Can Machines Be Creative?

A machine paints an image no human has seen before.

A program discovers a strange chess strategy.

An algorithm designs an antenna that looks unlike anything an engineer would sketch.

Is that creativity?

The answer depends on what we mean by:

creative.

Creativity Is Not One Thing

A useful starting point is to separate several ideas:

• novelty,
• usefulness,
• surprise,
• originality,
• intention.

A product can satisfy some without satisfying all.

Novelty

A result is novel if it is not merely an exact copy of something already present.

This is the easiest requirement.

Random noise can also be novel.

So novelty alone is insufficient.

Value

Creative outputs are usually valued within some context.

They may be:

• useful,
• beautiful,
• explanatory,
• interesting.

Creativity is partly evaluative.

Surprise

Some creative acts violate expectation while still making sense.

The output is not merely different.

It is unexpectedly appropriate.

Boden’s Framework

Margaret Boden distinguished important forms of creativity.

These include:

• combinational creativity,
• exploratory creativity,
• transformational creativity.

The distinctions are especially useful for machine creativity.

Combinational Creativity

Combinational creativity brings familiar ideas together in a new way.

Examples:

• a new metaphor,
• a fusion genre.

Machines can clearly perform this kind of recombination.

Exploratory Creativity

Exploratory creativity searches a structured conceptual space.

The rules remain fixed.

The system discovers an unusual point inside that space.

Game strategies are a good example.

Transformational Creativity

Transformational creativity changes the rules of the space itself.

A new artistic movement may redefine what counts as a valid work.

This is a stronger notion of creativity.

Conceptual Spaces

A conceptual space is a structured set of possible ideas or artifacts.

Creativity can be viewed as movement through such a space.

The important question is:

who defines the space?

Search and Creativity

If a machine searches a huge possibility space and finds something valuable, is that creative?

Humans also search.

The difference may be in:

• representation,
• evaluation,
• intention.

Search is not automatically uncreative.

Evolutionary Creativity

Evolutionary algorithms can generate unexpected designs.

They are especially interesting because designers do not specify the final form.

Selection discovers it.

This looks more creative than direct construction.

Evolved Antennas

NASA experiments famously used evolutionary computation to design unusual antenna shapes.

The resulting structures were functional but visually unintuitive.

The system found solutions outside ordinary engineering habits.

Is Surprise Enough?

Humans may call an output creative because it surprises us.

But surprise reflects:

our expectations.

A random mutation can surprise without being insightful.

Evaluation still matters.

Generative Models

Generative AI can produce:

• text,
• images,
• music,
• code.

It learns statistical structure from large datasets.

Generation combines learned regularities in new configurations.

Statistical Generation

Critics sometimes say:

It is only predicting.

But prediction can generate complex novelty.

The real question is not whether statistics are involved.

Human cognition is also shaped by statistical learning.

Memorization Objection

If a model merely reproduces training examples, creativity is weak.

But modern models can also generate outputs not present verbatim in training.

The challenge is measuring:

how much is retrieval, how much is recombination, how much is abstraction.

Training Data Matters

Machine creativity depends on cultural material produced by others.

But human creativity does too.

Artists learn from:

• teachers,
• traditions,
• previous works.

No creator begins from nothing.

Influence vs Copying

Influence and copying are different.

A creative system may transform patterns.

But exact or near-exact reproduction can raise:

• copyright,
• attribution,
• ethical concerns.

Technical novelty does not settle legal or moral questions.

Style Imitation

A model can imitate the style of an artist.

Is this creativity?

It may show learned structure.

But imitation alone is not strong evidence of originality.

Intentionality

One powerful objection is:

Creativity requires intention.

A person creates with:

• goals,
• reasons,
• aesthetic judgment.

Does a machine have such intentions?

Functional Intention

A machine may have functional goals:

maximize score, satisfy prompt, optimize reward.

This is not necessarily the same as human subjective intention.

We should distinguish them.

Phenomenal Intention

Human creativity can involve:

• desire,
• curiosity,
• emotion.

A machine may produce creative artifacts without any inner experience.

If experience is required by definition, machine creativity becomes impossible by stipulation.

Product vs Process

One can judge creativity at two levels:

Product creativity

Is the output novel and valuable?

Process creativity

Was the creation process itself creative?

A machine may succeed strongly on the first while remaining disputed on the second.

P-Creativity

Boden uses P-creativity for psychological creativity:

an idea is new to the person producing it.

H-Creativity

H-creativity means historically creative:

the idea is new to humanity.

A machine can potentially produce H-creative results.

That would be strong evidence of creative contribution.

AlphaGo Move 37

In 2016, AlphaGo’s move 37 in game two against Lee Sedol was widely described as surprising and creative.

It violated professional expectations.

