AI Convergence

Open-ended prompts (questions without a single correct answer) reveal that LLMs struggle to generate diverse responses, raising concerns that over time people's ideas and expressions may become homogenized; a risk characterized as an "Artificial Hivemind." (1) Within a single model, even with aggressive sampling, responses show significant intra-model repetition (e.g., average similarity remains high across multiple queries), and (2) across different models, surprisingly similar outputs demonstrate inter-model homogeneity. Even diversity-oriented decoding methods like min-p decoding reduce repetition but still leave substantial similarity, suggesting that "decoding alone cannot solve the problem." To diagnose and mitigate this issue, improvements are needed in data, training, and evaluation (pluralistic alignment reflecting diverse preferences).