Your team is probably using AI already. ChatGPT for drafting proposals, Claude for copy, Perplexity for research. Individual productivity is definitely up. But if someone at your last team meeting said "all our ideas feel the same" — that's not a coincidence.
What Is It?
This is a study by Wharton professors Christian Terwiesch and Gideon Nave, along with Mack Institute researcher Lennart Meincke, published in Nature Human Behaviour. The title is blunt: "ChatGPT decreases idea diversity in brainstorming."
The research re-analyzed experimental data from a prior study that had concluded "AI enhances creativity." In that original experiment (by Byung Cheol Lee and Jaeyeon Chung), participants used ChatGPT to complete creative tasks — and the AI-assisted group produced more original, useful ideas at the individual level.
Terwiesch's team spotted a dimension that had been overlooked — not individual quality, but the diversity of ideas across the whole group. Each idea might be solid on its own, but if everyone on the team independently arrives at the same idea — is that really good brainstorming?
How is this different from the Harvard "AI Brain Fry" research?
The Harvard AI Brain Fry study was about individual cognitive fatigue — the idea that using AI leaves people mentally drained. This Wharton study operates on a completely different level. Individuals are fine, but the team's collective diversity of thought collapses. Individual vs. team, fatigue vs. homogenization — these are two entirely different problems.
What Changes?
The results are pretty striking. Participants were given the task: "Use a brick and a fan to build a toy."
In the AI-assisted group, 9 participants worked completely independently — and yet they all came up with the same toy name: "Build-a-Breeze Castle." None of them had talked to each other. The group without AI? Every single idea was completely unique.
Across 5 experiments and 45 statistical comparisons, 37 of them (82%) showed significantly lower idea diversity in the AI-using group. The team used Google's semantic similarity tools to catch ideas that looked different on the surface but were conceptually the same.
| AI group | No-AI group | |
|---|---|---|
| Individual idea quality | High (AI refines ideas) | Average |
| Idea uniqueness | Only 6% unique | 100% unique |
| Concept overlap | 94% overlap | No overlap |
| Expression style | Similar language patterns | Diverse expressions |
| Brainstorming value | Repeated good ideas | Mosaic of varied perspectives |
Why does this happen? Researcher Meincke's explanation is clear:
"If you put the same prompt into the same model, the output comes from the same probability distribution. The more you repeat it, the fewer unique ideas you get — it's inevitable."
— Lennart Meincke, Wharton Mack Institute
Here's the thing — participants were also feeding the AI similar prompts. The convergence isn't just the model's fault; the way people talk to AI is itself uniform.
Getting Started: Using AI Right in Team Brainstorming
The researchers aren't saying to dump AI. The key message is: "Diversity doesn't happen unless you protect it deliberately." Three concrete ways to do that:
- Humans first, AI second
Have team members develop their own ideas first, then bring in AI. When human perspectives have already diverged, using AI from that point reduces convergence. Starting with AI means everyone launches from the exact same starting line. - Vary your prompts intentionally
Even for the same task, have each person ask from a different angle. "Approach this from a cost-reduction perspective." "Focus on user experience." "Suggest something a competitor would never do." Different prompts produce different outputs. As Professor Terwiesch puts it: "The cost of varying prompts is essentially zero, but the diversity payoff is enormous." - Mix your models
Don't rely on ChatGPT alone. Claude, Gemini, Llama, Grok — each model has different training data and a different probability distribution. Meincke's take: "Not trying all five models is just leaving ideas on the table. Mix them all up — go wild."
Chain-of-thought prompting works too
Another technique the research team recommends. Instead of asking AI to "give me an idea," break the thinking into steps. "What are the core constraints of this problem?" → "What are 3 ways to work around them?" → "What happens if you push each approach to an extreme?" Stepping through the reasoning reduces repetition and opens up variation.
Here's the biggest lesson from this research:
"The real value of successful brainstorming comes from the diversity of ideas — not from multiple people repeating the same thoughts."
— Meincke, Nave, Terwiesch (Nature Human Behaviour, 2025)
AI genuinely does turn individuals into superheroes. But if ten superheroes all fly in the same direction, that's not a team — it's a clone army. Deliberately designing structures for divergent thinking — that's the new job of a team leader in the age of AI.
