Getting AI to brainstorm ideas versus getting it to polish yours might push creative variety in opposite directions.

This explores whether the way you bring AI into creative work changes how varied the results are: AI as the source of ideas versus AI as the editor of ideas that are already yours. The short answer from the corpus is yes. The clearest evidence is a preregistered writing experiment in which AI-generated ideas made the pooled set of ideas less diverse for every group of writers, while AI that only refined ideas people had already come up with left that diversity intact Does AI assistance homogenize or preserve creative diversity?. The surprising detail is about who loses. Non-native English speakers contributed more distinctive ideas than native speakers, and AI ideation was the only condition that wiped out that advantage. Polishing someone's prose didn't flatten their perspective. Letting the model supply the starting point did.

The corpus suggests a reason the stage matters so much. When AI generates, it tends to produce many well-formed claims that come from roughly one point of view, because it follows the probable patterns in its training data instead of reasoning from competing positions Does AI generate diverse claims or diverse perspectives?. Scale that up across thousands of users and you get a quieter version of the old mass-media worry. Each output feels personalised, but separate models drift toward the same material, so the sameness is hard for any one person to notice Does AI homogenize culture the way mass media did?. Refinement escapes this because the variety is already present in the human input, and the model works within it instead of replacing it.

There is one caution in the corpus. A different study is often cited for the claim that AI narrows ideas, but it actually measured only the number and detail of ideas, which rose, especially for less experienced writers. It never measured diversity at all Does AI assistance actually narrow the diversity of ideas?. So "AI gives you more ideas" and "AI gives everyone the same ideas" can both be true, and a study that counts ideas can't tell you which is happening.

Research on model training points to the same split from a different angle: the effect on diversity depends on what the process rewards. Preference tuning (RLHF) reduces variety in code, where everything converges on correct answers, but increases it in creative writing, where distinctiveness gets rewarded Does preference tuning always reduce diversity the same way?. Critique-style feedback during training keeps a model's solutions from narrowing too early Do critique models improve diversity during training itself?. That roughly parallels the human finding. Generating tends to pull toward the most probable answer, while critiquing and refining can protect the range of options. In multi-agent ideation, varied viewpoints only improve results when the agents also have real expertise. Without it, diverse teams do worse than a single competent agent Does cognitive diversity alone improve multi-agent ideation quality?.

The takeaway you might not have expected: if you care about original work, the order of operations matters more than how much AI you use. Bring your own rough ideas first and use AI to sharpen them. The variety in a group's output comes from the people in it, and the evidence suggests AI-generated starting points are where that variety gets lost.