A four-study experimental paper from Columbia Business School (February 2026) documents that independent LLM sessions produce materially less idea diversity than equivalent human groups, attributing this to two mechanisms: within-session fixation (comparable in magnitude to humans) and cross-session knowledge aggregation (substantially worse than humans). The paper then demonstrates that combining ordinary diverse personas with Chain-of-Thought prompting causes LLM collective output to exceed human diversity by up to 26% on a combination metric, while CoT alone or personas alone each produce partial improvements. For financial institutions evaluating AI-assisted ideation, research, or strategy-generation workflows, this establishes that default LLM deployments carry a structural homogeneity risk at the portfolio level, and that specific prompt architectures materially change that risk profile.
293 other tracked developments touching the same claims.