Can AI Be Creative? The Deep Question Behind Every AI-Generated Image

In August 2022, a digital image called Théâtre D'opéra Spatial won first prize in the emerging artist digital arts category of the Colorado State Fair. It had been made with the image generator Midjourney, and the judges later said they had not known that. The episode exposed a question that no benchmark can settle: can AI be creative, or does it only produce things that look creative? This article separates the competing definitions of creativity, shows how AI performs on each, explains the consciousness argument, and looks at what courts and working artists have decided in practice.

Key takeaways

• "Creative" has at least three working definitions: producing new and valuable things, changing the rules of a field, and communicating something from one mind to another. AI fares very differently on each.
• On a standard divergent-thinking test, three chatbots averaged higher than 256 human participants, but the best human responses matched or beat the best from the chatbots.
• A 2023 review by 19 researchers concluded that no current AI systems are conscious, while noting no obvious technical barrier to building systems that meet its indicators.
• A study published in 2024 found that AI-generated story ideas made individual writers' stories better, but made the stories more alike.
• US courts and the Copyright Office require human authorship, so the legal answer to "who made it" currently names a person or no one.

Three things people mean by "creative"

The argument over AI creativity is often an argument about vocabulary. Each side is using a different definition, and each definition has a different test.

The first definition is about outputs. The cognitive scientist Margaret Boden, who wrote The Creative Mind: Myths and Mechanisms in 1990, defines creativity as producing ideas or artifacts that are new, surprising, and valuable (Computational creativity, Wikipedia). Under that definition, only the result matters. If a viewer finds an image novel and worthwhile, the image is creative, whatever made it.

The second definition is about what kind of newness is involved. Boden separates combinational creativity (joining familiar ideas in unfamiliar ways, as in a metaphor or a pun), exploratory creativity (searching a style or genre for possibilities nobody has tried yet), and transformational creativity (changing the rules of the space itself, as cubism did to painting). She also distinguishes P-creativity, which is new to the person who produced it, from H-creativity, which is new in history. A child who invents a pun for the first time is P-creative, while the person who invented the pun in the first place was H-creative.

The third definition is about communication. On this view, art is something one mind does to another. A poem is not only a pleasing arrangement of words but an attempt by someone to say something, with the risk of being misunderstood. If that is what creativity requires, a system that has nothing to say cannot be creative, however good the output.

Definition	What counts as evidence	How current AI looks
Output: new, surprising, valuable	Ratings of novelty and quality	Strong in several tests, with caveats
Combinational	Unusual but apt combinations	Strong
Exploratory	New moves inside a known style	Demonstrated in games and some design tasks
Transformational	Changing the rules of the field	Contested, with few clear cases
Intentional and communicative	A mind trying to say something	Not demonstrated; depends on open questions about consciousness

The table is a map of the disagreement. When someone says "AI is obviously creative," they usually mean the first row. When someone says "AI cannot be creative," they usually mean the last.

What a diffusion model does when it makes an image

To judge whether the process counts as creative, it helps to know what the process is. Most current image generators use diffusion models. During training, the model is shown a huge number of images, each progressively corrupted with random noise, and it learns to predict and remove that noise. To generate an image, the system starts from pure noise and removes it step by step, guided by a numerical representation of the text prompt. Our explainer on how diffusion models work walks through the details.

A useful analogy is a sculptor who starts from a block of marble and removes everything that is not the statue. The diffusion model starts with a block of static and removes everything that is not a plausible image of what the prompt describes. The learned part, the sense of what plausible images look like, comes from the training data.

Where does novelty come from? Partly from randomness. The starting noise is random, so the same prompt can produce different images, and the settings that control sampling can make results more conservative or more adventurous. The same principle applies to text, as our article on temperature and top-p sampling shows. The randomness is a feature of the algorithm, not a sign of intent.

This is also why the "just remixing" description is both right and incomplete. The model does not store and cut up its training images in a collage. It learns statistical regularities, and a new image is a new sample from what it learned. Whether a sample from a learned distribution counts as a creative act is exactly what the definitions disagree about. Older systems worked differently. Generative adversarial networks, covered in our piece on how GANs work, train one network to produce images and another to catch fakes, which pushes outputs toward realism.

