A designer starts with an unfinished sketch. A generative system offers six polished variations within seconds. One variation solves a layout problem the designer had not noticed; another quietly changes the audience, tone and purpose. The interesting question is not whether the tool produced images quickly. It is what the designer can now see, choose, reject and make that would not have been possible before.
Human creativity in the age of generative systems needs a wider account than a contest between a person and a machine. Creative work includes finding a worthwhile problem, exploring alternatives, judging fit, revising in response to others and taking responsibility for a result. A system can contribute to some of those activities. Its fluent output does not settle who set the direction, whose experience the work speaks from or whether a collection of finished works has become more varied.
Separate the product from the process
People often call an output creative when it seems new and useful for a purpose. That is a judgment about a product in a particular context. It does not automatically describe the process that produced it or the agency of the producer. A surprising image can emerge from a random procedure; a careful human exploration can end in an ordinary-looking result. Both cases show why an output score alone cannot answer every question about creativity.
Finnish researchers Pennanen and colleagues studied how observers judged robot-made drawings. Participants saw varying amounts of evidence about the product, the process and the producing robot. In two laboratory experiments, creativity judgments rose when more of that process and producer were visible. The effect concerns human assessment of robot creativity, not proof that a robot had an intention comparable to an artist’s. The distinction matters when people evaluate AI work only by a polished final image.
For a human creator, process also carries information about learning and authorship. The notes rejected on the way to a draft may show a change of mind; the conversation with a client may reveal what problem the work actually solves. An AI image generator may supply an unexpected formal option, while the person decides whether that option respects a community, fits an accessibility requirement or deserves to exist. These contributions are different. Calling both simply “the creative part” obscures the work.
Nor is the person always the lone originator. Creative practices have long depended on materials, tools, collaborators, traditions and audiences. A generative model enters that ecology with unusual speed and breadth of recombination, and with its own limitations. The practical question is how to assign room for exploration, judgment and accountability across the system of people and tools. This is an Institute framing of the field, not a claim that research has found one universal division of labor.
What experiments find when a person gets AI ideas
Lee and Chung asked participants in five experiments to generate ideas for everyday and innovation-related problems, including gifts and designs. Participants given ChatGPT produced ideas that were rated as more creative than those produced without technology or with conventional web search. Their reported advantage was stronger for incrementally new ideas than for radically new ones. That is encouraging for a stuck designer seeking plausible starting points, but the tasks and model version do not establish a general law for every form of art or invention.
Doshi and Hauser studied short story writing. In their experiment, some writers could request one or five ideas from a GPT-4 system, while others wrote without the aid. Separate readers evaluated 293 resulting stories. AI access improved average evaluations of some individual stories, with especially large gains for writers who scored lower on the study’s prior creativity measure. Yet stories in the AI-assisted conditions were more similar to one another. A stronger individual result and a narrower collective range can coexist.
This result is not a verdict that AI always homogenizes culture. The experiment used a specific prompt, a specific version of a model, an eight-sentence story format and a particular recruitment pool. It measured similarity within that collection, not a society’s future literature. It does, however, identify a plausible coordination problem: if many creators draw from similar suggestions, the same attractive starting points may recur even when each person’s first draft improves.
A later analysis by Meincke, Nave and Terwiesch of brainstorming data around Lee and Chung’s experiments sharpened the same distinction between the quality of individual ideas and diversity across a pool of ideas. Their contribution was a Matters Arising article, not a stand-alone general population experiment. Together with the original work, it cautions against declaring a brainstorming session successful solely because the mean rating of each suggestion rises. A group also needs options that differ enough to reveal an unexpected path.
The right measure depends on the purpose. A clinic designing a patient letter may value clarity and safety more than radical novelty. An arts collective asking what a neighborhood could become may need a broad range of voices, languages and aesthetic choices. An evaluation should say whether it is rating individual outputs, the best idea in a set, the set’s diversity, the creator’s learning or the audience’s response. One number cannot stand in for all five.
The interface can make a person an editor or a co-creator
When a tool presents a complete, confident draft, it creates an easy role for the person: accept, polish or correct. That may be exactly what is wanted for routine work. It can be a poor invitation to explore a question whose terms are still open. The timing of AI contribution changes the task. Asking for a finished poem before writing a line is unlike asking for an alternative image after building an initial concept.
