Chris Gash/theispot.com
What does artificial intelligence do to creativity? Are generative AI tools making us more creative, or less? Given that creativity is often the engine behind the most successful ideas and ventures, and that 83% of senior executives rank innovation among their top three priorities, understanding how using AI affects human creativity is critical for businesses.1 On the one hand, generative AI can act as a valuable brainstorming partner, enabling inventors and designers to rapidly prototype ideas and concepts. On the other hand, it risks inadvertently constraining creativity by narrowing the search space too early and encouraging users to anchor on AI-generated suggestions that seem “good enough.”
Across four recent studies, our research reveals that the truth lies beyond this simple binary. We have found that although AI can enhance individual creativity, it reduces collective creativity. To explain why this occurs, we should first clarify what we mean by creativity.
From Individual Creativity to Societal Innovation
Scholars typically define creativity as the intersection of novelty and usefulness.2 Novelty is the degree to which an idea or artifact is original or rare, and usefulness is the degree to which it is valuable or effective in achieving its purpose. An idea that is novel but useless or useful but unoriginal is not creative.
While these dimensions capture the creativity of a single idea, the dimension that best captures the creativity of a group of ideas is its diversity. Any collection of ideas may contain a few that are novel to some but obvious to others, or novel yet not useful. But a highly diverse set is more likely to contain a few highly original outliers that are both genuinely original and potentially valuable. This breadth provides teams with more raw material to recombine, compare, and refine over time, yielding products that better match the full range of customer preferences. After all, ideas that initially appear to be impractical can turn out to be breakthroughs once they have been refined or recombined: for example, the “failed” adhesive that became the Post-it Note or the abandoned video game whose internal communication tool became Slack. In other words, creativity requires more than just quantity and quality of output. It also requires diversity of output, where different ideas can spark new lines of inquiry, new speculation, and new seeds that breed new innovations.
By casting a wider net, organizations can guard against premature convergence on safe, conventional options, increasing the odds of surprising, high-impact breakthroughs. This, however, is where our research reveals an interesting paradox. Individually, AI often enhances creativity, particularly by enabling less experienced or less inherently creative individuals to generate more novel and useful ideas. But collectively, AI often “compresses” the idea space. Because many people anchor on similar AI-generated suggestions, outputs converge. A typical output produced with AI assistance is more creative, but the variance of the full set of outputs decreases. In short, even if an AI-inspired idea looks good, it may turn out to be similar to everyone else’s AI-inspired ideas.
For managers, the implication is profound. The challenge is to harness AI’s productivity and quality benefits while preserving the diversity of ideas that fuels long-term innovation.
Impact of AI on Idea Diversity
To understand how AI affects creative diversity, we analyzed four recent studies we worked on that span different creative domains: short-story writing, circular-economy solutions, humor caption contests, and collaborative storytelling. Despite the varied contexts, a consistent pattern emerged: AI assistance improved individual output quality while narrowing collective diversity.
In the first of these studies, two of us (Anil and Oliver) examined how access to AI influences the creative process and the diversity of collective output.3 In this experiment, participants were asked to write short, eight-sentence stories. Some people wrote entirely on their own, while others were given up to five three-sentence story seeds generated by an AI model. Independent evaluators rated the individual creativity of each participant’s story. We also used AI-based text analysis to measure the degree of semantic similarity among the stories, comparing those written with versus without AI assistance.
The results revealed the core tension. We found that AI assistance improved story novelty, especially for writers with a lower baseline level of creativity (as measured beforehand with an existing paradigm, the Divergent Association Task). Yet at the collective level, diversity declined. Stories from the AI-assisted groups converged on more similar beats or structures, showing less variance than those written without AI. (See the story-writing graphics.) This suggests a social dilemma: While individuals gain from AI, especially those who struggle most with creative tasks, widespread reliance risks narrowing the collective pool of ideas, leaving us with higher average quality but fewer distinctive outliers.
