the six meta-analyzed studies relied on self-report scales, which research suggests are vulnerable to social desirability bias and may correlate only moderately with demonstrated creative performance. Notably, the studies reporting the largest effects tended to use self-report measures or researcher-developed rubrics, raising the possibility that measurement characteristics inflated the observed magnitudes. None of the included studies employed widely validated, standardized creativity instruments with established psychometric properties, and all follow-ups were short-term, with none exceeding six months, leaving the durability of any creativity gains unknown.

Despite these caveats, the review identifies three mechanisms that appear to amplify problem-based learning’s effect on creativity. First, integrating the arts introduces non-verbal modes of expression, allowing students to externalize tacit knowledge through concept mapping, collage and dance that is difficult to articulate in words alone. Second, digital augmentation through platforms such as MATLAB, Genially and gamification tools creates interactive environments that scaffold creative risk-taking and sustain motivation, as seen in studies of matrix algebra and gamified mathematics training. Third, hybrid models combining problem-based learning with project-based learning generate synergistic gains across the so-called 4C competencies of creativity, collaboration, communication and critical thinking. The authors note these mechanisms are not mutually exclusive, though because most interventions bundled multiple components, the unique contribution of each cannot yet be isolated.

For educators and curriculum designers, the practical message is that problem-based learning shows genuine promise as a creativity-oriented strategy in higher education, particularly when enriched with digital tools or arts integration, but recommendations must remain tentative pending stronger validation. The authors call for future research built on randomized controlled designs, validated standardized creativity assessments and delayed post-tests to establish both causality and durability of effects. They also highlight gaps in the evidence base: disciplinary concentration in engineering and STEAM, sparse representation from the liberal arts and natural sciences, an exclusively undergraduate population, and incomplete reporting of key moderators such as intervention duration and facilitator expertise. Until such rigorous studies arrive, the dramatic effect size of 1.64 should be read not as a guarantee but as an encouraging, preliminary signal that when students are handed real problems to solve together, their creative capacities may genuinely flourish.

Subject of Research: The effect of problem-based learning on creativity outcomes in STEAM higher education

Article Title: Enhancing creativity in STEAM higher education through problem-based learning: a systematic review and meta-analysis

Article References: Lu, S., Phang, S. M., Judge, S. K., Liu, Y., & Fu, K. (2026). Enhancing creativity in STEAM higher education through problem-based learning: a systematic review and meta-analysis. Discover Education, 5(1), Article 1075. https://doi.org/10.1007/s44217-026-02125-1

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02125-1

Keywords: problem-based learning, creativity, STEAM education, higher education, meta-analysis, systematic review, pedagogy, divergent thinking, digital technology, arts integration, quasi-experimental design, educational assessment