The polished idea trap

A model can return twenty plausible concepts before a human has finished naming the problem. That speed changes the visible bottleneck. When candidate production was costly, a long list looked like creativity. Once candidate production becomes cheap, the scarce work moves elsewhere: deciding which constraint matters, distinguishing mechanisms hidden beneath similar language, and accepting responsibility for an irreversible choice.

The familiar definition of creativity joins two requirements: originality and effectiveness. Runco and Jaeger traced the standard formulation to novelty plus usefulness, while Amabile’s consensual assessment tradition operationalised creativity through informed judgement rather than a context-free meter. Those ideas do not settle every dispute, but they block a common mistake. A surprising output is not yet a creative contribution. It must alter a possibility space and survive the demands of a situation.

The bottleneck moves from making to judging

Suppose a team can manually develop four concepts in a day. A model raises the number to forty. If every concept receives equal attention, the team has not gained ten times more creative capacity. It has created a queue. Review becomes shallower, near-duplicates look independent, and the first polished option can become the reference point against which all later options are judged.

The practical quantity is not raw count. It is the number of causally distinct candidates that receive enough examination to reveal their promise and failure. Ten surface variations of a battery-powered refrigerator are one mechanism family. A phase-change sleeve, an evaporative liner and a reversible adsorption bed are three families because an intervention on humidity, ambient temperature or material state changes them differently. The relevant unit is a tested mechanism family, not a paragraph that sounds different.

The order of contribution is causal

Human–AI interfaces often begin by showing a complete draft. That choice is not neutral. The draft supplies vocabulary, structure and a local optimum before the person has externalised a frame. Editing then asks the human to react within a model-created coordinate system. Co-creation can instead expose partial materials, questions, mechanism families or counterexamples, leaving the person to establish direction.

McGuire, De Cremer and Van de Cruys tested this distinction in poetry tasks. In one study, participants who received an AI poem to edit were less creative than people writing alone. In a second interface designed for co-creation, that deficit dissipated, with creative self-efficacy implicated as a mechanism. The point is not that poetry generalises to engineering. It is that the order of contribution changes what the human is asked to do: originate, choose, repair or merely approve.

Evidence says “design the combination”, not “add AI”

A systematic review by Vaccaro, Almaatouq and Malone examined 106 experiments reporting 370 effect sizes with human-only, AI-only and combined conditions. Human–AI systems improved on humans alone on average, yet underperformed the better solo performer on the stricter synergy baseline. Task type mattered. Decision tasks showed a negative pooled synergy effect, while creation tasks showed a positive but statistically uncertain estimate. The experiments were heterogeneous, and creation studies were comparatively few.

This result does not prove that creativity is uniquely suited to AI. It does something more useful. It rejects the assumption that combination automatically yields complementarity. Division of cognitive labour must be designed and measured against both solo baselines. A system that makes an average participant better may still be worse than simply deploying the model, or worse than letting a skilled person work without an intrusive default.

flowchart TD
  A["Two teams, one impossible cabinet"] --> B["The polished idea trap"]
  B --> C["Analogy crosses the constraint"]
  C --> D["Fluency changes the population"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#5B5CF6,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

Analogy crosses the constraint

Recombination is often described as taking pieces from existing ideas and joining them in a new arrangement. That description is too loose to guide a design. A collage may be novel yet irrelevant. A useful analogy does more than import a visible feature. It transfers a pattern of relations from a source situation to a target situation, then predicts what should happen when the target is changed.

Gentner’s structure-mapping account makes the crucial move. The source and target need not share the same objects. They must support corresponding relations. In the clinic cabinet, wax, salt hydrate and ice are different materials. What matters is a relation: while a material changes phase, added energy is absorbed with limited temperature rise. That relation can be mapped to thermal protection. The cabinet does not need to resemble a freezer.

This is why analogy can carry creativity across a constraint. It retrieves a mechanism from outside the obvious solution vocabulary. It also creates a liability. A remote source can seduce the team through imagery while the causal relation fails. Biomimetic language, for example, can hide a decorative resemblance where no useful intervention transfers. The map must therefore be allowed to lose.

