Why Group Brainstorming Fails, and AI Doesn't Have the Same Problem

In 1957, advertising executive Alex Osborn popularized brainstorming with a specific, testable claim: a group working together could think up roughly twice as many ideas as its members working alone. Groups have been run this way ever since. The claim is also, as far as anyone has been able to measure it, backwards.

Taylor, Berry, and Block tested it directly in 1958: four people brainstorming together versus four people brainstorming separately, with their individual idea lists simply pooled afterward. The pooled individuals produced nearly twice as many distinct ideas as the real group did — the opposite of Osborn's claim. It has been replicated so often since that by 1987, a review found 18 of 22 comparison studies showed the same pattern, and none showed the reverse.

The 1987 study that found out why

Michael Diehl and Wolfgang Stroebe (Universität Tübingen) set out to test the three competing explanations on offer: free riding (people coast, expecting others to carry the group), evaluation apprehension (people censor themselves, afraid of sounding foolish in front of others), and production blocking (only one person can talk at a time, so everyone else just… waits).

Across four experiments, the first two explanations didn't hold up well. Manipulating how people expected their contributions to be assessed (targeting free riding) explained less than 8% of the difference between individual and group output. Evaluation apprehension had some effect, but inconsistently — it didn't reliably interact with the conditions the theory predicted it should.

Experiment 4: isolating the real cause

The decisive test controlled for everything else by simulating group brainstorming with a signal-light system: subjects sat in separate rooms, could only speak when their light was green, and in one condition couldn't even hear each other — they just had to wait their turn in silence, with nothing to actually listen to.

• Individual, no waiting: 106 ideas
• Simulated group, no blocking imposed: 102.67 ideas
• Blocking imposed, but no actual content to hear: 45.67 ideas
• Blocking imposed, with real group communication: 37.67 ideas

"The presence or absence of blocking accounted for most of the variance created by the experimental manipulations in Experiment 4. […] The one interpretation our data supported most strongly was production blocking."

Note that hearing nothing (condition 3) hurt output almost as much as hearing everything (condition 4) — the damage wasn't from being distracted by other people's ideas. It was from being forced to sit silently and wait.

The actual mechanism: short-term memory, not distraction

This is the part most retellings of "groupthink slows brainstorming" get wrong. The paper is specific that it isn't re-evaluating your idea against what someone else said, or being distracted by listening. It's simpler and more mechanical than that:

"[S]torage space in short-term memory is fairly limited, and individuals will only be able to store a small number of ideas at a given time."

While you wait your turn, the idea you had thirty seconds ago has to compete for space with whatever you're thinking about now — and it frequently loses. By the time you can actually speak, it's gone. Multiply this by every person in the room, every time someone else is talking, and you get a room that is measurably worse at generating ideas than the same people working alone would be.

Their own prescription — decades before it became an AI question

The authors' recommendation, in their own words:

"It might be more effective to ask subjects first to develop their ideas in individual sessions and next have these ideas discussed and evaluated in a group session. The task of the group would then consist of evaluation rather than production of ideas."

Generate separately, in parallel. Evaluate together. That's the fix — and it's a fix for a problem that simply doesn't exist when the participants generating ideas are AI models instead of people. An AI model composing a response isn't sitting silently, watching a light, waiting for its turn while an idea decays in short-term memory. Every model can compose its answer during the exact same window as every other model. There's no queue to block.

A 2026 study measured this gap directly

A recent study out of Cambridge and Microsoft Research compared 4,541 ideas from multi-agent LLM teams against 341 ideas from 114 real human teams, across six creative problem-solving tasks, scored blind by trained human judges on novelty and usefulness.

Overall creativity: LLM teams scored roughly twice as high as human teams (Cohen's d = 1.50), driven almost entirely by novelty (d = 1.29) — usefulness scores were statistically indistinguishable between the two.

The paper's own explanation for the gap lines up with the 1987 mechanism almost exactly: human teams' ideas cluster because people have to maintain "smooth conversational flow" to function as a group at all — the very same constraint that causes production blocking. LLM teams showed no such constraint, achieving wider, more efficient exploration of the idea space without needing to preserve a coherent back-and-forth.

What this means in practice

None of this says group discussion is worthless — the 1987 paper's own conclusion is the opposite: discussion is exactly where a group adds value, just not during idea generation. The practical lesson is sequencing: broadcast a question to every model at once for the divergent, generate phase — so answers stay genuinely independent instead of queuing behind each other — then bring the group together to argue, refine, and converge on what actually holds up.

See our Best Practices guide for how Brainstorming Rooms are actually structured around this — quantity and specific, firm pushback over the classic "defer all judgment" rule, with the intended shape running goal → divergence → convergence → decision.

Sources

• Diehl, M., & Stroebe, W. (1987). Productivity Loss in Brainstorming Groups: Toward the Solution of a Riddle. Journal of Personality and Social Psychology, 53(3), 497–509. https://doi.org/10.1037/0022-3514.53.3.497
• Hu, T., Jiang, Y., Li, H., Hernández-Orallo, J., Xie, X., Collier, N., Stillwell, D., & Sun, L. (2026). Multi-agent AI systems outperform human teams in creativity. arXiv:2605.17885.