If you build AI characters from real people's backgrounds, do they come up with the wide range of ideas a real crowd would?

This explores whether AI personas built from real people's backgrounds produce the full spread of ideas real people would, meaning variety across a whole population and not just plausible answers from each individual persona. The corpus has no study that measures creative diversity directly, so what follows is inference from nearby evidence. That evidence suggests personas capture the big, sturdy patterns of human behavior and are much shakier on the quiet variations where creative range lives.

Start with what personas do well. Viewpoints AI reproduced 84 of 111 main effects from published marketing experiments, and how well it replicated tracked how strong the original result was Can AI personas reliably replicate human experiment results?. On marginal effects it produced both false positives and false negatives. Creative diversity is mostly made of weak, low-frequency signals, such as the odd idea a small share of people would have. A method that is reliable on strong signals and noisy on weak ones is not obviously a diversity engine. The paper tested replication, not creativity, so this is a reading of the result rather than a finding.

There is also a reason to doubt that many distinct personas add up to many distinct outputs. Across three models, persona prompts made the models follow trait instructions but left between-group bias gaps unchanged. The persona works at the level of output and redistributes what comes out, without changing what is underneath Can persona prompts actually reduce bias in language models?. A cultural-critique note in the collection makes a related point. Independent LLMs converge on similar outputs even when nominally competing, and personalization hides this because each user only sees their own customized result Does AI homogenize culture the way mass media did?. Post-training also installs stable dispositions that persist under adversarial pressure Are RLHF personas performed characters or realized dispositions? Are LLM personas realized or merely simulated through training?. So a hundred personas could each look distinct on their own while being a hundred variations on one model's taste.

Some work does try to engineer diversity in, and it suggests one persona attribute is never enough. Realistic synthetic dialogue needed subtopic specificity, Big Five personality variation, and 11 contextual characteristics working together Can synthetic dialogues become realistic through layered diversity?. The 90.48% figure there measures in-domain dialogue performance, not creative range, so it shows layering helps realism and not that it recovers human variety. Grounding personas in real documents, as MAJ-EVAL does, anchors them in actual stakeholder perspectives instead of arbitrary roles Can personas extracted from documents generalize across evaluation tasks?. It was tested for reproducible evaluation, not for generating novel ideas.

The direction of effort in these results is telling. Consistency training cuts persona drift by over 55% Can training user simulators reduce persona drift in dialogue?, and scripted layered personas produce more human-like dialogue by restricting the model to reactive responses Can layered persona architecture sustain coherent character behavior?. Both reward staying in character, and none of the retrieved work rewards spread across a population. Real backgrounds seem to help personas reproduce what groups predictably do. Whether they reproduce what individuals surprisingly do is untested here. The missing experiment is to compare the spread of ideas from persona-conditioned models against real people's, and until it exists, persona diversity is claimed rather than shown.