A new distinction: semantic distance versus constraint distance

Creative systems usually treat remoteness as semantic distance. A hospital cabinet is close to a refrigerator and far from a fire door. That can be useful for retrieval, but it confuses two geometries. Semantic distance asks how unlike the source sounds. Constraint distance asks how much the source changes the relation by which the target’s hard constraint is met.

A spacecraft refrigerator is semantically remote from a rural clinic, yet it may remain constraint-near if it still assumes reliable electrical control. A cheap chemical indicator is semantically closer to packaging than refrigeration, yet constraint-far because it replaces powered logging with an irreversible material transition. The second transfer changes the design’s dependence structure. It removes an entire class of failure rather than decorating the same architecture.

A transfer test, not a resemblance score

There is a temptation to score an analogy from one to five and move on. A single resemblance number conceals which relation is preserved. A better transfer record names five things: the source structure, the target structure, the relation being preserved, the intervention that should commute across the map and the point at which the map fails.

“A team is like an orchestra” is weak until the relation is specified. If the preserved relation is differentiated roles coordinated through a shared temporal signal, removing the conductor may have different effects depending on whether musicians have a score, rehearsal and local cues. Those break conditions generate organisational questions. If the analogy is only “many people make one thing”, it generates none.

Gick and Holyoak’s work on analogical transfer also warns that possession of a useful source does not ensure spontaneous application. People often require a hint or comparison before they extract the common schema. Models can help by retrieving distant sources and making candidate mappings explicit. The human contribution is to decide whether the target really supports those relations and whether the source hides a condition absent from the target.

Fluency changes the population

AI assistance can improve the quality of one person’s output while making many people’s outputs more alike. These claims sound contradictory only if the individual artefact and the population of artefacts are treated as the same unit. They are not. A model can lift weaker drafts towards a strong default and simultaneously increase dependence on a small set of common patterns.

Doshi and Hauser found this combination in short-story writing: access to generative AI ideas increased judged novelty and usefulness for individual stories, especially for less creative writers, while stories within the assisted conditions became more similar to one another. A later meta-analysis of generative AI and creativity reported a positive effect for human–AI creative performance alongside a negative pooled effect on diversity. Yet Ashkinaze and colleagues found that high exposure to AI ideas increased collective idea diversity in a large dynamic experiment, though it did not improve individual creativity. Wan and Kalman reported that deliberately varied AI personas could mitigate homogenisation in collaborative ideation.

The responsible conclusion is not “AI homogenises creativity”. The direction depends on the task, prompting, exposure pattern, population, metric and interface. Homogenisation is a property of a coupled search process, not an intrinsic substance emitted by a model. The same generative capacity can act as a common attractor, a bridge to remote regions or a source of noise.

Why twenty better drafts can cover fewer possibilities

Imagine four mechanism families have probabilities of appearing in one unconstrained generation: powered cooling 0.70, insulation 0.20, phase-change buffering 0.08 and irreversible witnessing 0.02. With one draw, the expected number of represented families is one. With twenty independent draws, the expected count rises, but not to four:

Here K is the number of mechanism families observed, pⱼ is the probability of family j, and n is the number of draws. For twenty independent samples, the expected count is about 3.13. The rare witness family appears only about one-third of the time. Independence is the generous assumption. In a chat, later generations inherit the vocabulary, examples and approvals of earlier turns, so samples are correlated. Twenty polished variants may cover less than the formula predicts.

Why ranking fails

Teams often ask a model to generate ideas and then rank them. This creates an elegant but fragile loop. The same statistical tendencies shape both candidate production and evaluation. A proposal close to familiar exemplars may be described more fluently, receive a more coherent rationale and win the ranking, even when a rougher candidate changes the dependency structure more radically.

Human ranking is not immune. Rietzschel and colleagues showed that idea generation and idea selection are distinct abilities. People can generate creative ideas yet select less creative ones, especially when obvious feasibility dominates. The remedy is not to remove feasibility. It is to apply it at the right stage. Hard constraints should gate impossible candidates. Within the feasible set, the team should preserve a portfolio of different mechanisms long enough to learn which uncertainty matters.

stateDiagram-v2
  [*] --> Claim
  state "The bottleneck moves from<br/>making to judging" as Claim
  state "Same artefact, different generator" as Test
  state "The order of contribution is<br/>causal" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Preserve judgement by design

The central design problem is not how to make a model generate more unusual text. It is how to arrange a joint system so that cheap generation expands search without displacing the judgements that define the work. A useful workflow has five moves: Frame, Range, Map, Break and Commit. These are not stages that must occur once in a line. They are functions that must remain visible whenever the team cycles.

