AI models from rival labs are becoming more alike in their creative output, according to new research from Duke University, raising questions about whether LLMs are getting less useful as creative partners.
The Duke team put 69 models from 12 provider families, spanning releases between 2023 and 2026, through both open-ended real-world questions and a classic creativity exercise that asks for unusual uses of everyday objects, and measured a statistically significant slide in output diversity.
The team lays out the case in a paper whose title poses the question directly, “Are LLMs becoming similarly creative? Evidence from three years of models,” and reports that responses to their open-ended prompts have grown increasingly similar over time.
Western and Chinese labs both populate the roster: Anthropic, Cohere, DeepSeek, and Google are joined by Meta, MiniMax, Mistral AI, Moonshot AI, OpenAI, Qwen, xAI, and Z.ai. The analysis gauged how far apart model answers sat in meaning, and whether each newer generation of releases drew those distances tighter.
They caution that even strong creative performance could still confine users to a narrower band of possibilities, and with it, narrower thinking of their own.
The trend extends earlier findings of algorithmic monoculture, including a Cornell University study in May that found unrelated models sharing strange fixations on lighthouse keepers and anachronistic names. The Duke team calls its results preliminary but warns that converging creativity could narrow the ideas users encounter, and the thinking those ideas inspire.
