A generated image surprises a designer. It contains a visual solution the designer had not considered, and it helps a team move forward. Was the system creative? Mark Runco’s paper, The discovery and innovation of AI does not qualify as creativity, answers no. Its title is a thesis to examine, not a finding that a laboratory experiment has settled. The paper argues that useful or novel output is insufficient if the underlying process lacks qualities central to human creativity.
This research brief treats Runco’s 2024 online paper as a conceptual argument. It identifies the claim, the process distinction that carries it, the evidence that can and cannot bear on it, and serious alternative definitions. A reader need not accept Runco’s conclusion to learn from its insistence on asking what exactly is being credited when a machine produces something surprising.
The paper’s question is conceptual
Runco asks whether the discovery and innovation exhibited by artificial intelligence should be called creativity in the same sense as human creative activity. His answer depends on more than the novelty of a product. The publisher abstract says AI output can appear creative and be attributed with creativity, yet the processes that produce it may be better described as artificial creativity, discovery or innovation. In his account, extending the full term to AI risks diluting what it names in human life.
That is not a randomized comparison of artists and models. The article is a review and argument drawing on distinctions among human memory, perception and learning, and machine processes. It proposes a criterion for interpreting outputs. Its empirical references may support claims about particular cognitive operations, but the final boundary between “authentic” and “artificial” creativity is partly a definitional judgment. Readers should assess the argument as such.
The word “qualify” can make the question seem like an exam with one universal rubric. In fact, creativity has multiple research uses. An assessor might ask whether a product is novel and valuable in a domain. A psychologist might ask what a person intended, how a problem was found and how a solution developed. A designer might ask whether an interaction helped a team think. Different questions can produce different legitimate answers without making the evidence meaningless.
Why outcome alone can mislead
A product-focused test has a practical advantage: observers can compare outputs without needing access to private mental states. Amabile’s consensual assessment technique made expert judgment of products a useful research method. If knowledgeable judges agree that an artifact is original and apt, that is real information about the artifact. The method does not automatically reveal who deserves credit for every step or whether its producer had intention in a human sense.
Runco’s worry begins at that jump. If a generated poem receives a high creativity rating, the observation concerns the poem and its evaluators. It does not, without more evidence, establish that the model has a lived aim, notices an unsolved human problem or cares whether its words are true. The output may be valuable to a reader while the attribution to a producer remains disputed. The paper asks us to resist using an easy-to-measure product as a complete theory of creativity.
There is a possible overcorrection, however. Human creativity is also frequently judged from products because the full process is hidden. An artist can improvise, use chance, borrow a tool or collaborate. A process criterion should not become a romantic demand that each human work arise from isolated, fully conscious intention. The challenge is to specify which process features are necessary, why they matter and how they could be investigated fairly.
Runco’s process distinction
Runco contrasts machine operations with human processes of memory, perception, problem discovery and learning. The publisher abstract describes generative AI as potentially innovative but not truly creative, with its processes suggesting discovery rather than creativity. That distinction matters because novelty can arise in many ways. A system can search a possibility space and return an unexpected pattern. A person can confront a meaningful problem, revise an aim and decide that a new form expresses something worth communicating.
The argument is strongest when it identifies a feature missing from the machine case that is important to the concept, such as autonomous problem finding, authentic intention or participation in a life of meanings. It is weaker if it assumes that every output from statistical learning is therefore unoriginal. Recombination occurs in human work too. Describing the mechanism as “recombination” alone does not settle whether the result is novel or valuable. Runco’s conceptual distinction needs the richer claim that certain human relations to purpose and problem are part of what the word creativity should preserve.
One can test some component claims. Researchers can compare how systems and people respond to an altered brief, sustain a goal over time, explain revisions or transfer a method to a new domain. Such tasks do not directly measure subjective experience, but they can make process claims more precise. A sweeping assertion that AI can never be creative is harder to establish empirically, because systems and definitions can change. The present paper should be read as a reasoned position about the current concept and technologies, not a time-proof measurement.
What competing accounts say
Margaret Boden’s influential discussion of creativity and artificial intelligence distinguishes forms of novelty and treats computational systems as a way to probe creative mechanisms. Her approach shows why the phrase “AI creativity” can be analytically useful even when machine and human creativity are not assumed identical. One can ask whether a system explores a space, combines elements or transforms rules, then examine the quality of the resulting work.
Gilhooly’s paper on definitions and computability presses the point that the answer depends on the chosen definition. Under an applied, output-focused definition, some computed responses may be externally creative. Under a definition that requires particular internal processes or personal meaning, the same response may fail to qualify. The disagreement is not merely about taste; it identifies different variables that a study must measure. An argument can be strong under one definition and inapplicable under another.
Runco’s reply, as represented by his paper’s thesis, is that a broad output definition gives away too much. If “creative” means only surprising and useful, a search procedure might receive the same conceptual standing as a person’s imaginative engagement with a problem. His concern is partly about explanation and partly about the social consequences of language. Whether that concern outweighs the value of a broader technical term is a live debate, not an established experimental result.
Empirical studies can inform, but not end, the debate
Pennanen and colleagues’ 2025 robot-drawing experiments show that observers gave higher creativity ratings when they could see more of the process and producer, even though the underlying act being rated was held comparable. That finding demonstrates how attribution depends on perceptual evidence. It neither refutes Runco nor proves him right. People can attribute creativity to a robot while a philosopher argues the robot lacks a required process; the disagreement concerns the interpretation of the attribution.
Doshi and Hauser’s 2024 experiment studied generative-AI assistance in short-story writing. AI-generated ideas raised some individual evaluations of creative writing, while stories across assisted participants became more similar. This demonstrates an important two-level tradeoff: a tool can improve a person’s immediate product score while narrowing the diversity of a collective set. It does not show that the AI itself has human creativity, nor that human authors lose all creative agency when using it.
