According to Ethan Mollick on X, a new peer-reviewed working paper shows GPT-4 can produce idea sets with diversity approaching that of human groups when guided by better prompting and contextual scaffolds, countering the claim that AI is inevitably homogenizing. As reported by the SSRN paper by Mollick and colleagues, default GPT-4 outputs tend to be similar, but structured prompts, role instructions, and iterative selection significantly increase variance while maintaining high average quality (source: SSRN working paper 4708466). According to the authors, this creates practical opportunities for product ideation, marketing concept generation, and R&D portfolio exploration where firms can scale both quality and variety at low marginal cost, provided they use prompt engineering and human-in-the-loop review. As reported by the paper, teams can operationalize this by running multiple GPT-4 prompt regimes in parallel, seeding with distinct contexts, then using ranking and clustering to assemble diverse, high-quality idea pools for downstream testing.