Using artificial intelligence for creative tasks tends to make human output more uniform on a collective level. A recent preprint study provides evidence that while these tools might boost individual performance, they contribute to an overall reduction in the diversity of ideas across different users. This widespread reliance on automated assistance could lead to a narrower range of concepts in collaborative environments.
Generative artificial intelligence refers to computer programs capable of creating new text, images, or other media based on user instructions. The most common of these tools rely on large language models. Developers build these models by feeding them billions of sentences from the internet, allowing the software to recognize patterns and predict how words should follow one another.
Since many users interact with similar systems trained on overlapping data, scientists have raised concerns about how this technology shapes human thought. Researchers Alwin de Rooij, assistant professor in creativity research at Tilburg University and associate professor at Avans University of Applied Sciences, and Michael Mose Biskjaer, associate professor in design creativity and innovation at Aarhus University, designed a new study to assess these concerns. They noticed that previous research often focused on how these tools help individuals work faster or overcome temporary mental blocks.
They wanted to know if this individual assistance comes at a collective cost. “There are growing concerns that using Generative AI may lead people toward similar creative ideas,” the authors explained. “While AI can enhance creativity at the individual level, these benefits might come at a cost for creativity at a collective, or even societal, level.”
The authors sought to answer whether generative software makes people think alike. “We sought to address this by conducting a systematic review and meta-analysis of 19 empirical studies,” they noted. “More concretely, we wanted to examine whether and to what extent generative AI use is associated with convergence at the level of creative output, such as people’s ideas, designs, and creative writing.”
A meta-analysis is a statistical technique that combines the results of multiple independent studies to find common patterns or overall trends. By pooling data from various experiments, scientists can draw more robust conclusions than they could from a single test. The authors searched academic databases for studies published between 2022 and early 2026.
This time frame covers the period following the public release of popular chatbots, capturing the first wave of empirical research on this topic. The researchers selected 18 eligible articles containing 19 distinct experimental studies. These studies provided a total of 61 individual effect sizes, which are mathematical values indicating the strength of a specific phenomenon.
To be included in the analysis, the original experiments had to compare humans working with generative software against humans working alone. The original studies measured homogenization using several techniques. Many relied on advanced text analysis tools that translate written responses into mathematical coordinates.
This process allows computers to measure the semantic distance between words, essentially calculating how closely related different ideas are to one another. Other studies used human experts to rate the variety of meanings produced by participants. The analysis revealed a statistically significant homogenization effect associated with the use of artificial intelligence.
When people co-created with these systems, their final products tended to be more similar to the work of other users. “The meta-analysis shows that using generative AI can indeed lead people to think alike,” the authors noted. “Across individuals, AI use tends to make ideas, designs, and creative texts more similar to one another.”
“This suggests that AI may contribute to a form of homogenization of creative thought at the collective level,” they continued. “Importantly, this does not necessarily reflect a failure of human-AI co-creation but may instead be an inherent feature of how these systems currently support creative work at scale.”
The scientists also evaluated whether the type of task influenced the degree of uniformity. They categorized the experiments into four groups, which included divergent thinking, idea generation, writing, and visual art. Divergent thinking tasks are highly open-ended exercises, such as asking someone to list creative uses for a paperclip.
Idea generation tasks provide more specific constraints, such as asking for solutions to improve public transportation. The analysis showed that the homogenization effect was strongest in the idea generation tasks. Because these exercises require specific solutions to defined problems, users likely rely more heavily on the predictable suggestions provided by the computer algorithms.
The researchers did not find strong statistical evidence for differences among the other three categories, suggesting that open-ended tasks lead to less convergence. They also checked if these patterns only happen in highly controlled laboratory settings. The authors compared traditional laboratory experiments with real-world scenarios, such as analyzing published essays and visual artworks created before and after the widespread adoption of automated writing tools.
The analysis of these real-world conditions showed a small but significant reduction in idea diversity. “In many ways, the findings resemble classic fixation effects from the psychology literature, where exposure to examples constrains later thinking, but here they appear amplified by the scale and synchronicity of generative AI model use,” the researchers stated. “This homogenization effect was observed not only in controlled lab studies but also in real-world quasi-experiments. This suggests that it is not merely a lab-based phenomenon, but a practical concern affecting concrete creative processes and practices.”
De Rooij and Biskjaer also investigated whether this narrowing of ideas persists after a person stops using the software. They isolated a subset of studies that tested participants on new creative tasks after their initial interaction with the computer models. The results suggest that the homogenization effect carries over into these subsequent activities.
“The findings also provide preliminary evidence that homogenization effects may persist beyond moments of direct AI use,” the researchers told PsyPost. “In other words, interacting with these generative AI systems may shape how people think and generate ideas even after the interaction has ended. This potential ‘rub-off’ effect on creative cognition warrants further research and is something we would like to explore in more depth.”
These results closely align with another recent study published in the journal PNAS Nexus. Scientists Emily Wenger and Yoed N. Kenett tested how large language models affect human creativity by evaluating 22 different commercial chatbots. They recruited 102 human participants to complete a series of verbal creativity tests, including the alternative uses task, and then asked the chatbots to complete the exact same assignments.
Wenger and Kenett found that individual language models performed at or slightly above the level of the average human on most exercises. When viewed in isolation, a single chatbot provided highly original and creative responses. However, when the scientists compared all the responses from the different models, a stark pattern of similarity emerged.
Across all tasks, the computer programs produced answers that were significantly more alike than the answers provided by the human participants. Both sets of researchers point to similar underlying mechanisms for this phenomenon. Because the major technology companies train their models on massive, overlapping datasets scraped from the internet, the programs naturally gravitate toward the most statistically common word associations.
When thousands of people use these tools to generate ideas, the software acts as a semantic anchor. The models pull human users toward a shared set of typical concepts, reducing the overall variety of ideas. Wenger and Kenett attempted to fix this issue by adjusting the internal settings of the chatbots to force more random text generation, but this caused the models to produce nonsensical sentences.
Readers should avoid interpreting these findings as proof that human beings are becoming entirely uncreative. De Rooij and Biskjaer note that the reduction in collective diversity does not equal a total loss of individual ability. “A key point is that our findings do not show that using AI reduces creativity,” the researchers emphasized.
“Rather, they point to a shift in where and how creative diversity occurs, and where it may be constrained,” the authors said. “Individual output can improve in creative quality while becoming more similar across people. While these effects are often subtle in single instances, they may become meaningful when considered at the scale at which generative AI is now being used.”
The authors point out some limitations to their current analysis. The review primarily focuses on text-based tools and large language models, meaning the findings might not apply to other types of computer systems. For instance, adaptive machine learning programs or tools used for music composition were not adequately represented in the available data.
This restricts how broadly the scientific community can apply these conclusions across different artistic domains. Additionally, the analyses regarding long-term persistence and real-world applications relied on relatively small groups of studies. The limited data makes these specific conclusions tentative and open to revision.
Future research should explore different forms of human and machine collaboration over extended periods of time. “An important next step is rethinking how generative AI systems are designed and used in creative contexts to mitigate homogenization effects,” the authors noted. “This includes exploring alternative workflows, interaction designs, and creative strategies that sustain diversity rather than encourage early convergence.”
“One step in this direction has already been taken by mapping creative strategies for working with generative AI and machine learning, based on analyses of AI art practices,” they added, referencing a recently published article outlining this approach. “We believe these strategies can transfer to other creative domains.”
The preprint study, “Does Generative AI Make Us Think Alike? A Systematic Review and Meta-Analysis of Homogenization Effects in Human-AI Co-Creation,” was authored by Alwin de Rooij and Michael Mose Biskjaer.