Overview
This series explains why AI-assisted brainstorming can look broad while narrowing too early: shared model priors concentrate independent starts, sequential context anchors later ideas, and agent interaction can synchronize trajectories through authority and visibility. It separates these failure points, tests what temperature, seeds, personas, staged revision, model diversity, and topology actually change, and proposes a workflow that preserves independent alternatives before critique and synthesis. The evidence comes mainly from controlled GPT-4o fitness-product ideation and multi-agent scientific-proposal studies, so diversity measures are not direct measures of creativity, usefulness, or safety.
Learning outcomes
- Fluent outputs do not prove broad exploration; many polished ideas can represent one underlying direction.
- In the tested fitness task, the largest human–LLM gap appeared across sessions and first ideas, not within ten-idea sessions.
- In the tested GPT-4o procedure, human seeds did not widen downstream diversity; ordinary personas widened sessions but increased fixation.
- Different interventions target different losses: personas or models vary starting regions, staged revision interrupts fixation, and topology or authority changes coupling.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Why can AI brainstorming produce many ideas but little real variety?
Because fluency can elaborate a shared prior or early premise, producing many phrases that occupy one conceptual direction.
How can you tell whether AI brainstorming is fixated or merely convergent?
Measure where narrowing occurs: slow new-category accumulation suggests fixation, clustered first ideas suggest shared starting points, and post-interaction collapse suggests coupling.
Does increasing temperature make AI brainstorming more diverse?
Not reliably: in the tested fitness task it produced only a small gain, while a higher setting damaged interpretability and did not create independent knowledge regions.
How should teams structure AI brainstorming before consensus?
For open-ended tasks, generate independently with varied cues or models, delay sharing, use staged comparison to interrupt fixation, then synthesize after diversity and quality checks.