In a second study one of us worked on (Léonard, with four collaborators), this pattern held in a very different domain.4 We asked participants to propose circular-economy solutions to address sustainability challenges, such as repurposing waste materials. A human-only crowd produced a broad range of ideas, from conventional recycling proposals to unique, unconventional ones such as innovative bricks made from foundry dust and waste plastic with a Lego-like interlocking design to reduce construction-related air pollution. In contrast, a single human working with AI often surpassed the crowd in independent evaluators’ ratings of overall quality, strategic viability, and financial and environmental value. But the human crowd scored higher on novelty, and the unusual ideas that might spark breakthroughs emerged mostly from the human-only group. Once again, AI raised the floor of performance but narrowed the variance in outputs. (See the circular-economy graphic.)
A third study pinpointed where in the creative process this narrowing occurs.5 One of us (Kartik, with a co-researcher) conducted a randomized experiment modeled on the New Yorker cartoon caption contest to test whether AI reduced the diversity of creative outputs during the idea-generation or idea-evaluation stage. We tested four collaboration designs: human-only, AI use for idea generation alone, human idea generation with AI use for idea selection alone, and AI support during both phases. The study found that AI boosted both the quantity and average quality of ideas, with the greatest gains when it supported both the generation and evaluation stages. Importantly, the diversity effects diverged depending on when AI was used: AI in idea generation consistently reduced diversity, whereas AI in idea selection preserved variety at levels comparable to those of human-only work. (See the humor caption contest graphic). This finding suggests a potential way forward: The stage at which AI enters the workflow may matter as much as whether it is used at all.
A fourth study that one of us worked on (Kartik and a co-researcher) tested this insight directly.6 It examined how different human-AI collaboration models affect both the quality and the diversity of story writing, as well as the impact of those collaboration models on self-reported writer satisfaction. Four designs were tested: human-only, human-led ideation with AI drafting, AI-led creation with human approval, and continuous human-AI collaboration throughout (the “copilot” scenario). The results confirmed the pattern observed in the previous study: Ceding creative control to AI produced the most homogeneous outputs; having humans and AI work together as copilots mitigated the effect to some extent; and keeping humans in charge of early creative tasks preserved significantly more diversity — approaching the variety seen in fully human work. (See the Hosanagar and Ahn story-writing graphic.) This indicated a clear design principle: whether diversity survives depends on where humans are introduced into the workflow.
The graphs above visualize these findings across all four studies. To measure the diversity effect, we used AI techniques to calculate the similarity between the various outputs within each group and then averaged those scores. Think of it as a clustering metric: A higher similarity score means that ideas bunched together; a lower score means that they were spread out across a wider creative space. In each panel, the horizontal axis represents the average similarity score, and the distribution curve shows the frequency of that score among participants. When AI is involved, the curves consistently shift to the right — toward higher similarity — signaling that the outputs become more alike.
How to Use AI in Creative Workflows (Without Sacrificing Diversity)
The evidence across studies points to one conclusion: How you use AI in creative work matters as much as whether you use it at all. Leaders who seek the efficiency gains of AI while preserving or enhancing originality must intentionally design their workflows. Here are some practical strategies you can use.
1. Keep humans in the driver’s seat for ideation. The fourth study described earlier, which tested different modes of human-AI collaboration in story writing, offers direct guidance here. Participants who retained responsibility for ideation produced stories that were rated higher by independent evaluators in terms of interestingness and overall quality, and they reported greater satisfaction. The diversity effects were equally important: Ceding creative control to AI produced the most homogeneous outputs, while keeping humans in charge of early creative tasks resulted in significantly more diversity, approaching the variety seen in fully human work. Even the copilot model, which involved AI throughout, narrowed the diversity of output compared with human-led ideation.
This has immediate practical implications for managers: Let humans take the lead in any creative and innovative workflow to capture more unique ideas. Start by having a team sketch ideas or draft early outlines before integrating AI into the process. This sequencing preserves variety while still capturing efficiency — and ensures that AI complements, rather than substitutes for, the uniquely human capacity to make messy, surprising leaps.
Video game producer Ubisoft’s in-house AI tool, Ghostwriter, offers a concrete illustration of this sequencing in action. Designed to assist scriptwriters working on large open-world games, Ghostwriter takes on one of the most repetitive narrative tasks: generating first drafts of short lines of dialogue spoken by nonplayer characters. Crucially, the tool does not replace the writer’s role in shaping character or story; scriptwriters define the character and context first and then select and edit from among the AI’s generated variations. Human judgment remains in the driver’s seat throughout the process. The result is a workflow that frees writers to invest their creative energy where it matters most while AI absorbs the volume work downstream.
