Rival AI Models Are Producing Increasingly Similar Outputs

Researchers from Duke University have discovered that the creative outputs of major large language models are converging, making their responses increasingly homogenous over time. In a paper titled "Are LLMs becoming similarly creative? Evidence from three years of models," the team evaluated 69 models across 12 provider families, including Anthropic, Cohere, DeepSeek, Google, Meta, MiniMax, Mistral AI, Moonshot AI, OpenAI, Qwen, xAI, and Z.ai. Spanning releases from March 2023 to July 2026, the study suggests that competing AI systems are beginning to offer highly similar ideas when prompted for creative tasks.

To measure this convergence, the researchers tested the models using two distinct datasets: the Alternate Uses Task, which asks for unconventional uses for everyday objects like a book, shoe, or hammer, and 100 open-ended prompts from the Infinity-Chat100 dataset. Operating through OpenRouter's API with temperature and top-p both set to one, the researchers converted the responses into numerical vectors using the all-MiniLM-L6-v2 sentence-transformer. Across 1,000 rounds of balanced resampling, the semantic distance between different providers' answers shrank. For the Alternate Uses Task, the average semantic distance fell from approximately 0.50 in the earliest models to below 0.40 in the newest versions. Similarly, the Infinity-Chat100 distance dropped from about 0.34 to just over 0.32.

For practitioners and developers, this shift toward an "algorithmic monoculture" presents a significant hurdle when building applications that rely on unique or diverse generative outputs. The researchers hypothesize that this homogenization stems from overlapping training datasets and identical optimization objectives, which force models to organize concepts in similar ways. As a result, developers utilizing LLMs for brainstorming, creative writing, or academic assistance may find that different model families yield nearly identical suggestions. This lack of variety could ultimately "bound the range of possibilities LLM users are exposed to," as the authors warn, limiting the utility of these systems as true creative partners.