Duke University researchers tested 69 AI models across 12 provider families—including Anthropic, OpenAI, Google, Meta, and DeepSeek—released between 2023 and 2026, and found a statistically significant decline in creative output diversity. Using the Alternate Uses Task and 100 real-world prompts, responses became increasingly similar over time. The authors attribute possible convergence to overlapping training data and similar optimization objectives, while noting models may later diverge. They caution the trend could limit users' exposure to diverse ideas.
No score is assigned. Sources and their independence are shown in the citation chain below.