Creativity Rejectionism

A procedure for appraising claims of the form "system S is creative" (or insightful, intelligent, understanding, reasoning) when S is non-human — an animal or an AI system. It encodes the argument of Shevlin (2021), Rethinking creative intelligence: comparative psychology and the concept of creativity (EJPS 11:16).

The skill exists because these claims feel tractable but aren't. "Creativity" is not a scientific natural kind with an agreed definition; it is a folk-psychological bundle of criteria whose membership conditions researchers actively dispute. Asserting or denying it of a non-human system inherits every one of those disputes, plus new ones unique to the non-human case. The job of this skill is to stop you from adjudicating the unanswerable question ("but is it really creative?") and redirect you toward the answerable one ("which operationalized capacities does it exhibit, on which measure?").

When this lens applies

Run the diagnostic before engaging the substance. The lens applies when all of these hold:

1. The claim ascribes a cognitive capacity (not a behavior) — "is creative", not "bent the wire".
2. The concept is a bundle of criteria with no consensus definition: people invoke different necessary/sufficient conditions and reach different verdicts. Creativity, insight, understanding, reasoning, intelligence, consciousness all qualify. Memory and learning are closer to settled and usually do not.
3. The subject is non-human (animal or machine), so the human escape hatches — agreed exemplars, first-person phenomenological access — are unavailable or contested.

If the request is really about a behavior ("did the crow raise the water level?") or a settled construct, skip the skill and answer directly. The skill earns its place only when someone is reaching for the contested umbrella term.

Step 1 — Name the criterion being smuggled in

Whoever makes the claim is implicitly using some theory of creativity. Surface it. There are four broad families (full taxonomy in references/theories-and-cases.md):

• Experiential — creativity requires surprise (Boden) or an "aha!"/insight phenomenology (Kounios & Beeman).
• Capacity/agency — creativity requires flair, judgement, intentional agency directed at the task (Gaut; Paul & Stokes).
• Spontaneity — creativity requires independence from prior knowledge and intentional control; it is unlike ordinary goal-directed behavior (Kronfeldner; Amabile).
• Process/stage — creativity is a temporal sequence: prepare→incubate→illuminate→verify (Wallas), or blind-variation→selective-retention (Campbell, Simonton), or generate→explore ("geneplore").

Two conditions are near-universal across families: the output must be novel and valuable. State which family the claim depends on, because the verdict will flip depending on the choice — a behavior that counts as creative under a spontaneity theory may fail an experiential one. Do not let a verdict ride on an unstated criterion.

Step 2 — Run the hazard checklist

For each hazard, say whether it bites this specific claim and how. These are the load-bearing arguments of the paper; cite them when you flag them.

1. Definitional disagreement (§4.1). Different theories partition nature differently. Worse than ordinary scientific dispute (wave/particle, kin selection) because here there is no agreement on how to identify the phenomenon at all, and — unlike memory research, which agrees on the concept and on tasks like n-back — creativity research lacks even that floor. Note that researchers disagree on whether value is required (Weisberg; Hills & Bird say no), whether surprise is required, whether the process is conscious. In the non-human case there are not even agreed exemplars to anchor on (some deny animal creativity wholesale — Fuentes 2017).
2. The novelty grain problem (§4.2). "Novel to the individual" is the least controversial criterion and still breaks. Individuate behavior finely and everything is novel (no two performances identical → novelty is ubiquitous and trivial). Individuate coarsely and nothing is (a painter trying a new form is still "painting" → novelty is near-impossible). There is no principled grain. For animals: is a crow's tool-bending novel, given wild crows bend twigs (Rutz et al. 2016)? The fine/coarse choice decides the verdict and nothing decides the choice. Contextualism about novelty leaks into a radical contextualism about creativity itself.
3. Value-ladenness (§4.2). Value varies across cultures and individuals. Anchoring value in the agent's own attitude fails (artists despised masterpieces — Michelangelo's Pietà, Virgil's Aeneid). For animals, metacognitive evaluation is mostly unavailable, so value must be cashed out objectively (fitness/survival) — but creativity-associated traits (neophilia, exploration) are frequently dangerous and fitness-reducing in the relevant environment (keas), and captive specimens can be "creative" in ways that would be lethal in the wild. Fitness and creativity come apart.
4. The criterion / measurement problem (§4.3). Human psychometrics (Alternative Uses, Remote Associates, Torrance) may measure divergent thinking rather than creativity per se (Amabile 1982). Arden et al. (2010): 45 neuroimaging studies, little overlap in implicated brain areas — the criterion problem had worsened, not resolved. If we can't pin a measure, neural signature, or criterion in humans, the prospect is far worse across the sensorimotor and cognitive diversity of animals.
5. Serendipity / reinforcement confounds (§3.2, §3.3). Many showcase cases reduce to chance + reinforcement + social learning (great tits and milk-bottle lids; meta-analysis of Aesop's-fable performance as trial-and-error + innate object preferences — Hennefield et al. 2019). Note though (footnote 6) that a reinforcement-learning origin does not automatically disqualify a behavior as creative — the Thorndike-to-Skinner-to-connectionism tradition grounds even human cognition in association. So "it's just RL" is a confound to flag, not an automatic defeater.