Frame: state the conflict before requesting a solution

A frame is not a verbose prompt. It states the desired change, hard constraints, observables and prohibited dependencies. “Design an innovative cabinet” provides a topic. “Keep mock vials below 8°C for twelve hours during a 35°C outage, reveal any breach afterwards, and use no powered sensing” provides a causal demand.

The person or accountable team owns this move because a frame embodies values and exclusions. A model can challenge it by asking what has been omitted, but it cannot authorise the trade. The frame should exist before the first polished proposal so that later outputs can be rejected without negotiating against their fluency.

Range: sample mechanism families, not adjectives

“More creative”, “wilder” and “less obvious” are weak controls. Ask instead for solutions that depend on different physical, social or computational relations. In the cabinet case, the families might include stored thermal capacity, phase transition, evaporative heat removal, reversible adsorption, thermal isolation and irreversible witnessing. The list can come from experts, retrieval, models or all three.

Ranging is where models offer particular leverage. They can search across vocabularies, translate a relation into multiple domains and instantiate many variants. Yet the output should be grouped by dependency. The team stops requesting variants from a family once marginal samples cease adding a new intervention or failure mode.

Map: expose what transfers and what does not

For each remote source, record objects, relations, mapped intervention, predicted target effect and residue. Residue is what refuses translation. A plant’s transpiration may inspire evaporative cooling, but roots, metabolism and active regulation are not imported merely because the word “biomimetic” appears. Naming residue prevents a comparison from expanding into a false equivalence.

The mapping receipt also guards against post-hoc stories. Before a prototype, it states what result is expected if the borrowed mechanism matters. After the test, the team can compare the observation with the prediction rather than rewriting the analogy to fit any outcome.

Break: search for the intervention that separates stories

A break test changes one causal property. Increase humidity to challenge evaporation. Replace the phase-change pack with equal thermal mass. Apply a brief high-temperature spike to distinguish a cumulative chemical witness from an instantaneous maximum indicator. The aim is not to destroy every candidate. It is to create different predictions from rival explanations.

This move protects the team from fluent rationalisation. A model can assist by proposing failure modes, but an independent path should review them. In consequential work, the person who benefits from a proposal should not be the only person defining its test.

Commit: choose a reversible learning step

Creative work becomes consequential when resources, reputation or safety are placed behind an idea. Commitment therefore needs its own judgement. A promising candidate may deserve a small prototype but not a deployment. A high-scoring pair may be less informative than a slightly weaker pair that contrasts two mechanisms.

The commitment receipt states the chosen experiment, rejected alternatives, evidence used, unresolved assumptions, owner, stop conditions and recovery path. Creativity without a commitment boundary is theatre; commitment without preserved judgement is automation of taste.

Three receipts make judgement inspectable

The frame receipt records the problem conflict, hard constraints, observables, prohibited dependencies and who may alter them. The mapping receipt records the source, target, preserved relation, predicted intervention and break condition. The commitment receipt records the chosen experiment, portfolio rationale, evidence, owner and stop condition.

Receipts do not turn taste into bureaucracy. They preserve the load-bearing moves so that disagreement can attach to something. A reviewer can say the humidity constraint is missing, the phase-change analogy has no capacity calculation or the portfolio contains two versions of one mechanism. Without receipts, the discussion drifts towards prose quality and confidence.

How to interpret the audit

The audit tests whether a proposed portfolio contains candidates that are individually balanced, constraint-compliant and mechanistically different. It encodes three assumptions. First, a hard constraint is a gate rather than a negotiable penalty. Second, weakness on one core dimension should lower the whole candidate, which is why the geometric mean is used. Third, learning value rises when the chosen pair represents different mechanism families.

A positive result permits a bounded next step: build or simulate the selected pair under declared stop conditions. It does not permit deployment or a claim that the idea is objectively creative. A negative result means either every candidate violates a hard condition, quality is unbalanced, or the portfolio adds little causal contrast. The response is to reframe, retrieve a new family or improve evidence, not to lower the gate silently.

