Ask a language model to brainstorm marketing slogans and you will receive a competent list in seconds. Ask it again and you will receive a similar list. Ask a hundred people to do the same with AI assistance and you will find that their collective output is measurably more creative than what they would have produced alone — but also measurably less diverse. This paradox — better individual performance, worse collective variety — sits at the center of a growing body of evidence about what generative AI actually does to human creativity.
The Meta-Analytic Picture
Holzner, Maier, and Feuerriegel (2025), in a systematic literature review and meta-analysis spanning 28 studies and 8,214 participants, provide the most comprehensive quantitative assessment to date. Their findings organize into three clear results.
The comparison between AI alone and humans alone shows no significant difference in creative performance (Hedges' g = −0.05). This null result challenges both camps in the debate: AI is neither the creative powerhouse its enthusiasts claim nor the sterile pattern-matcher its critics describe. On standard creativity assessments, AI and humans perform equivalently.
When humans collaborate with GenAI, they significantly outperform humans working without assistance (g = 0.27). The effect is moderate but consistent across diverse tasks including creative writing, business ideation, and divergent thinking exercises. AI acts as an augmentative tool — a collaborator that lifts individual creative performance above what either party achieves alone.
The third finding is where the trouble lies. GenAI collaboration produces a large negative effect on the diversity of ideas (g = −0.86). When many people use the same AI system, their outputs converge. The ideas are individually better but collectively more similar. This homogenization effect is substantial — nearly four times larger than the positive effect on individual creativity — and it poses a fundamental challenge for organizations that depend on diverse perspectives for innovation.
The Experience Paradox
Mei, Marrone, and colleagues (2025) surface a different dimension of the same tension. Their study, titled "If ChatGPT can do it, where is my creativity?", finds that generative AI boosts creative writing performance while diminishing the subjective experience of creativity. Writers produce objectively better text with AI assistance, but they report feeling less creative, less ownership over their work, and less engagement with the creative process.
This experience-performance gap is not trivial. If the psychological rewards of creative work — the sense of flow, authorship, and personal expression — are eroded by AI assistance, then the performance gains may come at a cost that productivity metrics do not capture. A marketing team that generates better slogans but feels alienated from its own output may eventually produce less, not more, because the motivational engine of creative work has been undermined.
The Shifting Model of Co-Creativity
Medeiros, Cropley, and Marrone (2025), writing in the Journal of Creative Behavior, argue that the prevailing view is shifting from "AI is creative" toward a more balanced model of human-AI co-creativity. Their research investigates what happens when humans and AI collaborate on creative tasks, examining not just whether the output improves but how the collaborative process works.
The emerging picture suggests that the value of AI in creative work depends heavily on how the collaboration is structured. When AI is used as an idea generator — providing raw material that humans then evaluate, combine, and refine — the results tend to be positive. When AI is used as a finisher — producing polished output that humans merely approve — both quality and diversity suffer. The difference lies in whether the human remains the active creative agent or becomes a passive curator of AI suggestions.
This distinction has implications for how organizations integrate AI into creative workflows. A design team that uses AI to generate a hundred rough concepts and then selects, merges, and reimagines the most promising ones is leveraging AI's combinatorial power while preserving human judgment and diversity. A team that asks AI to produce finished designs and then chooses among them is outsourcing not just execution but creative direction — and the homogenization data suggests this approach will narrow rather than expand the creative space.
The Diversity Problem Is a Systems Problem
The homogenization effect identified by Holzner et al. deserves particular attention because it operates at a different level than individual performance. A single writer may produce better work with AI assistance. But when every writer in a newsroom uses the same model with similar prompts, the newsroom's collective voice converges. When every startup in an accelerator generates business models through the same AI, the portfolio loses the diversity that makes venture investing rational. When every student writes essays with AI assistance, the distribution of ideas in a classroom narrows.
This is not a failure of the technology but a structural consequence of how it works. Large language models are trained on the same data and optimized for the same objectives. Their outputs reflect statistical regularities in that training data. When these outputs become inputs to human creative processes at scale, they impose a subtle uniformity — not by preventing original thought but by anchoring it around a shared set of AI-generated starting points.
The implication is that the question "does AI make us more creative?" has two correct answers. For any individual, probably yes. For a population of individuals all using the same AI, probably no — because the gain in individual quality is offset by a loss in collective diversity. Managing this tension — capturing the augmentative benefits while preserving the diversity that drives genuine innovation — is emerging as one of the central design challenges for human-AI creative collaboration.