Your AI brainstorming tool and your competitor’s AI brainstorming tool are producing the same ideas
Here is the contradiction sitting at the centre of every content team’s AI adoption right now: the tool that makes your average output better is making your output look more like everyone else’s. That is not a theoretical concern. It is what four separate studies found when researchers tested generative AI across short-story writing, circular-economy solution generation, humour writing, and collaborative storytelling.
The research, published in MIT Sloan Management Review and led by Léonard Boussioux at the University of Washington Foster School, alongside Anil Doshi, Oliver Hauser, and Kartik Hosanagar, documents the mechanism precisely. Participants who used generative AI produced ideas of higher average quality. The typical output improved. But the full set of ideas produced across all participants shrank in diversity. The outliers — the genuinely strange, the counterintuitive, the ideas that become category-defining campaigns — largely disappeared.
The cause is anchoring. When people see an AI-generated suggestion early in the ideation process, they cluster around it. Different people, using the same tools, prompted in broadly similar ways, land on broadly similar ideas. The middle of the distribution rises. The edges collapse.
How the industry walked straight into this
The adoption pattern followed a familiar logic. AI reduced the cost and time of producing content. Teams under pressure to publish more, faster, adopted the tools that removed friction. Brainstorming with an AI assistant became standard practice because it works — for an individual, in isolation, measured against their own previous output. The research confirms this: AI meaningfully helps less-experienced and less-creative people generate more novel, useful ideas. That is a real benefit and it should not be dismissed.
But adoption decisions were made at the individual or team level, not at the market level. Nobody modelled what happens when every content team at every competing brand runs the same prompts through the same models during the same quarterly planning cycle. The answer, which the research now makes explicit, is convergence. Fewer breakthrough outliers emerge across the collective pool of ideas, even as individual outputs improve.
This is not a failure of the tools. It is a failure of how the tools were positioned and adopted. The productivity story was true and legible. The homogenisation cost was real but invisible until researchers controlled for it across hundreds of participants.
The strongest case for the other side
The opposing view deserves a fair hearing. If AI raises the floor of creative quality across an industry, perhaps the homogenisation is tolerable. A sector where most content was mediocre, now producing competent work uniformly, may serve audiences better than a sector with a few brilliant outliers surrounded by noise. There is also an execution argument: differentiation comes from distribution, targeting, and production quality, not just from the raw idea. Two brands could start with the same concept and produce wildly different results.
Both points have merit. Neither resolves the core problem for brands that compete on thought leadership, creative reputation, or audience trust built over time. In those categories, the outlier idea — the campaign, the report, the content series that nobody else would have made — is precisely what builds durable audience relationships. A rising floor helps nobody when the ceiling is what determines market position.
The research also offers a specific and practical rebuttal to the optimistic case. Studies found that keeping humans in charge of early-stage ideation preserved idea diversity at levels close to fully human creative work. The diversity loss is not intrinsic to using AI at all. It is specifically tied to using AI during the generative, divergent phase of the process. That is a distinction with direct operational implications.
What this means for marketing and content teams
The practical adjustment is a sequencing change, not a tool ban. Run your divergent ideation phase without AI assistance. Put people in a room, or a document, or a structured async process, and generate raw directions before any model sees the brief. Protect that phase deliberately, because it is the phase where your ideas are most likely to diverge from your competitors’ ideas, and divergence is what differentiation actually requires.
Bring AI in during convergent phases: drafting, refinement, variant generation, optimisation against a defined direction. That is where the quality uplift is most useful and where anchoring to similar outputs causes the least strategic damage.
Audit your current process against this sequencing. If your team is opening a chat window before they have written down their own instincts about a brief, you have already handed the divergent phase to a model trained on the same corpus as every other team doing the same thing. The output will be competent. It will not be distinctive.
One further implication: the brands that treat early-stage ideation as a protected, human-led capability will, over the next two to three years, accumulate a library of genuinely differentiated creative directions. The brands that do not will find themselves publishing work that is indistinguishable from their competitors’ work at the same moment those competitors are wondering why their content performance is plateauing. The research explains why that plateau arrives. The sequencing adjustment is how you avoid it.