These studies also suggest that authorship is distributed. A human may select a problem, ask for alternatives, reject many suggestions and choose one to develop. The resulting work may be shaped by both person and system. A binary question of “human or AI” may hide that process. Runco’s concern about protecting human creativity becomes more practical when it prompts careful account of which choices were made by whom, rather than a blanket dismissal of assisted work.
Discovery, innovation and creativity are not synonyms
Discovery can mean finding a pattern or possibility that was not previously noticed. Innovation can mean introducing a new product or practice that works in a setting. Creativity can refer to original and valuable ideas, or to a process involving agency, intention and meaning. The boundaries overlap in ordinary speech, but research benefits from making them explicit. A system might discover a molecular candidate without independently deciding which disease deserves attention. A team might turn that candidate into an innovation through validation, manufacturing and care decisions.
Runco’s title asks readers to withhold one label even when the other two apply. This is useful as a conceptual challenge. It warns that a machine’s surprising output may be evidence of effective search, learning and design without evidence of a human-like creative life. It also asks whether the label changes how institutions value the people who frame problems, contribute data, set standards and interpret outcomes.
Yet withholding the label is not a substitute for evaluating the output. A helpful idea still needs testing. An inaccurate one still needs correction. A compelling painting can still affect an audience. The philosophical classification should not be used to avoid questions about quality, provenance, labor, ownership or downstream harm. Those questions remain whether one says “AI creativity” or “AI-generated innovation.”
What counts as originality?
An image can be new to a viewer while resembling many existing works. A phrase can be unusual but useless. Research often treats originality and appropriateness as separate dimensions. The output may also be new within a local task yet familiar across a wider field. Runco’s process account asks what the producer did with possibilities, but product evaluation still needs comparison with the domain and attention to provenance.
Novelty alone is insufficient for most practical decisions. A school should ask whether a student learned and contributed; a journal should ask whether a claim is supported; a museum should ask what a work means within its practice and community. A model’s ability to generate many variations does not automatically meet these tests. A human’s intention, likewise, does not guarantee that a result is good or original. Process and product need separate questions.
Tools can assist with divergent search. A designer might use them to see alternatives and then exercise judgment in selecting, revising and situating a proposal. The value of the tool should be assessed through the team’s actual work. If every designer receives similar suggestions, the tool may improve average local output while reducing variation. This is one reason the Doshi and Hauser finding matters to a discussion that could otherwise stay purely verbal.
Implications for education and institutions
A teacher can explain what an assignment evaluates: a finished artifact, a student’s learning process, a contribution to a group or all three. If a student uses AI, process evidence may include what question they chose, which suggestions they rejected, why a revision improved the result and what they can explain without the tool. This is an Institute design proposal, not a scoring rule validated by Runco’s paper.
An organization can credit human work that otherwise disappears behind an output. Data curation, problem framing, evaluation and correction are creative or enabling activities under many definitions. The question is not solved by assigning every achievement to a model or every achievement to its operator. A useful record describes actual contributions and decision authority. That record can also expose when workers are expected to take responsibility for results they cannot meaningfully control.
Public language should be precise. “The system generated an image that judges rated original” is a narrower claim than “the system is an artist.” “A team used AI to explore solutions” says something different again. Such descriptions leave room for empirical tests and conceptual disagreement. They can also reduce the temptation to use anthropomorphic language as a marketing shortcut.
What the evidence does not show
Runco’s paper does not demonstrate through an experiment that no AI system can ever satisfy any definition of creativity. Its conclusion follows from a process-centered account. Gilhooly and Boden show why other definitions and research programs can yield different conclusions. The existence of a disagreement is not evidence that all definitions are equally useful for every purpose; each must be judged by what it explains.
Pennanen’s creativity ratings do not establish machine intention or consciousness. Doshi and Hauser’s writing results do not establish that AI is either a universal creativity enhancer or a universal threat to human imagination. They study particular tasks, people, interventions and measures. A product score, a diversity measure and a theory of authentic creation are not interchangeable forms of evidence.
The Institute does not need to settle the metaphysics of creativity to make a responsible editorial decision. It can identify who made consequential choices, what the output does, which definition is being used and where evidence ends. The practical value of Runco’s challenge is its demand for precision, not a license to call every machine-assisted work inauthentic.
Return to the surprising image
The designer can keep the image that opened a new path and still ask who identified the problem, selected the direction and decided what the work should mean. Runco would resist calling the system itself creative. A product-focused assessor might use the word for its output. The disagreement becomes clearer when each specifies what it is measuring.
The paper’s durable contribution is to prevent a useful surprise from carrying more philosophical weight than it can bear. Discovery and innovation may be real. Human creativity involves questions of purpose, perception, learning and responsibility that deserve explicit attention. The strongest account of AI-assisted work can describe both the tool’s contribution and the human capacities that guide its use.
Research references
- Runco, M. A. (2024, online). The discovery and innovation of AI does not qualify as creativity. Journal of Cognitive Psychology. View source
- Gilhooly, K. J. (2025, online). Creativity: Definitions and computability. Journal of Cognitive Psychology. View source
- Boden, M. A. (1998). Creativity and artificial intelligence. Artificial Intelligence, 103(1–2), 347–356. View source
- Amabile, T. M. (1982). Social psychology of creativity: A consensual assessment technique. Journal of Personality and Social Psychology, 43(5), 997–1013. View source
- Pennanen, N., Linkola, S., Kantosalo, A., et al. (2025). From product to producer: The impact of perceptual evidence and robot embodiment on the human assessment of AI creativity. ACM Transactions on Human-Robot Interaction, 14(3), Article 41. View source
- Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. View source