Where AI already scores well

The output-based definition can be tested, and the results are more interesting than either side usually admits.

Mika Koivisto and Simone Grassini compared 256 human participants with three generative AI chatbots in a 2023 Scientific Reports study. They used the alternate uses task, a classic test of divergent thinking in which people propose unusual uses for an ordinary object such as a rope or a pencil. On average the chatbots scored higher than the humans. However, the best human responses matched or surpassed the best from the chatbots, and the authors concluded that AI can be a useful tool for boosting creativity while human creativity may be hard to fully replicate.

Two cautions apply. A divergent-thinking test measures one slice of creativity, namely how many unusual and fitting ideas a respondent can list, and it does not measure whether any idea changes anything. And the test is scored by raters, so it measures how ideas look, not what produced them.

A second study points to a cost. Anil Doshi and Oliver Hauser ran an online experiment, published in Science Advances in 2024, in which some writers received story ideas from a large language model and others did not. Stories written with AI ideas were rated more creative, better written, and more enjoyable, with the largest gains among writers who were less creative to begin with. But the AI-assisted stories were also more similar to one another. The authors describe a social dilemma: individuals benefit, while the group's collective range of novel content narrows.

The result is a useful reminder that creativity can be measured at two levels. At the individual level, AI help can raise quality. At the level of a whole culture, a shared tool can pull outputs toward a common middle. A model that suggests the same plausible ideas to a million people is a source of convergence as well as novelty.

Game 2, move 37: exploratory creativity in practice

The best-known case of a machine surprising experts is in the game of Go. In the second game of the 2016 match between AlphaGo and Lee Sedol, played on March 10, 2016, AlphaGo's 37th move was a shoulder hit that most professional players would not have considered. The professional Michael Redmond called the move "creative" and "unique," while An Younggil called it "a rare and intriguing shoulder hit." AlphaGo won the match 4 to 1 (AlphaGo versus Lee Sedol, Wikipedia).

In Boden's terms, this looks like exploratory creativity: a search of a rule-bound space that found a place experts had not looked. It was not transformational, since the rules of Go did not change. But the game shows that machine novelty can be real, valuable, and surprising to experts. Our article on AlphaGo explains how the system learned.

Go is also an unusually friendly case. The goal is unambiguous, the rules are fixed, and a win is a clear signal of value. Art offers none of those. There is no score for a painting, so a system cannot find "good" moves by self-play in the same way. That difference is one reason AI success in games does not automatically transfer to the arts.

The Lovelace objection and its limits

One of the oldest objections to machine creativity comes from Ada Lovelace, who argued that a machine can only do what it has been programmed to do, and so cannot originate anything. Computational creativity researchers note that not every theorist accepts the premise, since a learning system can produce results its programmers did not foresee.

Modern systems make the objection less simple. Nobody wrote a rule telling AlphaGo to play move 37, and nobody wrote the specific image that a diffusion model produces. What the programmers did specify is the training process, the data, and the objective. A fairer version of the objection is that the system's goals come from outside: it does not choose what matters, and it does not decide to make something because it has something to say.

This is where the output and communication definitions part ways. A system can produce surprising and valuable results without choosing its own goals. Whether the choosing is essential to creativity is a philosophical judgment, not an experimental result. Boden's own framework is deliberately neutral on it, since her three types describe what a result does to a space of possibilities rather than the inner life of the maker.

The consciousness question

Underneath the intention argument is a harder one: is there anything it is like to be the system? Thomas Nagel posed the question in a 1974 essay in The Philosophical Review, "What Is It Like to Be a Bat?" His point was that an organism has conscious states only if there is something it is like to be that organism, from the organism's own point of view, and that a purely objective description of the brain cannot capture that point of view. Our overview of the hard problem of consciousness covers the debate that followed.

If the communication definition of creativity is right, then consciousness, or at least some kind of intention, matters. How would one find out whether a system has it? A 2023 report by 19 researchers, including Patrick Butlin, Robert Long, Yoshua Bengio, and Jonathan Birch, tried to do this by deriving "indicator properties" from leading neuroscientific theories of consciousness, such as global workspace theory and recurrent processing theory, and checking AI systems against them. Their conclusion was that no current AI systems are conscious, but that there are no obvious technical barriers to building systems that satisfy the indicators (Butlin et al., arXiv).

