Generative AI creativity can make an individual’s work stronger while pulling different users towards more similar ideas. That is the central finding from a new peer-reviewed meta-analysis of human-AI co-creation. Across 19 studies and 61 measured effects, people working with generative systems produced outputs that were modestly but reliably more alike than work created without AI.

The result complicates the familiar argument that AI either unlocks creativity or destroys it. Both claims are too crude. A tool may help one person develop a better story, image or concept while simultaneously reducing the range of ideas produced by a group. Personal improvement and collective uniformity can happen at the same time.

How generative AI creativity was compared

The review was conducted by Alwin de Rooij of Tilburg University and Michael Mose Biskjaer of Aarhus University. Their article, published in Behaviour & Information Technology on 31 August 2026, examined controlled studies that compared people creating with generative AI against people completing comparable tasks without it.

The researchers synthesised 19 studies containing 61 effect sizes. The tasks included divergent thinking, idea generation, writing and visual creation. Rather than asking only whether an AI-assisted output was good, the analysis measured homogenisation: whether outputs from different people became more semantically similar to one another.

The pooled effect was small but statistically significant and remained present across sensitivity checks. The authors found no evidence of publication bias in the available literature. An institutional record for the peer-reviewed paper reports that the effect was stronger in semantically constrained ideation tasks and also appeared in preliminary real-world and post-use analyses.

Small does not mean trivial, particularly when the same systems are used by millions of people. A slight convergence in one brainstorming session may barely register. Repeated across advertising, education, product development and entertainment, it could reduce the variety of ideas entering the wider culture.

Better work can coexist with collective sameness

Generative tools often help users overcome a blank page, produce more options or polish an unfinished concept. That individual benefit is compatible with homogenisation because quality and diversity are different measurements. Ten people may each improve their first draft while moving towards overlapping themes, phrases or visual conventions.

This helps explain why studies of human-AI collaboration can appear contradictory. LiveAIWire previously found that generative AI was already reshaping creative industries through faster production and new forms of experimentation. Those gains do not guarantee a broader range of final outputs.

The meta-analysis does not conclude that people lose creativity as a personal capacity. It examines similarity among their products. A user can still make original decisions, reject suggestions and transform material substantially. The concern emerges when many users begin from related model outputs or accept the same high-probability associations.

That distinction matters for businesses assessing AI productivity. Counting how many concepts a team generates or how quickly it completes a campaign will not reveal whether the ideas occupy genuinely different territory. A faster workflow may conceal a narrower portfolio.

A separate meta-analysis reached a compatible warning about performance: working with AI was better than working alone, but not always best. Assistance changes the system around a task. Measuring only the assisted individual’s improvement can miss what happens to the team, the strongest alternative or the range of outputs.

Structured brainstorming showed stronger convergence

The clearest homogenisation appeared in ideation tasks with defined semantic constraints. When participants were asked to solve a particular problem or generate ideas inside a narrow brief, AI assistance pulled their answers closer together more strongly than it did across the other broad task categories.

That pattern is plausible because a model responds to common prompts by drawing on similar learned associations. If several people ask for ideas within the same constraints, each may receive variations from a densely populated part of the model’s output space. The model can offer many answers while still directing users towards related conceptual neighbourhoods.

This is an interpretation of the evidence, not a mechanism the meta-analysis proves. Similarity could also reflect prompt design, the tendency to accept early suggestions, limited time or evaluation rules that reward conventional answers. Different models, interfaces and user groups may produce different levels of convergence.

The finding nevertheless gives practical weight to an older philosophical question explored in LiveAIWire’s examination of whether machines can be truly creative. The operational issue is no longer only whether a model can surprise us. It is whether a shared model surprises different people in sufficiently different ways.

Shared systems can become semantic anchors

An AI suggestion does more than add an option. It creates an anchor. Once a response names a theme, structure or metaphor, users may explore around it instead of searching elsewhere. The output can shape what feels relevant before the user has formed an independent view.