McGuire, De Cremer and Van de Cruys tested this in poetry-writing interfaces. In their first experiment, participants editing a generated poem were judged less creative by professional poets than people writing without the AI. Their second experiment redesigned the interaction so a person selected topical words and wrote the first line, then built the poem with subsequent AI suggestions. The deficit in expert ratings did not appear in the same way, and reported creative self-efficacy was part of their explanatory analysis.
That finding does not make a line-by-line poetry interface the right tool for architecture, music or public-service design. It shows that “AI assistance” is not one treatment. A system that arrives after the person sets a direction may support a different activity from one that replaces the initial act. Interface studies should describe what the human and model actually did, rather than comparing vague labels such as “with AI” and “without AI.”
A practical creative workflow can preserve an unassisted first pass when that pass is valuable. The designer names a question, sketches several directions and records why each might matter before requesting machine alternatives. They can ask the system for contrasts, counterexamples or constraints it has neglected, rather than another polished version of the same answer. They then test the suggestions against the brief, intended audience and material realities. These are design hypotheses informed by the studies, not a validated recipe for universal creative gain.
Sometimes the initial pass should be assisted. Someone working in a second language may need a phrase to begin; a person with limited mobility may use voice and image tools to make an idea visible. Preserving human direction does not require withholding assistance. It requires that the person can set or revise the purpose and can recognize where the tool’s proposal is only a proposal.
Ideas belong to a social setting
Creative value is rarely determined by novelty alone. A public campaign can be visually striking and still misrepresent the people it depicts. A product concept can be efficient and still exclude people who use assistive technology. A poem may be formally familiar but deeply resonant in its community. The creator’s judgment connects an output to a relationship, place and use; a model’s suggestion needs that contextual work.
Kumar, Tu and Zallio reviewed 50 papers and preprints on generative AI in interface and experience design. Their review found substantial use in prototyping and visual asset production, but more limited treatment of early conceptualization and UX evaluation. It also identified fragmented methods and thin treatment of bias, transparency and privacy. Because the review includes preprints and focuses on one design field, it should guide questions about practice rather than serve as a census of all creative industries.
For a Nordic public-service design team, the review suggests a concrete inquiry. If an AI tool makes interface mockups easier to produce, does it also help residents describe the service problem? Can people using different languages or access methods evaluate the alternatives? Who decides whether a design reflects local needs? These questions move the evaluation from output speed to the quality of participation and testing. They are Institute applications, not results directly measured in the review.
Teams can protect diversity by gathering independent human ideas before showing a shared AI proposal. They can deliberately compare multiple prompts, datasets, languages and perspectives, and preserve outlying suggestions long enough to test them. The point is not to manufacture arbitrary difference. It is to avoid mistaking a polished consensus for a well-explored problem. If the task is collaborative, the distribution of ideas across participants matters as much as the quality of any one person’s page.
This also changes what counts as credit. A creative project may include a person who framed the problem, a community that supplied knowledge, a designer who selected and transformed outputs, and a software service that generated candidate forms. Transparent records can describe those contributions without pretending the machine has the same obligations as a person. The appropriate attribution rules vary by context and jurisdiction; the cited psychology and design studies do not settle copyright or contractual rights.
Capability includes the ability to disagree with a fluent output
Generative systems can produce convincing language and images before a novice knows how to assess them. This creates a temptation to measure success by how quickly something publishable appears. Yet a creator who cannot explain the intended audience, identify a borrowed assumption or correct an inaccurate detail remains dependent on the tool’s first framing. Fast production and developed judgment are separate outcomes.
In education, one test is whether the learner can generate and defend an alternative after using the tool. A student might compare their own first sketch with an AI suggestion, describe what each solves, and make a third version from that comparison. The teacher can assess the reasoning and revision, not only the surface finish. This does not require a ban on tools. It requires an assignment in which the student’s decisions remain observable and open to feedback.
In professional settings, the same principle may involve a creative brief, version history and an explicit review stage. The person can record why a model option was selected, what evidence about users supports it, and what was changed after a critique. Such documentation can become bureaucratic if demanded for every minor variation. The level of record should match the stakes: an internal mood board needs less than a public campaign that affects real people.