Step 3 — Evaluate the rescue strategies (don't endorse uncritically)

People will offer fixes. The paper grants each some value but finds each insufficient for rigorous comparative/AI science:

• Creativity comes in degrees. Intuitive, removes the binary, but if still tied to a unified kind it inherits every hazard above — degrees of an ill-defined thing don't help you apply it.
• Divide and conquer (Boden's combinational/exploratory/transformational; Kaufman & Beghetto's four-C). Helpful, but be explicit whether you're doing folk-conceptual analysis or building operational tools — conflating them breeds confusion and a needless proliferation of distinctions. If the goal is measurement, this strategy is really a step toward rejectionism.
• Prototype concept (Chen 2018, via Rosch): membership by similarity to exemplars, not necessary/sufficient conditions. Good for modeling folk usage. But culturally/context-relative similarity judgments are exactly what makes it ill-suited to rigorous comparative psychology — and Chen concedes it may fracture into multiple natural kinds.

Step 4 — Apply the rejectionist recommendation

Distinguish two positions:

• Weak rejectionism — a comparative science of creativity may be possible eventually, but isn't viable now given the open controversies.
• Strong rejectionism (the paper's tentative recommendation) — we already have good reason to think "creativity" is the wrong kind of concept for comparative psychology: more like wit, charisma, or good taste (folk, value-laden, context-bound) than like learning, memory, or g-factor (operationalizable). So "is animal/machine X creative?" is not a substantive scientific question. This is not human exceptionalism or mysterianism — Shevlin grants the component capacities are likely widespread in nature; the objection is to the umbrella concept, not to the abilities.

The constructive move: drop the umbrella term and substitute stipulatively-defined, operationalized constructs you can actually measure — behavioral plasticity, neophilia, exploration, innovation, divergent thinking, insight-as-restructuring (success after an impasse involving reconceptualization). Define them by fiat for the study at hand, with no deference to the folk concept.

Two payoffs to state explicitly:

1. Pluralism without essence-hunting — you can use typologies cutting across many dimensions of behavioral variability without theorizing what they all share.
2. No dangling question — if a scientist says "this satisfies my stipulated definition of behavioral-plasticity-type-3," there is no leftover "yes but is it really creative?" The austere vocabulary dissolves the trap. (Compare: episodic memory → "episodic-like memory" operationalized as what/when/where; intelligence → g-factor.)

Acknowledge the honest objection (§5): this may "move the bump in the carpet" — novelty is scarcely more perspicuous than creativity. The reply is that stipulation + pluralism still beats chasing a folk essence, even if residual hazards remain.

Step 5 — For AI / LLM claims specifically (footnote 16)

This is where the skill most often fires in practice. Shevlin flags that strong rejectionism "might also be the best path" for whether a machine is creative: we can sensibly aim to build systems that produce novel and valuable outputs and respond flexibly and appropriately to environmental change, and the narrowly conceptual question of whether the machine is really creative is "largely irrelevant to achieving these ends."

Operational consequence for benchmark/eval design:

• Don't build a "creativity benchmark." Build benchmarks for stipulated, measurable sub-capacities (output novelty under a fixed individuation grain; value under a declared, task-specific rubric; behavioral flexibility under distribution shift; divergent generation; restructuring after impasse).
• Pre-register the individuation grain and the value rubric — Step 2's grain and value-ladenness hazards are the two that most silently inflate or deflate model "creativity" scores.
• Treat "but does it understand/reason/create?" as out of scope for the eval, by construction. The eval reports construct scores; it does not adjudicate the folk term. This is the strong-rejectionist discipline applied to evaluation.
• The same recipe transfers to other contested LLM capacity terms (reasoning, understanding, agency, planning): substitute operational constructs, declare measures, refuse the dangling essentialist question.

Output template

When appraising a concrete claim, structure the response like this:

## The claim
[Restate it, and name the system S and the behavior(s) at issue.]

## Criterion in play
[Which of the four theory families the claim relies on — and whether it's stated or smuggled.]

## Hazards that bite
[For each relevant hazard from Step 2: how it applies here, with the cite. Skip ones that don't bite.]

## Rescue strategies, if invoked
[Only if the source offers degrees / divide-and-conquer / prototype — evaluate per Step 3.]

## Rejectionist reframing
[Replace the umbrella verdict with operationalized constructs + how you'd measure each.
 For AI: phrase as an eval-design recommendation.]

## Bottom line
[Whether the original "is it creative?" question is substantive here, or should be dissolved.]

Scope notes

• Stay agnostic on the human case. The paper confines strong rejectionism to non-humans; humans retain agreed exemplars and phenomenological access. Don't over-extend the conclusion to "humans aren't creative either" — that's not the argument.
• Don't smuggle in human exceptionalism. The component capacities are likely abundant in nature and in machines; the target is the concept, not the abilities.
• The four-family taxonomy, the full animal case studies (Köhler's chimps, Pryor's porpoises, Betty the crow, Goffin cockatoos, Aesop's-fable corvids), and the strategy details live in references/theories-and-cases.md. Load it when you need the specific theories, citations, or examples rather than the bare procedure.