AI Models' Creative Outputs Becoming Increasingly Similar Across Major Providers

New research from Duke University has found that 'creative' output across a wide range of leading AI models is becoming more similar over time. [1] The authors tested 69 models across 12 provider families, covering releases from 2023 to 2026. [1] This testing revealed a statistically significant decline in output diversity across both real-world open-ended questions and a standard creativity test. [1] The provider families tested included Anthropic, Cohere, DeepSeek, Google, Meta, MiniMax, Mistral AI, Moonshot AI, OpenAI, Qwen, xAI, and Z.ai. [1]

The authors of the new paper state that LLM responses to the open-ended prompts they tested have become increasingly similar over time. [1] They suggest that similar ideas may be offered in response to 'creative' requests, increasingly, among all the major LLM providers. [1] This suggests that LLMs are becoming less creative in tasks that involve generating open-ended responses. [1] According to the authors, this development demands scrutiny of the models' long-term usefulness as creative assistants. [1]

The Duke University study points to a concerning trend for professionals relying on AI for ideation: the potential homogenization of thought, sometimes referred to as 'algorithmic monoculture'. As models from diverse providers like Anthropic, OpenAI, and Google converge on similar outputs for open-ended tasks, the utility of these tools for genuine creative brainstorming may diminish. Users might need to employ more aggressive prompting strategies or look beyond standard LLMs to avoid echoing standard, similar ideas. What the sources don't address: whether this convergence is due to shared training data sets, similar alignment techniques, or another underlying technical factor.