Yet it emerged from training and search.

Does that make it less creative?

Human Creativity Also Has Mechanisms

If discovering a mechanism removes creativity, then neuroscience could eventually explain human creativity away.

That seems mistaken.

Mechanistic explanation and creativity can coexist.

Mechanism Does Not Eliminate Achievement

A human poem depends on:

• neurons,
• memory,
• linguistic habits.

Knowing the mechanism does not make the poem uncreative.

The same principle may apply to machines.

Evaluation Problem

A creative system must somehow distinguish:

interesting novelty

from:

meaningless novelty.

Evaluation can come from:

• humans,
• reward models,
• environmental success.

Who evaluates matters.

Autonomous Evaluation

Stronger machine creativity would involve the system developing or revising its own criteria.

It might reject outputs not because humans told it to, but because of internally maintained standards.

This begins to resemble agency.

Curiosity

Artificial systems can be given intrinsic motivation for:

• novelty,
• information gain.

This can produce exploratory behavior.

But engineered curiosity is not necessarily felt curiosity.

Creativity as Constraint Satisfaction

Creative work often solves problems under constraints.

A composer must respect:

• rhythm,
• harmony,
• style.

Creativity occurs inside structure.

Machines are well suited to constrained search.

Constraint Breaking

Exceptional creativity sometimes violates established constraints.

The system must recognize:

which rule is essential, which can be broken.

This is harder.

Transformational Systems

A truly transformational machine might invent:

• a new game,
• a new mathematical representation,
• a new aesthetic grammar.

Not just a new instance.

It changes the space.

Scientific Creativity

Science requires more than generating hypotheses.

A creative scientist:

• notices anomalies,
• proposes explanations,
• designs tests.

Machines may increasingly participate in all three.

Automated Discovery

AI systems have been used to search:

• molecules,
• materials,
• proofs.

If a system proposes a genuinely useful structure no human considered, calling the contribution creative becomes reasonable.

Mathematical Creativity

A theorem prover may discover an unexpected proof.

Does creativity lie in:

• finding the proof,
• inventing the lemma,
• choosing the representation?

Different levels support different judgments.

Language Models and Creativity

Language models can generate:

• metaphors,
• stories,
• concepts.

Some outputs are banal.

Some are unexpectedly strong.

Capability is statistical, not uniform.

Prompt Dependence

If a human specifies:

write a sonnet about black holes in the style of X,

how much creativity belongs to:

• user,
• model,
• training culture?

Authorship becomes distributed.

Co-Creation

Human–AI systems may be better understood as co-creative systems.

The human supplies:

• goals,
• taste,
• selection.

The model supplies:

• variation,
• recombination,
• speed.

Creativity emerges from interaction.

The Curator Role

Selection is itself creative.

A photographer chooses one image.

An editor chooses one sentence.

When a human selects machine outputs, the creative process includes curation.

Tool or Collaborator?

A paintbrush is a tool.

A generative model proposes alternatives autonomously.

The boundary between:

tool

and:

collaborator

becomes less clear.

Creativity and Responsibility

If a machine creates something harmful, who is responsible?

Possibilities include:

• developer,
• deployer,
• user.

At present, legal and moral responsibility remains human-centered.

Creativity Without Consciousness

Could an unconscious process create?

Evolution does.

Natural selection produced:

• eyes,
• wings,
• brains

without foresight.

This suggests conscious intention may not be necessary for generative novelty.

Evolution as Counterexample

If we call biological evolution “creative” metaphorically, then creativity can exist without a creator.

But metaphorical creativity is not identical to artistic agency.

The distinction matters.

Creativity as Emergence

A complex system may generate structures no component explicitly planned.

Creativity can emerge from:

• search,
• recombination,
• selection.

This fits earlier themes of emergence.

The Anthropocentric Trap

If creativity is defined as:

what humans uniquely do,

machines cannot qualify by definition.

A better definition should specify measurable properties.

Then we can ask whether machines satisfy them.

A Practical Criterion

A system looks more creative when it can:

• generate novelty,
• satisfy meaningful constraints,
• evaluate alternatives,
• transfer ideas,
• surprise competent observers.

This avoids requiring human likeness.

Stronger Criterion

A stronger notion might also require:

• self-generated goals,
• self-evaluation,
• conceptual transformation.

Current systems vary greatly on these dimensions.

The Philosophical Lesson

Machines can clearly generate outputs that are:

• novel,
• useful,
• surprising.

Whether this counts as full creativity depends on whether creativity requires:

• intention,
• consciousness,
• autonomous values.

The product question is increasingly easy.

The agency question remains open.

The Next Question

Creativity is often tied to:

• feeling,
• mood,
• desire.

Could machines possess anything analogous to emotion?

Can an artificial system:

have emotions?