Do Not Let AI Give Everyone the Same Idea

The brainstorm starts unusually well. Five people arrive with polished options, tidy rationales, and almost no awkward silence.

Ten minutes later, someone notices the problem: the options are different versions of the same idea.

Several participants used AI to prepare. Each asked a sensible question, received a sensible answer, and brought the answer into the room. The meeting gained speed, but it lost range before the conversation even began.

This is not an argument for banning AI from creative work. It is a reason to separate two jobs that brainstorms often blur together: expanding the space of possible ideas and selecting the ideas worth pursuing. AI can help with both. When everyone asks it for the best answer too early, however, selection quietly begins before the team has done enough exploration.

The fix is a short divergence round: a deliberate period in which the meeting rewards distance between ideas before it rewards agreement.

The research is less simple than "AI makes ideas generic"

Evidence about AI and creativity now points in more than one direction.

An April 2026 preregistered experiment on AI-assisted creative writing found that incentives changed how people used the tool. Participants rewarded for originality relative to their peers produced a more diverse collection of writing than participants rewarded for quality alone. The originality group did not abandon AI. They accepted fewer suggestions verbatim and used it more selectively for brainstorming, proofreading, and targeted edits.

That result matters because it moves the problem out of the model alone and into the surrounding workflow. If a meeting rewards speed, polish, and immediate plausibility, people have good reasons to converge. If it explicitly rewards an option that changes the frame, they use the same technology differently.

There is also evidence that AI systems can generate substantial novelty when the discussion is designed for exploration. A May 2026 study compared 4,541 ideas from multi-agent AI teams with 341 ideas from human teams across six problem-solving tasks. The multi-agent teams scored higher on the study's creativity measures, driven by novelty while maintaining comparable usefulness. For both people and AI, more creative results were associated with conversations that ranged widely rather than remaining centered on one theme.

Those are controlled tasks, not proof that a group of agents should replace a product workshop. The practical signal is narrower: range can be designed. Model choice and discussion structure affected how the AI teams explored, just as incentives affected how people incorporated AI suggestions.

A recent Anthropic study offers a useful caution. In multi-agent experiments, identical agents sometimes made strikingly similar choices, including repeated project ideas and even the same story title. The researchers argue that low variance can turn one weak choice into a system-wide pattern. These were artificial agent environments, not workplace brainstorms. Still, the example captures a recognizable risk: multiplying outputs does not guarantee multiplying perspectives.

Run a 12-minute divergence round

Put this round near the beginning of an ideation meeting, after the problem is clear but before anyone presents a preferred solution. Twelve minutes is enough to change the shape of the conversation without turning creativity into ceremony.

Minutes 1-2: Start from private human observations

Ask each person to write two things without seeing anyone else's answers:

1. A constraint the obvious solution ignores.
2. An option that feels slightly uncomfortable, impractical, or unfashionable.

This is not a purity test about unaided thought. It creates independent starting points before the room's first polished answer becomes an anchor. A support lead may notice a training burden. An engineer may question the assumed timeline. A new employee may challenge a term everyone else has stopped examining.

Do not ask people to write "their best idea." Best invites premature ranking. Ask for a neglected fact and a strange direction.

Minutes 3-5: Use AI to stretch the frame

Now bring in AI, but change the assignment. Do not ask, "What are good solutions to this problem?" That produces a menu that looks finished.

Ask for pressure from several distinct lenses:

We are exploring how to reduce customer onboarding time. Do not recommend a final solution. Give us one hidden constraint, one counterexample, and one unconventional move from each of these perspectives: a first-time customer, an accessibility specialist, a support manager, and a security reviewer. Keep each contribution to one sentence.

The roles should reflect real differences in knowledge, incentives, or exposure to consequences. Inventing four colorful personalities is not the same as adding four perspectives. "Bold innovator" and "visionary disruptor" may sound different while rewarding the same behavior.

Treat AI suggestions as provocations, not participants with authority. The purpose is to disturb the current frame, especially when the relevant human expert is not in the room.

Minutes 6-9: Cross ideas instead of voting

Each person chooses one human observation and one AI-generated provocation that do not naturally belong together. Their task is to create a new option from the collision.

A support constraint paired with a security counterexample might lead to a reversible, limited pilot. An accessibility perspective combined with an odd pricing idea might produce an onboarding service rather than another interface feature.

For these four minutes, ban duplicate proposals and defer feasibility objections that can be tested later. Questions are welcome when they open the option: "What would make this possible?" or "Which assumption does this remove?" A statement such as "Legal will never allow it" closes the path without showing the actual boundary.

Capture each option in one or two sentences. A long explanation gives familiar ideas an advantage because they are easier to defend.

Minutes 10-12: Preserve difference before selecting

Group the options by the assumption they change, not by superficial similarity. You might find one family that changes who does the work, another that changes when the work happens, and a third that removes the need for the work entirely.

Then record three things:

• the most immediately useful option;
• the option that most changes the team's frame;
• the unresolved question that would alter the choice.

Only after those are visible should the meeting move into evaluation. The unusual option may not win. Its value may be exposing a constraint, creating a fallback, or showing that the team's preferred idea rests on a fragile assumption.

Watch for polished convergence

An AI-heavy brainstorm can look healthy by ordinary meeting signals. Participation is high. Notes are detailed. Nobody is stuck. The risk appears in the relationships between ideas.

Look for repeated verbs, identical customer assumptions, and options that differ only in channel or packaging. Notice when every proposal preserves the current business model, workflow, or power structure. Ask which idea could have come only from someone in this room. If the answer is "none," the team may be assembling rather than thinking.

The same warning applies after the meeting. A summary that compresses ten proposals into three themes can erase the outlier that matters later. Keep the raw option and the assumption it challenged alongside the cleaner recap. Compression is useful for follow-through, but it should not rewrite exploration as consensus.

The point is better choices, not permanent divergence

Teams eventually have to choose. Endless novelty is another form of avoidance, and a surprising idea is not automatically a good one. The divergence round earns its keep only when it improves the options that enter evaluation.

Afterward, switch the criteria openly. Test value, evidence, cost, reversibility, and risk. Bring the relevant expert into the decision. Let AI retrieve context, compare options, or surface a contradiction, but keep the team's reasoning visible.

Caspi is built for real-time meeting support through live recap, suggested questions, contextual chat, proactive flags from connected tools, post-call action items, and persistent meeting memory. Used during a divergence round, those capabilities can help a team recover missing context and ask the question that widens the room. The goal is not more ideas on the page. It is a meeting that explores broadly enough to make a real choice.