That is a careful claim, and it rewards a close reading. It does not say machines can never be conscious, and it does not say the theories are correct. It says that if those theories are roughly right, today's systems do not meet the conditions. The question is therefore open in principle and, by that report's assessment, answered in the negative for current systems.

When a language model writes a sentence about grief, nothing in the evidence suggests that it feels grief. A model's description of an experience is built from patterns in human descriptions. A reader can still be moved by the sentence, which is a separate fact about the reader. The distinction between what a text does to a reader and what produced it is the heart of the creativity debate.

Who is the author? Courts, prizes, and one expensive portrait

Law has to answer the "who made it" question in practice, and its answers are instructive.

Jason Allen, who made Théâtre D'opéra Spatial, said he used at least 624 text prompts and revisions, then edited the image in Photoshop and upscaled it. The Colorado fair's judges said they would have given him the top prize anyway, but the fair began requiring disclosure of AI use in 2023. When Allen applied for copyright, the US Copyright Office initially refused to register the work, and in September 2023 its Review Board ruled it ineligible because the human creative input was too minimal (Théâtre D'opéra Spatial, Wikipedia).

The wider legal picture is that US courts and the Copyright Office require human authorship. The D.C. Circuit affirmed in March 2025 the refusal to register works generated by a machine. In January 2025 the Office said that AI-assisted works can be protected when a human's creative expression is evident, for example through creative arrangement or modification, while works whose expressive elements are determined by a machine remain unprotected. A composite work, A Single Piece of American Cheese, was registered that month on the basis of human selection, arrangement, and coordination of AI outputs (Artificial intelligence and copyright, Wikipedia). Our article on AI art, copyright, and ethics looks at the ownership questions in more detail.

The art market has had its own version of the argument. In October 2018 Christie's sold Edmond de Belamy, a GAN-generated portrait from the Paris collective Obvious, for US$432,500 against an estimate of US$7,000 to US$10,000. The model was trained on about 15,000 portraits from the WikiArt database. But the team had used an open-source codebase written by Robbie Barrat, who was not affiliated with Obvious, and a member of Obvious admitted that there was not a big percentage of the code that had been modified. Barrat publicly questioned the sale (Edmond de Belamy, Wikipedia).

Notice that nobody in these stories credited the machine as the author. The debate was over which humans deserved credit: the prompt writer, the code author, the artists whose work trained the model, or the buyer who paid for the story. Authorship was treated as a property of people even when the process was automated.

Artists who treat AI as a medium

Working artists have mostly set aside the question of whether the system is creative and asked what it can do. The German artist Mario Klingemann, born in 1970 and based in Munich, works with neural networks, code, and algorithms to examine creativity, culture, and perception. His Memories of Passersby I was the first AI-made work auctioned at Sotheby's, in 2019, and his The Butcher's Son won the Lumen Prize Gold Award in 2018. In 2023 he presented A.I.C.C.A., a robotic dog that produces art critiques (Mario Klingemann, Wikipedia).

What these practices share is an understanding of the tool as a material with its own properties, much like clay, ink, or a camera. The artist chooses what to train on, which results to keep, how to frame them, and what the work is for. The creativity, on this account, is in the decisions around the system, and the model supplies a distinctive kind of variation. The same division of labor applies in AI music and in style transfer, where the human sets the material and the selection, and the system supplies the variation.

Photography offers a useful comparison. Early critics of photography made a similar argument, that a machine made the image, so photographers were not artists. Over time the decisions around the machine, such as what to point it at, when to press the shutter, and how to print, came to be recognized as the art. Whether generative AI follows the same arc is an open question. A camera records the world, while a generative model reproduces patterns in other people's work, which is a harder case.

Does it matter whether the machine is "really" creative?

There is a practical response to all this: if the images are indistinguishable from human work, the label does not matter. For many uses, that is right. An advertising team that needs a background image for a campaign cares whether it works, and not whether the process involved intention.

