Five AI Ideas May Still Be One Idea

A team has twenty minutes to improve a service. Someone asks an AI assistant for five ideas. The answer looks generous: a dashboard, a chatbot, personalised alerts, predictive routing and an automated summary. Five bullet points have appeared, but all five assume the same basic move—collect more data, process it centrally and place another interface between the person and the service.

The list is varied in wording and narrow in structure. That distinction matters because brainstorming is supposed to expose different ways the world could work, not merely produce more sentences. Generative AI is useful for widening an option space, but only if people inspect whether the options differ in mechanism, resources, authority and failure. Asking the model to choose the “best” idea too early collapses exploration into recommendation.

Fluency can disguise convergence

A language model generates text by extending a prompt through patterns learned from large collections of prior text. When many people ask a similar, underspecified question, the model has strong reasons to return recognisable solution patterns. Different nouns and examples can make the outputs feel independent even when their causal logic is identical.

Suppose a school wants to help pupils who miss homework deadlines. “Send reminders”, “build a progress dashboard” and “use personalised nudges” differ as features. Structurally, each detects lateness and pushes information towards the pupil. A tutoring hour, a quieter place to work, fewer simultaneous deadlines and a change in how work is assigned belong to different families: human support, material access, coordination and institutional redesign. The first list may be useful, but it has not mapped the problem.

Experimental evidence gives this concern a sharper shape. In a 2024 Science Advances study, people writing short stories with access to generative-AI ideas produced work evaluated as more creative, better written and more enjoyable, with larger gains among less creative writers. Yet AI-assisted stories were also more similar to one another. The tool improved an individual output while narrowing variation across the group.

That result does not establish a universal law. A large experiment later presented at ACM Creativity & Cognition found that high exposure to AI-generated examples increased collective idea diversity in its alternate-uses task, without improving the creativity of individual ideas. A 2026 systematic review and meta-analysis of 19 studies reported a small overall homogenisation effect, stronger in semantically constrained ideation tasks than in minimally constrained divergent-thinking tasks. The sensible conclusion is not that AI always homogenises thought. It is that diversity is an outcome to measure, not a benefit to assume.

Count mechanisms, not bullet points

A practical option audit can classify each suggestion along five dimensions.

• Mechanism: What changes first, and through which causal path?
• Resource: Does the option mainly use software, money, time, space, equipment, rules or human attention?
• Authority: Who must decide, permit, maintain or reverse it?
• Beneficiary and burden: Who gains capability, and who receives new work or risk?
• Failure mode: How does the option fail differently from the others?

Two suggestions belong to the same family when their answers are substantially the same, even if their surface forms differ. A chatbot and an automated email may both be self-service information channels. A trained navigator shares information too, but adds judgement, conversation and responsibility. A rule simplification changes the need for information in the first place.

This classification is deliberately human-readable. Semantic-similarity scores and embedding clusters can help with very large collections, but they can also reproduce the model’s own sense of similarity. For a small decision, a table and a conversation are often more revealing. The point is not to calculate a universal diversity number. It is to notice which causal families are absent before a familiar family fills the page.

Alkemata’s earlier article Do Not Make Human–AI Agreement the Goal argued for preserving an independent human view before reconciling conclusions. Option diversity addresses a different stage: before agreement or disagreement, did the workflow create genuinely different choices to judge?

Do not confuse diversity with quality

The strongest objection is straightforward. A team does not need maximum variety; it needs something that works. In a mature domain, several independent analyses may converge because evidence, standards or physical limits favour the same design. Forcing novelty can waste time, produce theatrical “alternatives” or give an unsafe option equal status merely because it is different.

Convergence can therefore be information. If a problem has strict constraints, the useful question may be why the options converge and whether any constraint has been assumed rather than verified. A diversity audit should not reward distance for its own sake. Every additional family still needs a plausible mechanism, an authorised owner and a way to test it. Ideas that violate rights, budgets, safety requirements or established evidence should not survive merely to keep the board colourful.

Nor does a list of perspectives guarantee social diversity. Asking one model to “act as” several professions or demographic groups can create stereotyped voices rather than situated knowledge. A 2025 large-scale comparison of humans and language models found mixed to negative results from several prompt-engineering strategies intended to raise model creativity, while humans showed greater variability and stronger performance at the most creative end of the distribution. Real people affected by a decision remain sources of evidence and values, not personas for a model to simulate.

Run a two-round option-space experiment

Choose one low-stakes, reversible problem: organising a shared room, reducing a recurring administrative delay or improving a team handover. Do not use a live medical, legal, employment or safety decision.

1. Write the boundary. State the desired outcome, hard constraints, affected people and what the AI may not decide.
2. Create a human seed. Without the model, write three possible mechanism families—not finished solutions. “Change information”, “change the physical arrangement” and “change the rule” are enough.
3. Generate round one. Ask the AI for ten feasible options. Do not ask it to rank them.
4. Build the diversity table. For each option, record mechanism, resource, authority, burden and failure mode. Merge superficial variants.
5. Generate round two. Give the model the table and ask only for options from missing families or for a reason why a missing family is infeasible. Require assumptions to be labelled.
6. Choose with evidence. Only now apply criteria such as effect, cost, reversibility, access and responsibility. Keep “combine two families” as a possible result.

Compare the rounds by counting structurally distinct families, not total ideas. Record whether round two adds a new causal route, exposes an untested constraint or merely produces stranger wording. Success is not ten unprecedented inventions. It is one worthwhile option or missing question that the first fluent list concealed. Failure is useful too: if every credible option converges after assumptions are checked, the team has learned that the space may be narrower than it looked.

The final decision remains human and accountable. AI can cheaply populate the board and challenge its empty regions. People still decide which differences matter, which burdens are acceptable and whose experience is missing. The next time five polished suggestions arrive, the useful question is not “Which one wins?” It is “How many different ways of changing the situation are actually here?”