A person can also decide that a system is the wrong collaborator for a particular task. Private stories, culturally specific work or unshared drafts may call for a different tool or no model at all. This is not a romantic claim that unassisted work is inherently more creative. It recognizes that creative practice involves relationships and permissions as well as outputs. A genuine option to abstain helps a team discover which parts of the process need care that the tool does not provide.
Evaluate the whole creative system
A useful evaluation compares more than one finished artifact. It can include an unaided condition, a search condition, an AI idea condition and an iterative co-creation condition. It should describe the model, prompt access, time allowed and who could edit what. Independent evaluators may rate usefulness and novelty; participants may report agency and learning; domain experts may check constraints; a team may measure diversity across outputs. No single measure captures all of these dimensions.
Timing matters. A short experiment can show what happens on today’s task, but not whether a creator becomes more skilled after months of use. A team can revisit the same people later with a changed brief, ask them to work without assistance and examine whether they can still choose and justify a direction. It should also check who benefits. A novice may gain a starting point; an experienced designer may find suggestions repetitive; a minority-language creator may encounter narrower coverage. These are questions for study, not assumptions about every user.
Better average output can coexist with lower ownership. Conversely, a slow, difficult process may teach little if the task is needlessly obscure. Evaluation should ask what the tool enabled the person to express, not merely how much effort it removed. The aim is a creative ecology in which more people can contribute with understanding and where the resulting work remains varied enough to respond to different lives.
What the evidence does not show
The cited experiments do not prove that generative AI makes all people more creative or less creative. They use defined tasks, model versions, interfaces and rating methods. Short stories, brainstorming prompts, poetry and robot drawings are not interchangeable measures. A rating of novelty is not a measure of truth, social value or long-term skill. A higher score for one output does not establish that a whole culture gains expressive diversity.
The Aalto robot study investigates observers’ judgments, not the subjective experience of an AI system. The design-workflow review maps a developing literature and includes preprints, so it cannot establish causal effects of every tool in practice. Lee and Chung’s outcome evidence concerns specified idea-generation tasks, while the diversity analysis highlights a different outcome; the two can be jointly informative without one invalidating the other. The poetry experiments show that interaction design matters in that task, not that every creator must begin unaided.
No study here establishes who deserves legal authorship, what compensation is fair for training data or how a particular organization should handle confidential material. Those questions require additional sources and decisions. The Institute’s proposed practice is therefore modest: retain a human purpose, document consequential choices, compare individual quality with collective range, invite affected people into evaluation and preserve the ability to reject a machine suggestion. This is a synthesis for responsible work, not a validated creativity score.
Return to the unfinished sketch
The designer looks again at the six variations. One reveals a better way to organize the page. Two share an attractive but unsuitable assumption about the audience. The rest are useful mainly because they show what the project is not. The next step is neither to accept the cleverest image nor to refuse all assistance. It is to make a more deliberate sketch that responds to the people and purpose at hand.
Human creativity can grow with generative systems when people have room to frame the work, encounter genuinely different possibilities, develop judgment and remain answerable for what they make. The research also gives reason to watch for convergence, dependence on finished suggestions and loss of process visibility. The value of the tool is not only that it produces another candidate. It is whether a person or a group can now imagine, question and realize something worth making.
Research references
- Lee, B. C., & Chung, J. (2024). An empirical investigation of the impact of ChatGPT on creativity. Nature Human Behaviour, 8, 1906–1914. 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
- Meincke, L., Nave, G., & Terwiesch, C. (2025). ChatGPT decreases idea diversity in brainstorming. Nature Human Behaviour, 9, 1107–1109. View source
- McGuire, J., De Cremer, D., & Van de Cruys, T. (2024). Establishing the importance of co-creation and self-efficacy in creative collaboration with artificial intelligence. Scientific Reports, 14, 18525. View source
- Pennanen, N., Linkola, S., Kantosalo, A., Hiillos, N., Männistö, T., & Guckelsberger, C. (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
- Kumar, T., Tu, X., & Zallio, M. (2025). How generative AI is reshaping UI/UX design workflows: A systematic review. Human Factors in Design, Engineering, and Computing, 2438–2452. View source