The sliders remain judgements. Experts may disagree, and that disagreement is informative. Record independent scores before discussion, compare the reasons and test the dimension producing the largest decision change. The tool makes judgement inspectable; it does not replace judgement.

function balancedQuality(candidate) {
  if (!candidate.hardConstraintPass) return 0;
  const dimensions = [
    candidate.novelty,
    candidate.fit,
    candidate.analogy,
    candidate.evidence,
    candidate.testability
  ];
  const product = dimensions.reduce((p, x) => p * x, 1);
  return Math.pow(product, 1 / dimensions.length) / 5;
}

function pairValue(a, b, lambda) {
  const diversity = a.mechanism === b.mechanism ? 0 : 1;
  return balancedQuality(a) + balancedQuality(b) + lambda * diversity;
}

The first function exposes the hard gate and balanced-quality assumption. The second exposes the portfolio policy. Replace the binary diversity term with a domain-specific distance only when that distance has a defensible meaning. Do not infer precision from the decimals; they support comparison and sensitivity testing.

Test the workflow, not the mythology

The strongest contribution of the framework is an experimental comparison. Randomly assign participants to three conditions. The human-only group receives the frame and ordinary research tools. The AI-first editor group receives a complete model proposal and may revise it. The judgement-preserving group must freeze a frame receipt, request mechanism families, complete one mapping receipt and specify a break test before seeing a polished synthesis.

Measure immediate output quality through blinded domain judges. Measure originality at both artefact and population levels. Code mechanism-family coverage and semantic similarity. Record review time, rejected candidates and confidence calibration. Then, after a delay, give participants a structurally related problem without AI. Measure unaided framing, transfer, ability to produce a discriminating test and willingness to reject a fluent default.

The prediction is not that the judgement-preserving group wins every immediate measure. AI-first editing may be faster and may produce the smoothest average artefact. The framework predicts that the judgement-preserving condition will cover more causal mechanisms, provide clearer break conditions and retain more unaided transfer. A result in which AI-first editing equals or exceeds it on those measures would weaken the claim.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: Two teams, one impossible cabinet
  M->>E: The bottleneck moves from making to judging
  E-->>M: compare rival explanations
  E->>D: Delegate expansion, keep commitment accountable
  D-->>Q: revise scope or act

Delegate expansion, keep commitment accountable

When a team says, “Let the model be creative,” the phrase hides four different choices. Should the model widen candidate supply? Should it search remote sources? Should it judge whether a transfer is valid? Should it decide what receives commitment? Treating these choices as one capability claim produces weak systems and confused evaluations.

The decision to change is simple: do not buy fluency as if it were creative judgement. Buy it as search capacity. Require the human–AI system to preserve the constraint, the mapping and the commitment boundary. That means writing the frame before requesting a polished answer, sampling across mechanism families, recording what an analogy predicts, designing a test that can make it fail, and funding a contrast rather than a single eloquent winner.

For an individual creator, begin with a fragment of your own judgement. State what must remain true, what you are unwilling to trade and what would surprise you. Ask the model for different relational routes, rather than only more versions. Delay synthesis until you can name at least two mechanisms. After the model helps, attempt one unaided transfer so you can see what capability remained yours.

For a team, change the review object. Do not review twenty paragraphs as twenty independent ideas. Cluster them by causal dependence. Gate hard failures. Select a portfolio that contains an informative contrast. Preserve frame, mapping and commitment receipts so a later reviewer can reconstruct why an idea crossed from possibility into action.

For a product designer, evaluate the interface as a cognitive intervention. Compare human-only, AI-only and combined baselines. Measure population diversity as well as average artefact quality. Vary the order of contribution. Test whether the interface develops, preserves or erodes the user’s ability to frame, map and reject. A co-creation product should not be judged only by what users can produce while the product is present.

For a research programme, stop asking whether AI is creative in the abstract. Specify the unit, mechanism and consequence. Is the claim about an artefact, a generator, a person-plus-tool system or a population? Is novelty semantic, causal or social? Does the proposed analogy predict an intervention? Does the process create a useful outcome, a new capability or merely a more fluent default?