Different AI models keep giving strikingly similar answers, and each one is narrow on its own. So does switching models really buy you variety?

This explores whether cross-model sameness and within-model narrowness are one phenomenon or two. The corpus suggests they stack. Each model is narrow on its own, and the narrow regions overlap, so switching to a different model buys much less variety than you'd expect.

Start with the cross-model side. One study ran 70+ models on 26K open-ended queries and found an 'Artificial Hivemind': models independently produced strikingly similar or even identical responses, which the authors trace to overlapping training data and alignment procedures (Do different AI models actually produce diverse outputs?). One consequence is that an ensemble of different models is a weaker diversity hedge than it sounds. The same pattern shows up in values. Across 106 models and 625 scenarios, the models bunched into a tight region of value space while human respondents scattered widely (Do large language models actually reflect human value diversity?). Each model has a coherent worldview, and the worldviews look alike.

Inside a single model the picture is different. A model isn't one voice. It holds a superposition of many possible characters, and each response samples one of them, which is why regenerating can give you a different personality that still fits the conversation (Does an LLM commit to a single character or maintain many?). That range shrinks as the conversation goes on, and it also seems to shrink with scale. For synthetic data generation, models around 500M parameters produced more unique outputs per sample than larger ones, because larger models pile probability onto their preferred outputs (Why aren't bigger models better for generating diverse outputs?). The opposite extreme is temperature zero, which returns the same answer every time. That answer is still just one draw from the distribution, so consistency doesn't mean it's representative (Does setting temperature to zero actually make LLM outputs reliable?). Within a model, narrowness is about how probability mass is spread. Across models, the finding is about where different models put that mass, and it lands in similar places.

The corpus points to a shared pipeline as the reason the two line up. Controlled experiments show RL post-training amplifying one dominant format from pretraining within the first epoch while collapsing the alternatives. Which format wins depends on model scale, and this is mostly hidden when you start from proprietary pretrained models (Does RL training collapse format diversity in pretrained models?). Homogenization also starts before generation. Users rephrase their prompts toward the higher-frequency forms a model handles best, so distinctive inputs get flattened on the way in (Does high-frequency text homogenize user input before generation?). Similar data, similar alignment and similar user habits would each narrow every model in the same direction.

How narrow a model looks depends on what you count. Preference tuning reduces lexical-syntactic diversity in code but increases it in creative writing, because each domain rewards something different (Does preference tuning always reduce diversity the same way?). When diversity is measured only among outputs that pass a quality check, tuned models come out more diverse than base models. Base models look varied mainly because their variance includes incoherent text (Does preference tuning actually reduce the diversity of model outputs?). The corpus has no study that measures cross-model and within-model diversity on the same yardstick. So it can't say which problem is bigger. It does suggest that neither more samples from one model nor more models on the same task escapes the region they all share.