There is a documented phenomenon in AI-assisted brainstorming that doesn't show up in any single session. It shows up at scale: when many people independently brainstorm with the same AI model using similar prompts, the ideas they produce converge toward the same cluster. A 2024 study found that people who brainstormed with AI scored 39.94 on average on creativity tests compared to 11.35 for those who brainstormed without it — a dramatic improvement at the individual level. But the individuals within the AI-assisted group produced ideas that were more similar to each other than the individuals in the non-AI group. Better individual performance. Lower collective diversity. The population of ideas narrowed even as each individual's output improved.
This is the homogenization problem. It isn't a bug in any particular AI — it's a structural consequence of how large language models work. They generate responses that are statistically likely given the prompt and the training data. Many different people asking similar questions receive responses that draw from the same statistical center. The ideas that look creative to any individual may be systematically similar across the population, because they all look creative relative to the same baseline.
This module examines the tension between two legitimate goods: using AI to dramatically improve individual brainstorming output, and maintaining the creative diversity that comes from genuinely different people thinking independently. This is not a reason to avoid AI in brainstorming. It's a reason to know what you're gaining and what you might be trading away — and to build practices that preserve diversity even as you use AI for individual lift.
A product innovation team at a large consumer technology company recently moved to AI-assisted brainstorming. Each product manager now prepares for quarterly strategy sessions by spending 30 minutes with an AI assistant generating ideas before the team meets. Attendance at the brainstorming sessions is up. People come prepared. The discussions are sharper and more productive. Leadership is pleased.
Six months in, a researcher on the team notices something unexpected: the ideas coming out of each product manager's pre-session AI conversations are more similar to each other than they used to be. The "wild card" ideas that used to appear occasionally — the ideas no one else had thought of — have nearly disappeared. The overall idea quality is up. The idea variance is down.
The team is facing a choice that doesn't have an obvious right answer. Position A: the homogenization effect is a real systemic risk — it's reducing the creative diversity that produces breakthrough ideas, and the organization should limit or restructure AI-assisted brainstorming to preserve variance. Position B: individual brainstorming quality has improved dramatically, and the right response is to accept that some diversity is traded away in exchange for substantially better average ideas — the loss is acceptable and perhaps not real (better average ideas may be more valuable than higher variance).
Both positions have serious arguments behind them. That's what makes this a debate worth having.
The scenario illustrates a problem that is invisible in any individual session and only visible at scale. No product manager opened an AI chat and received bad ideas. Each individual session looked like a success. The problem assembled itself gradually, across many sessions, across many people — until the researcher compared outputs and noticed that the population of ideas had quietly narrowed.
This is the structure of the homogenization problem: local improvement, collective convergence. Understanding that structure is the prerequisite for deciding what, if anything, to do about it.
The homogenization problem is not a reason to avoid AI in brainstorming. It's a reason to understand what AI-assisted brainstorming optimizes for (individual quality) and what it doesn't optimize for (collective diversity) — and to make intentional choices about both.
The 2024 research finding is specific and worth holding precisely. People who brainstormed with AI assistance scored 39.94 on creativity tests on average. People who brainstormed without AI scored 11.35 on average. That is a large, real improvement at the individual level — not a rounding error, not an artifact of measurement. AI assistance substantially lifted individual brainstorming quality.
But within the AI-assisted group, the ideas were more similar to each other than the ideas within the non-AI group. The individuals who brainstormed without AI produced lower-quality individual outputs — but those outputs were more varied. The population of ideas in the AI-assisted condition was higher quality and lower variance.
This is a structural consequence of how the model works, not a fixable bug. AI generates responses by predicting what is statistically likely given the prompt and the training data. Many different people asking similar questions receive responses that draw from the same statistical center of the training distribution. The result is that ideas that look creative to any individual — because they exceed what that individual would have generated alone — may be systematically similar across the population, because they all look creative relative to the same baseline. Making the AI "more creative" shifts the center, but the center is still a center. Many similar prompts will still produce similar outputs.
Creative diversity produces its primary value not in any individual idea but in the combination and collision of different ideas across people who think differently. If ten people independently arrive at similar ideas, the meeting where they share those ideas produces less combinatorial novelty than if ten people arrived at genuinely different ideas. The value of diversity is not in any single idea — it's in the interactions between ideas that no single person would have generated.
The loss from homogenization isn't visible in any individual's output. It's visible in what the group discussion fails to generate. The "wild card" ideas that a diverse population occasionally produces — the ideas that disrupt categories, that no one had thought of, that force the group to reconsider its assumptions — become rarer when the population's idea-generation process draws from the same statistical well.
There are concrete practices that preserve collective creative diversity while using AI for individual lift. They require intentionality — they don't happen automatically.
Stagger AI use. Have some team members brainstorm without AI assistance before any AI-assisted session runs. Preserve their outputs separately so they aren't absorbed into the AI-assisted pool before the meeting.
Vary prompts deliberately. Assign different team members meaningfully different problem framings before they use AI. Reducing the correlation between prompts reduces the correlation between outputs.
Preserve independent outputs. Create time and space for ideas generated completely independently of AI — not as a criticism of AI use, but as a distinct input stream. Treat AI-assisted and non-AI-assisted outputs as two different data sources, not one merged pool.
Use two streams explicitly. In the brainstorming session itself, distinguish between AI-assisted ideas and independently generated ideas. The collision between the two streams is where the combinatorial value lies.
This is a debate module, which means your job is not to arrive with a pre-formed conclusion — it's to be able to argue either position with rigor. That requires understanding what makes each position genuinely strong, not just what makes it superficially plausible.
The systemic risk is real and measurable. Creative diversity is a form of intellectual biodiversity — it's what produces the ideas that disrupt categories and solve intractable problems. If AI systematically reduces the variance of ideas across a population of knowledge workers, the loss may not be visible in any quarterly review but will compound across years. The organizations that preserve genuine creative diversity while their competitors homogenize may have a significant long-term advantage. The cost of prevention is low; the cost of ignoring it could be high.
The measured improvement in individual brainstorming quality is substantial and real. A team of people producing substantially better individual ideas who then share and combine those ideas is likely producing better total output than a team of people producing diverse but lower-quality individual ideas. The concern about collective diversity may be theoretical — there's little evidence that the kinds of ideas lost to homogenization (the statistical edges) were the ideas that actually drove breakthrough products. Average quality matters more than variance in most real product decisions.
Both positions are defensible because they're evaluating different things. Position A is concerned with what brainstorming is for at scale: producing genuinely novel directions. Position B is concerned with what brainstorming produces at the individual level: ideas good enough to act on. These are both real goods, and a complete answer has to reckon with both.
You'll apply this in the lab — you'll be assigned a position and required to defend it rigorously across multiple rounds. The AI will argue the other side. At the end, you'll assess which position holds up better under pressure, and under what conditions each is strongest.