Why ChatGPT Makes Your Brainstorms Worse (And How to Fix It)

In a 2025 paper in Nature Human Behaviour, researchers ran five experiments asking people to brainstorm new product ideas with and without ChatGPT. The AI-assisted ideas scored higher on individual creativity. That's the headline most coverage stopped at.

The buried result is more interesting. Across the AI-assisted experiments, 94 per cent of ideas shared overlapping concepts. In one task (invent a toy using a brick and a fan), nine different participants, working independently, named their toy "Build-a-Breeze Castle." Without the AI, the same task produced unique ideas every time.

If your team's been using ChatGPT or Claude or Copilot to "brainstorm faster," this is the result you should care about. AI tools don't just help you think faster. They quietly narrow what you think about. And the narrowing is invisible, because each individual idea looks fine on its own.

What's actually going on

Three different research groups have now found the same pattern from different angles.

The Meincke, Nave and Terwiesch (2025) work above is the cleanest version: AI-assisted brainstorming produces higher average per-idea quality and lower diversity across ideas. The trade is real and consistent across tasks.

A team at TU Berlin (Bouschery et al., 2024) found that hybrid configurations behave differently from solo-AI ones. Groups of humans plus AI, working iteratively, generated 172 per cent more creative ideas than groups of humans alone, and 201 per cent more than individuals working alone. The configuration is what matters.

Wharton's Christian Terwiesch put it bluntly in interviews: idea variance comes from using AI to generate ideas, then bringing your own thinking, then collecting other people's thinking, then iterating. If you sit back and let the model do the work, you get the same answer everyone else got.

The reason is straightforward when you say it out loud. Large language models are pattern-matching machines trained on similar data. Ask the same question and they converge on similar answers. Two teams in two different companies asking ChatGPT the same brainstorming prompt today will get two strikingly similar lists. The fluency feels like creativity. It isn't.

Three patterns that work

The hybrid configuration is the one the research keeps validating. Here's how to set it up.

Think first, prompt second. Spend ten minutes on the problem yourself before you go to the model. Stanford's Jeremy Utley calls the alternative "cognitive closure": the moment you accept the AI's first reasonable-looking output, you stop thinking. The team that wins is the team that brings its own raw material into the prompt window, not the team that asks the prompt window to do the raw thinking.

This is the move that most teams skip. ChatGPT is open, the deadline is real, the temptation to ask it for ten ideas right away is huge. Resist the first ten minutes. Write down what you already know, what you're confused about, what assumptions you're carrying. Then prompt.

Iterate, don't accept. Ask for ten ideas. Then ask for ten more that are explicitly different from the first ten. Then ask the model to argue against its own suggestions. Then ask what a sceptical board member would push back on. The model is not a vending machine that gives you the answer. It's a sparring partner that gets better the harder you push it.

Most people use AI as a sentence completer. The teams that get the 172 per cent gain use it as an interlocutor.

Bring the team in afterwards, not instead. This is the one most teams get backwards. They use the AI to generate fifty ideas, then bring those into the team meeting as the starting point. The result: the team anchors to whatever the model said first. The original Wharton finding is that human-only groups outperform AI-only groups when the AI output is treated as the floor.

A better sequence: AI as your private thinking partner before the meeting (with the discipline above). Walk into the meeting with two or three of your own genuinely new ideas, sharpened by the AI conversation but not produced by it. Then run a human ideation session that doesn't reference the AI work. Bring the AI output in only at the end, as a comparison.

The configuration question

Most of the published advice on "using AI for brainstorming" misses the central point: the configuration of human, AI, and group time determines whether you get the upside or the downside.

A useful test: if removing the AI from your process would produce roughly the same output, your team is doing it right. The AI was a sparring partner that made the human thinking sharper. If removing the AI would gut the output, you've outsourced the thinking. That's the configuration the research finds is worst.

Three diagnostic questions to ask after your next AI-assisted session:

The first is whether your team's ideas, looked at side by side, are noticeably different from each other. If everyone's ideas have the same shape and use similar phrases, the AI converged the room. Five different people should produce five recognisably different sets.

The second is whether anyone pushed back on the model in writing. Not "make it more creative" (that's lazy) but specific, constraining pushback: "these all assume the customer is enterprise. Give me ten that assume early-stage startup." If nobody steered, the model drove.

The third is whether you'd be embarrassed if another team in your industry produced your shortlist. Convergence at the population level is the real risk. If a competitor running the same workflow would reach the same answers, your AI use is producing parity, not advantage.

A practical pattern for next week

For your next ideation session, try this configuration.

Two days before the meeting: send the question. Ask participants to spend ten minutes thinking by themselves, then to use AI for thirty minutes as a sparring partner, with the explicit instruction to push back on the model at least three times before settling. Ask them to bring three ideas they believe are genuinely theirs, not the model's.

In the meeting: run a normal silent-ideation-first session. Five minutes solo, then round-robin sharing. Cluster, vote, decide. Don't reference the pre-meeting AI work yet.

In the last fifteen minutes: invite people to add anything from their AI sessions that the human discussion didn't surface. The order matters. If you start with the AI material, you've anchored the room. If you end with it, it acts as a useful check on what the team produced.

You'll know it's working when the team disagrees about which ideas to take forward. Disagreement is a sign of diverse input. Easy consensus, especially on AI-assisted ideas, is usually a sign that everyone has been quietly drinking from the same pipe.

The point

AI is a multiplier. It multiplies whatever process you bring to it. Bring a process that uses it as a sparring partner and you'll get sharper team output than you've ever had. Bring a process that uses it as a replacement and you'll get a faster, fluent, mediocre version of what every other team in your industry is producing right now.

The question your team needs to answer isn't "should we use AI in our brainstorms." It's the more uncomfortable one: how exactly are we using it, and would the output be any different if we weren't?