The AI Brainstorming Paradox: More Fluency, Less Range
This series examines a specific behavior of AI-assisted brainstorming: a language model can produce polished, plausible ideas quickly while the overall pool of ideas becomes narrower than it appears. The central question is not whether an AI can invent a novel-sounding answer. It is whether repeated use of the same model, prompt, or discussion structure explores enough different directions before people begin selecting one. That question matters to anyone using AI for product concepts, research proposals, strategy, or other open-ended work where the best option may be far from the first plausible one.
The practical judgment this series supports is simple to state but harder to apply: before asking an AI system to refine an answer, determine whether it has explored the problem broadly enough. Doing so requires separating three things that are often treated as one. A session can become anchored to its own early ideas. Independent sessions can draw from the same narrow region of the model’s knowledge. And agents that are supposed to challenge one another can converge because they share messages, roles, or authority. These are related forms of narrowing, but they call for different remedies.
When a good-looking idea pool is still too small
AI changes the economics of brainstorming. It can turn a vague request into a coherent list in seconds, connect concepts that an individual user might not have considered, and revise weak formulations without much effort. Those advantages are real. They also make a failure harder to notice: fluency can give a small idea pool the appearance of breadth.
The distinction is between the quality of an individual idea and the information contained in a collection of ideas. One answer may be original, useful, and well written. That tells us something about that answer. It does not tell us whether the system considered other assumptions, customer needs, mechanisms, or solution paths. If ten outputs are variations on the same underlying approach, the user may receive ten pieces of text but only one meaningful direction.
This matters because brainstorming is upstream of judgment. Screening can discard weak candidates, and development can improve a promising one. Neither process can select an alternative that was never generated. A narrow starting pool therefore creates a risk that is different from ordinary low quality: the organization may become confident in a local optimum simply because no competing path is present. In innovation, where uncertainty is unavoidable and distant solutions may matter most, the missing alternatives are part of the cost.
The risk becomes collective when many people use similar AI defaults. Each user may save time and receive an answer that feels better than what they could have produced alone. Across users, however, the outputs can cluster around the same familiar themes. This is a kind of tragedy of the commons in the marketplace of ideas: an individually efficient tool can reduce the variety available to the community that relies on it.
The warning is not that every AI-generated idea is derivative, or that human brainstorming is automatically diverse. Human groups also fixate, defer to authority, and repeat familiar approaches. The more precise claim is that AI-assisted brainstorming can hide where exploration was lost. A polished final list may reflect genuine discovery, local elaboration around a shared theme, or premature agreement. Those outcomes look similar if we inspect only the wording of the final answer.
The diagram captures the central problem. Narrowing can happen inside a session, when an early output shapes what follows. It can happen across sessions, when different runs repeatedly sample from a shared center. It can also happen after agents begin interacting, when their histories and judgments become coupled. By the time a collective pool reaches a human decision-maker, those losses may be difficult to distinguish from successful synthesis.
Two scales at which exploration disappears
The first scale is the individual session. Suppose an AI is asked for several fitness products, research ideas, or strategic options one after another. Once it proposes a familiar pattern, later responses are generated in the presence of that pattern. The model can continue elaborating the same structure: a tracking feature becomes another tracking feature, a research method becomes a slightly different version of the same method, or a strategic premise quietly becomes fixed. This is fixation in behavioral terms. It does not require claiming that a model has a human mind. It means that earlier material constrains the subsequent search.
Humans experience a related problem. Once a representation or solution path is activated, associated alternatives can become harder to retrieve. Examples shown at the beginning of a creative task can pull later work toward their features, even when people are told to be different. An individual’s knowledge can therefore be both a resource and a constraint: it makes some connections easy to reach while leaving other regions unexplored.
The second scale is the population of sessions or agents. Human contributors begin with limited, uneven, and personally shaped knowledge. A veterinarian, an artist, and a traveler may approach the same question from different starting points because their experiences make different associations accessible. No one person contains the whole knowledge space, but a group can cover more of it because knowledge is partitioned across people.
A language model is built differently. Training aggregates patterns from many people into a unified representation, post-training aggregates preference signals, and inference samples from that shared system. The result can be remarkably broad relative to one individual, yet similar across independent runs. The model has access to many perspectives without necessarily preserving the separation among the perspectives that would cause different contributors to start and continue along different paths.
This gives us a useful distinction. Fixation concerns how ideas accumulate within one session. Knowledge aggregation concerns how much independent sessions differ from one another. The two can coexist. A model might be no more fixated within a ten-idea session than a human participant while still producing a much narrower collection of first ideas across many sessions. Conversely, a technique might make each session more exploratory while leaving all sessions pointed at the same starting region.
Interaction adds a third layer to the same distinction. When agents exchange messages, each proposal is conditioned not only on the original problem but also on the collaborative history. A discussion can broaden the local treatment of an idea while synchronizing the group around one direction. This is why a transcript that becomes longer, more detailed, and more internally coherent may still represent less independent exploration than several separate first passes.
The practical consequence is that “more ideas” is an incomplete measure. We need to ask where the ideas begin, how they develop, and how much they remain distinct after agents see one another’s work. Those questions will organize the evidence that follows in this series. They also explain why a single intervention rarely solves the whole problem: changing the starting cues, changing the within-session reasoning process, and changing the communication structure affect different points of the search.
What the evidence can—and cannot—tell us
The evidence examined here comes from two complementary controlled settings. One study compares human and GPT-4o sessions generating fitness-product ideas, using categories based on industry context, psychological need, and product form rather than relying only on wording similarity. It examines both diversity within a participant’s sequence and diversity among participants’ first ideas. It then tests human seeding, different persona designs, and a staged generate–review–revise procedure.
The other study evaluates multi-agent systems generating structured scientific research proposals across twenty topics. It varies model capability and alignment, agent roles and authority, group size, communication topology, and model heterogeneity. Its measures include effective semantic modes, global concentration around a centroid, pairwise semantic spread, and lexical uniqueness, with limited expert validation. Together, the studies make it possible to connect a cognitive account of narrowing with a systems account of interaction-driven convergence.
Their measurements should be read as instruments, not as direct readings of creativity. A semantic diversity score is not the same as usefulness, feasibility, truth, or safety. The fitness study’s strongest results concern diversity and leave some quality trade-offs for further work. The multi-agent study finds that quality differences can be modest while diversity differences are large, but part of its quality assessment uses an automated judge. Neither study establishes that every distant idea is valuable or that every consensus is harmful.
The task also determines what diversity should mean. In a problem with a strong ground truth, forcing additional variation can produce noise or hallucination. In open-ended research or product ideation, suppressing alternatives too early can hide worthwhile possibilities. The goal is not permanent disagreement. It is a broad enough search followed by an informed convergence, with evidence that the narrowing occurred for a good reason rather than because the system’s defaults made other paths inaccessible.
That distinction gives the series its scope. It will not treat AI as a uniformly creative or uniformly homogenizing force. Instead, it will follow the points where range can be lost: early anchors inside a session, shared starting regions across sessions, and structural coupling among interacting agents. It will then examine which interventions target which point, what trade-offs they introduce, and why topology or prompt changes that work in one setting may fail in another.
The opening conclusion is therefore deliberately modest. AI brainstorming converges too early when shared priors and coupled histories contract independent exploration before enough alternatives have been generated. To understand and correct that failure, we first need to see how fixation, shared knowledge, and interaction differ in sequence and scale. Once those distinctions are clear, “the AI keeps giving us the same kind of idea” becomes a testable systems question rather than a vague complaint.