Using AI to brainstorm makes everyone's ideas converge, but using it to polish your own ideas doesn't. Where should it go?

This explores where in a creative process AI should and shouldn't be used, so that a group of people still produces a wide range of ideas. The corpus gives a specific answer: put AI after the idea, not before it. In a preregistered experiment, AI-generated ideas reduced collective diversity for every writer, while AI that refined ideas people had already written left diversity intact Does AI assistance homogenize or preserve creative diversity?. Non-native English speakers contributed more diverse ideas than native speakers. Only AI ideation erased that advantage. So the rule is about the stage of the workflow, not about how much AI gets used.

The difficulty is that ideation is where writers lean on AI most. In a study of 18 writers, LLM use was heaviest during ideation, then organizing thoughts, then drafting. Writers also returned to ideation whenever they were blocked, and unexpected outputs sent them in new directions How do writers use AI through different creative stages?. The risky moment is the stuck moment, when an AI suggestion is most tempting. A diversity-protecting workflow has to plan for that moment, not just assume writers will hold off.

Switching models or adding more of them doesn't fix this. INFINITY-CHAT tested 70+ models on 26K open-ended queries and found an 'Artificial Hivemind'. Different models independently produced strikingly similar or identical answers, because their training data and alignment procedures overlap Do different AI models actually produce diverse outputs?. The sameness is also hard to spot from inside a single chat. Each user gets an output that looks personalized, so the homogeneity only shows up in aggregate Does AI homogenize culture the way mass media did?. Simulated crowds of AI agents don't rescue you either: teams with varied perspectives but no real domain expertise did worse than one competent agent Does cognitive diversity alone improve multi-agent ideation quality?. One partial counter-move is inside a single model. Structuring its reasoning as a dialogue between distinct voices produced more varied problem-solving strategies than monologue Can dialogue format help models reason more diversely?. That result is about reasoning tasks, not group creativity, so it points a direction without proving anything. In practice, the variety has to come from people with real expertise, and AI should be supporting them.

There is one more reason to keep the idea in human hands. AI can produce the finished form of an intellectual product without the person's own reasoning behind it Does AI separate intellectual form from the thinking behind it?. Refining someone's idea keeps their thinking in the piece, while generating the idea replaces it.

Two visibility tools show up in the corpus. Paired writers strongly preferred shared editors that reveal each other's prompting, valuing awareness of when, how, and where AI was used, though some found full sharing intrusive Do writers want to see each other's AI prompts in shared editors?. Process data can also flag AI use. AI contributions arrive in concentrated bursts outside an author's normal rhythm, so wholesale delegation is detectable, while ordinary collaborative use looks the same as minimal assistance Can process data distinguish AI delegation from ordinary collaboration?. Together they could catch the case of AI writing the idea, but not quieter influence. The corpus doesn't test whether either tool actually preserves diversity, so treat them as ways to see what's happening, not as proven safeguards.