Can AI Be Creative? Define It First
4 min read · updated August 3, 2026
Arguments about machine creativity almost never disagree about a fact. They disagree about a definition, silently, while both sides believe they are talking about the world. The way out is to use the definitions the field studying creativity already uses, and then see which parts of the question survive.
The definition that is actually used
Creativity research has converged on a two-part criterion that is almost boringly standard: a creative product is novel and it is valuable — appropriate, useful or effective for some purpose. Novelty alone is noise. Value alone is competence. Both together is the thing.
A second standard distinction does most of the remaining work. Novelty can be relative to the individual producing it or relative to everything anyone has produced. The literature calls these little-c and big-C creativity, or historical versus psychological novelty. A child deriving a known theorem unaided has done something creative in the first sense and nothing new in the second. Almost every argument about AI creativity is one person meaning the first and another meaning the second.
Applied straightforwardly: a model producing a competent sonnet on a subject no one has written a sonnet about has produced something novel in the first sense and valuable by the standards of the form. Denying that requires either a different definition or a claim about process rather than product — and the process objection is a real one, treated below.
Three kinds, three different answers
A widely used taxonomy splits creative acts into three types by what they do to the space of possibilities. It is unusually useful here because the three come out differently.
| Type | Description |
|---|---|
| combinational | Unfamiliar combinations of familiar ideas — a metaphor, an analogy, a mash-up of two genres. Models are demonstrably strong here; the mechanism is well suited to it. |
| exploratory | Working within an established style or set of rules to find points in it that nobody has visited. Models do this well when the space is well represented in training data, and poorly when it is not. |
| transformational | Changing the rules of the space itself, so that what was previously impossible becomes possible. This is the hard case and there is no clear demonstration of it. |
The third row is where the honest uncertainty sits. There is a real argument that a system trained to model an existing distribution is structurally biased toward staying inside it, and a real counterargument that transformational moves in human history also came from people saturated in the existing tradition. Neither has been settled, and it is not obvious what experiment would settle it, since recognising a transformational move usually takes decades.
The measurable weakness: distribution collapse
Here is a concrete, checkable finding rather than a philosophical point, and it is the strongest empirical argument against calling current models creative in the ordinary sense.
Ask a model the same open-ended creative prompt many times. The individual outputs are often good. The set of outputs is far narrower than the set a comparable number of people would produce: the same handful of names, the same structural moves, the same openings. Individual novelty is high, collective diversity is low.
The mechanism is not mysterious. Sampling picks high-probability continuations, so it concentrates on the modes of the distribution. Post-training that optimises for a preferred answer sharpens that concentration further, since the preference signal rewards the response people rate highest rather than the spread of responses. You can widen the distribution with sampling parameters, and you pay for it in coherence — which is a real trade-off, not a bug to be fixed.
For anyone using a model to generate options, this is the practical finding in the whole page: ask for many and expect them to cluster, and treat divergence as something you have to engineer with varied prompts, varied seeds and explicit constraints rather than something you get for free.
The objection about process
The serious objection is not about the output at all. It is that creativity is a property of an agent doing something, not a property of an artefact: it requires intention, a reason for the choice, and something at stake in making it. On this account a model cannot be creative for the same reason a kaleidoscope cannot, however striking the pattern, because there is no one for whom the pattern is an attempt at anything.
This is coherent and it is not refutable by pointing at better output, which is why the two sides never converge. It is worth noticing what it costs, though: on a strict reading it also denies creativity to collaborative works, to processes that use randomness deliberately, and to any case where the maker cannot articulate why they made the choice — which includes a large fraction of what artists report about their own practice. Whether that consequence is acceptable is where the argument actually lives.
What this means if you make things
- Strong at combination, weak at variety. Use it to cross domains and to produce many candidates, and do the selecting yourself, because the selection is where the taste is.
- Expect the modes. If an output feels familiar, that is the distribution, not a failure of your prompt. Constrain explicitly — forbid the obvious answer by name — and the space opens.
- Novelty and value need separate checks. A model is better at producing something plausible than at telling you whether it is good. The value half of the definition still needs a judge who knows the field.