A new study argues that AI language models may be less creatively diverse than many users expect. Researchers at the University of Washington, Carnegie Mellon University, and the Allen Institute for AI found that different systems often move toward the same ideas, even when the prompt leaves room for imagination.
The team, led by Liwei Jiang, describes the pattern as the "Artificial Hivemind." The concern is not only that a single chatbot may repeat itself, but that models from different companies can produce outputs that look strikingly alike.
What the study found
The researchers examined how language models respond to open-ended prompts. These are the kinds of tasks where users might reasonably expect variety: metaphors, product descriptions, creative writing, educational help, or other responses that do not have one fixed answer.
The study separates the issue into two related patterns. One is intra-model repetition, where the same model returns very similar answers across multiple attempts. The other is inter-model homogeneity, where different model families produce similar responses despite coming from different developers.
One example is the prompt "write a metaphor about time." The researchers asked 25 different language models to generate 50 responses each. Instead of a wide spread of ideas, the results formed only two dominant clusters.
One cluster centered on "time is a river." The other used variations of "time is a weaver." The phrasing changed from output to output, but the concepts were repeatedly concentrated around the same images.
Why model similarity matters
Similarity inside one model is already important for users who ask for alternatives, drafts, or brainstorms. If many answers are only lightly different from each other, the model may create the impression of choice while offering a narrow set of ideas.
The study suggests the issue can extend across model families. To measure this, the researchers introduced Infinity-Chat, a dataset of real user queries. The source article reports that in nearly four out of five test cases, responses from the same model were so similar that they were barely distinguishable.
The more surprising finding is that overlap also appeared between different systems. When asked to write a product description for iPhone cases, DeepSeek-V3 and OpenAI's GPT-4o produced identical phrases, including "Elevate your iPhone with our," "sleek, without compromising," and "with bold, eye-catching."
The study reports an average similarity of 81 percent between DeepSeek-V3 and OpenAI's GPT-4o. DeepSeek-V3 and Qwen's qwen-max-2025-01-25 reached 82 percent overlap.
Those figures matter because the models were developed by different companies on different continents. The study does not claim a settled explanation for why this happens, but it raises the question of whether today’s AI systems are becoming more alike than their branding suggests.
Possible causes remain unresolved
The researchers do not present a confirmed cause for cross-family convergence. According to the source article, they point to several possibilities that still need causal analysis.
- Shared data pipelines could push models toward similar source material.
- Contamination from synthetic data could reinforce patterns already generated by AI systems.
- Overlapping alignment practices could steer different models toward similar wording, tone, or structure.
These explanations are presented as speculation, not settled findings. That distinction is important. The study identifies a pattern and warns about its possible effects, but it does not prove which mechanism is responsible.
For users, the practical takeaway is still clear: asking several AI language models for options may not guarantee real diversity. If the systems have converged in training data, synthetic data, or alignment behavior, their answers may cluster around the same familiar concepts.
The cultural risk
The authors worry that model-level convergence could eventually influence human expression. The source article frames the risk as a gradual homogenization of human thought through repeated exposure to similar AI outputs.
This matters because billions of users increasingly rely on language models for creative, educational, and decision-making tasks. If those tools repeatedly present similar metaphors, structures, and assumptions, users may absorb those patterns and reuse them.
The study points to existing evidence of measurable changes in human writing styles and creative thinking since ChatGPT's widespread adoption. It also raises a cultural concern: if models favor dominant expressions, other traditions and worldviews may receive less space.
The example given is a Western-centric metaphor such as "time is a river." If that kind of expression becomes the standard AI answer, alternative ways of describing time may be pushed aside. The issue is not only originality in a narrow artistic sense, but the range of cultural ideas that remain visible.
AI researcher Andrew J. Peterson made a similar argument in 2024, warning of a knowledge collapse driven by the AI boom. In the context of this study, the concern is that repeated AI outputs may not simply reflect culture, but also reshape what people produce next.
Limits for synthetic data
The findings also affect a common assumption about synthetic data generation. One strategy for increasing diversity is to use multiple models or ensembles, expecting that different systems will contribute different perspectives.
But if the underlying models are already homogeneous, that strategy may deliver less variety than expected. A multi-model setup can still look diverse on the surface while recycling the same concepts underneath.
For anyone using AI language models for brainstorming, education, writing, or synthetic data, the study points to a simple caution. More outputs are not always more ideas. The deeper question is whether those outputs actually widen the field of thought or quietly pull it toward the same center.