That influence can persist even when people edit heavily. Changing words, colours or surface details may leave the underlying concept intact. Two finished pieces can look distinct while sharing the same premise, emotional arc or solution strategy.

Scale intensifies the concern. A model used inside one team can concentrate that team’s ideas. A handful of dominant models used across an industry can also concentrate the reference points available to competitors, freelancers and customers. Homogenisation then becomes an ecosystem effect rather than a problem inside one document.

The authors describe the outcome as a reorganisation of creative diversity, not evidence that creativity has disappeared. That language is useful. AI can widen access to creative production and raise the floor for individuals while compressing differences across the resulting work.

What this means for creators and organisations

The first practical step is to separate divergent and convergent phases. Generate initial ideas independently, then use AI to expand, test or refine them. If the model appears before people have explored their own directions, its first suggestions may define the boundaries of the discussion.

Teams can also compare concepts by meaning rather than volume. Twenty slogans built around the same promise are not twenty strategic directions. Grouping outputs by premise, audience and emotional appeal makes hidden similarity easier to see.

Prompt variation helps, but it is not a complete answer. Asking the same model to act as five different experts may produce stylistic diversity while preserving shared assumptions. Greater variety may require different source materials, different models, human research and contributors with genuinely different experiences.

Organisations should preserve a route for unassisted work when originality is a core objective. The cost of doing so can look inefficient in a dashboard, especially when AI accelerates routine production. Yet removing every slower stage may also remove the moments in which unusual ideas emerge.

How to use AI without flattening every idea

A useful workflow begins with private generation. Each participant records a view before seeing model suggestions or colleagues’ answers. AI then becomes one contributor among several, rather than the common starting point for the entire group.

The next stage should ask the model to challenge its own centre of gravity. Users can request assumptions that have been ignored, concepts drawn from distant domains or reasons why the most obvious proposal might fail. The purpose is not to make prompts theatrical. It is to move exploration away from the first cluster of plausible answers.

Evaluation should reward difference where difference matters. A news headline, safety procedure or customer-service reply may benefit from consistency. A brand concept, scientific hypothesis or product strategy may suffer when everyone converges. There is no universal optimum for diversity.

AI use can also create fatigue when every task becomes a cycle of generating, sorting and correcting. LiveAIWire has reported on businesses pulling back from exhausting AI workflows. Adding diversity checks should improve decisions, not become another layer of low-value prompting.

Important limits remain

The evidence base is still young. Most included research concerned text-based systems and large language models. Music, embodied design, specialist professional tools and other creative forms were underrepresented. Results from current products may not generalise to future systems with different training data or interaction designs.

Real-world and post-use findings were based on smaller subsets of the evidence. They suggest that convergence may persist after AI assistance ends and may appear outside laboratory conditions, but those conclusions remain tentative. Longitudinal studies are needed to learn whether users adapt, resist or become more dependent over time.

Meta-analysis also combines studies that define tasks and similarity in different ways. The peer-reviewed publication improves confidence that an overall pattern exists, but it does not supply one universal percentage for every creative domain. The earlier open preprint record and final publication record allow readers to follow how the work developed.

Creative diversity becomes a design problem

The study’s most useful contribution is to shift the question from whether AI is creative to how human-AI systems distribute creativity. A model can help a person produce something better without helping a population produce something broader.

That makes diversity an engineering and editorial objective. Interface designers can delay suggestions, expose competing directions or help users compare semantic clusters. Managers can protect independent thinking before group convergence. Creators can decide when shared conventions are valuable and when they are a warning.

Generative AI does not automatically make every idea identical. The evidence points to a smaller, more subtle pull towards similarity. Because that pull can coexist with obvious individual benefits, it is easy to miss. The organisations that measure both quality and variety will be better placed to keep the speed of AI without surrendering the range of human imagination.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.