One other insight emerged consistently across the four studies: If what you are after is the greatest diversity of ideas possible, then humans are hard to beat. But diversity alone is not enough. In the circular-economy study, for instance, a single human iteratively working with AI produced solutions that scored higher in overall quality because targeted prompting enables rapid refinement toward practical value, even though a human crowd produces more diverse and novel ideas. Leaders should choose wisely when deciding how to employ human-only groups in their organization’s workflow.
2. Diversify AI inputs. Homogenization often stems from everyone using the same AI tool in the same way. Managers can push back against this by deliberately introducing variety: Rotate prompts, experiment with role-playing instructions (such as “Argue against this idea”), run parallel AI models, or integrate novel data sources.
Research supports the use of those tactics to increase idea diversity. For instance, chain-of-thought prompting, which involves asking the model to reason step by step before generating outputs, produces substantially greater dispersion in idea sets than plain-vanilla prompts and, in some cases, approaches the variance achieved by human groups.7
Complementing this, the circular-economy study demonstrated that when humans iteratively instruct a large language model to generate solutions distinct from previous iterations, they significantly enhance novelty without sacrificing value. That field study found that this human-guided differentiation approach, which explicitly prompted the model to “tackle a different problem than the previous ones and propose a different solution” after each output, produced solutions with novelty ratings comparable to those of human crowds while maintaining superior strategic viability and overall quality.
3. Deploy multi-agent and multimodel approaches. Just as diverse human teams surface a broader range of perspectives than homogeneous ones, diversifying AI “voices” can counteract convergence. Incorporating agentic workflows, where one AI generates ideas while another critiques them or multiple specialized agents tackle different aspects of a problem, can significantly broaden the search space.
Organizations can design systems where different models or agents address the same challenge from distinct perspectives. At the research frontier, multi-agent AI systems are beginning to assist with scientific discovery itself, generating hypotheses, critiquing them, and refining them in self-improving cycles. Some early use cases are already producing experimentally validated results.8 Colgate-Palmolive offers a practical illustration of this architecture in action. Rather than routing innovation work through a single AI interface, the company uses different AI systems in conjunction with one another: One mines consumer data to surface unmet needs, a proprietary AI generates product concepts, and a third uses “digital consumer twins” to simulate consumer reactions — with humans guiding each handoff.
The key insight is architectural: Rather than channeling all creative work through a single AI interface, organizations should build workflows that create productive tension across multiple AI perspectives. This approach takes advantage of AI’s efficiency while maintaining the divergent thinking that drives breakthrough innovation.
4. Build guardrails and mindful friction to protect human comparative advantage. The temptation will be to let AI act as a copilot everywhere. But if employees outsource their core creative tasks, they risk losing the very skills that make them distinctive. We recommend introducing some friction to AI use — small design choices that prevent people from becoming passive consumers of AI output. For instance, teams could be required to submit human-generated options before consulting AI, or justify why an AI-suggested idea should be selected.
There’s more at stake than just another good idea that might benefit the organization: These practices keep people’s creative muscles active while still harnessing AI’s efficiencies. Without such guardrails, efficiency gains will quickly become commoditized, leaving little basis for competitive differentiation and depriving the workforce of its ability to drive new ideas forward in an age when everyone will have access to AI. Research backs this up. A 2025 study found that students with unrestricted AI access performed significantly worse once that access was removed — but carefully designed guardrails eliminated this penalty.9 Similarly, in a different study, consultants who blindly adopted AI recommendations underperformed compared with those who maintained critical oversight.10
Ultimately, what will define organizations’ competitive edge in the years to come is their ability to cultivate a diverse set of creative ideas through human ingenuity, complemented by an efficient, research-backed workflow that uses AI’s capabilities at the right time to achieve superior quality and feasibility.
AI can support creativity, but only if humans engage actively and early in shaping the process. The organizations that stand out will not be those that use AI the most but those that use it most intentionally, designing AI use in ways that allow human originality and machine efficiency to amplify rather than cancel each other out.