For art as a cultural practice, the answer may differ. People value art partly because it tells them what it was like for someone else to live at a particular time and place. A work can still be meaningful to a viewer who knows a model made it, because meaning also arises in the viewer. But the sense of contact with another person's perspective, which is a large part of why many people value art, is missing when there is no perspective to contact.

The training data raises a separate issue. The systems depend on a large body of human creative work, and a technology that reduces the demand for that work could weaken the supply of the thing it depends on. The Doshi and Hauser result adds a related worry, that shared tools may narrow the range of what gets made. Those are questions about economics and culture, and the philosophy of mind does not resolve them. Our article on the rise of AI art covers the controversy and the artist response in more depth.

What people get wrong about AI creativity

"It is just copying." Diffusion models and similar systems learn statistical regularities from training data and generate new samples. The typical output is a new combination of learned patterns, which is why the more interesting question is whether novel combination is enough.

"If the output is good, the machine must be creative." That depends on the definition. Good output satisfies the output-based test. It does not by itself settle whether intention, choice, or experience were involved.

"AI is clearly conscious because it talks about feelings." Fluent descriptions of feelings are built from human descriptions. The 2023 expert review looked for structural indicators from scientific theories, and did not treat conversation as evidence.

"AI will never be creative." Machines have already produced exploratory novelty in games and rated well on some creativity tests. Whether they will meet the stricter definitions depends on facts about consciousness and agency that nobody yet knows.

"Human creativity does not depend on prior work." Human creators also learn from predecessors, and Boden's combinational type describes this directly. The difference lies in the combination of experience, motive, and accountability that a person brings, rather than in starting from nothing. For the human side of the story, see our overview of the psychology of creativity.

The takeaway

The honest answer is "it depends on what you are asking for." If creativity means new, surprising, and valuable results, current AI already qualifies in measurable ways, though the best humans still match it and heavy use of shared tools may make results more alike. If creativity means a mind choosing to say something, the evidence does not support it today: expert review finds no conscious AI systems, and the law still looks for a human author. The two answers are compatible. The more useful question for most people is not whether the machine is creative, but who chose the result, who gets credit, and what is lost if the human work that trained the model becomes harder to sustain.

FAQ

Q: Can AI be creative? A: By output-based definitions, yes: chatbots outscored 256 humans on average on one divergent-thinking test, and AlphaGo played a move that professionals called creative. By definitions requiring intention and experience, the evidence is much weaker. Which answer applies depends on the definition used.

Q: What are Boden's three types of creativity? A: Margaret Boden distinguishes combinational creativity (joining familiar ideas in new ways), exploratory creativity (searching a conceptual space such as a style), and transformational creativity (changing the rules of that space). She also separates P-creativity, new to the maker, from H-creativity, new in history.

Q: Does AI need consciousness to be creative? A: That is contested. Output-based definitions do not require it. Definitions based on communication or intention may. A 2023 review by 19 researchers concluded that no current AI systems are conscious, while saying there are no obvious technical barriers to building systems that meet its indicators.

Q: Can AI-generated art be copyrighted in the United States? A: Only to the extent that a human contributed creative expression. The Copyright Office refused to register Théâtre D'opéra Spatial, and its Review Board upheld that in September 2023 because the human input was too limited. In March 2025 the D.C. Circuit affirmed the refusal to register works generated solely by a machine.

Q: Does AI make human creativity more or less diverse? A: A 2024 Science Advances experiment found that AI-generated story ideas improved individual writers' stories, especially for less creative writers, but made the stories more similar to one another. The authors describe a tension between individual benefit and collective variety.

Q: Did AlphaGo's move 37 count as creativity? A: Professional players described it as creative and unique, and most would not have considered it. In Boden's terms it looks like exploratory creativity within the fixed rules of Go. Whether it reflects intention is a separate question, because AlphaGo did not choose its goal.

Q: Who is the artist when someone uses Midjourney? A: Courts and registrars treat the human as the possible author and ask how much creative control that person exercised. Jason Allen reported using at least 624 prompts and revisions, but the Copyright Office still held the human contribution too limited to register the work.

---

Read next:

• AI Music Composition: How Neural Networks Learn to Make Music
• Generative AI Art: How Diffusion Models Actually Work
• Neural Style Transfer: How AI Learns to Remix Artistic Style