Creative work usually develops in stages. You start with an idea, then build it, cut parts, rearrange it, question it, test it, and return to it.
Generative AI can enter that process in several roles. It might give you a finished artifact to edit, offer possibilities to develop, or help you test an idea you already have.
Recent research shows why those roles matter. Several experiments found creative gains with AI assistance. Other results changed when the studied interaction changed and when people kept different roles inside the creative process.
That leads to a useful question: What part of the creating are you still doing?
AI can improve creative output in some tasks
In 2026, Feng Zhu and Wenbo Zou published a pre-registered experiment, Generative AI adoption in human creative tasks: Experimental evidence, in the Journal of Economic Behavior & Organization. They recruited 302 university students to complete creative picture-idea tasks. A randomly selected group had access to ChatGPT.
Students with ChatGPT access produced ideas that received higher creativity ratings.
The researchers also compared the human-AI work with ideas produced by ChatGPT on its own. They found no evidence that combining a person with ChatGPT automatically created an extra creative advantage beyond the tool's own performance.
Those findings belong to one set of picture-idea tasks with university students. They still give us an important distinction. The researchers asked two separate questions: Did ChatGPT access improve the participant's score? Did the human-AI pair add something beyond ChatGPT alone?
Anil Doshi and Oliver Hauser found a similar reason to separate those questions in a 2024 Science Advances experiment, Generative AI enhances individual creativity but reduces the collective diversity of novel content, on short-story writing.
Some participants received story ideas from a large language model. Those writers produced stories that were rated as more creative, better written, and more enjoyable. The gains were largest among people who had scored lower on a baseline creativity measure.
At the same time, the AI-assisted stories were more similar to one another than the stories written without AI ideas.
One person can therefore produce a stronger result in an assisted task while a group of people produces a narrower range of results. Both outcomes can be true at once.
Better output and wider exploration are different goals
Creative work needs material to choose from. You may start with one idea, generate several variations, combine parts of two, discard three, and discover a new direction while comparing them.
Adam Grant makes a related point in Originals. His popular explanation is that producing more ideas creates more variety, which gives original ideas more chances to emerge.
That idea is useful here as a touchstone. The research on AI carries the evidence about AI. Grant's point helps explain why variety matters inside a creative process.
The Doshi and Hauser study shows a tension worth noticing. AI ideas helped individual writers produce stronger stories, while the set of AI-assisted stories became more alike.
Picture a communication problem with several reasonable ways to approach it. You could begin with reassurance, a clear explanation of what changed, a direct answer to a likely question, or a short scene that shows why the issue matters.
For my own practice, the finding that AI-assisted stories became more similar is a reason to keep exploring several directions before choosing one. A useful first suggestion can be one option among many. The search can continue after that first option appears.
For creative work, the useful question is broader than whether AI gave you a good idea. It also includes how many directions you considered, what you added yourself, and what you chose after seeing the options.
The role you keep changes the process
Jack McGuire, David De Cremer, and Tim Van de Cruys studied this question through poetry writing. Their 2024 paper, Establishing the importance of co-creation and self-efficacy in creative collaboration with artificial intelligence, in Scientific Reports, reported two experiments using purpose-built human-AI interfaces.
In the first study, people writing a poem on their own produced work rated as more creative than people who first received an AI-generated poem and then edited it.
The second study changed the interaction. Participants worked through a co-creation design that gave them a more active role in developing the poem with AI. Under that design, co-creators no longer showed the lower creativity seen in the editor condition.
The study examined poetry writing under two experimental designs. Within those tasks, receiving an artifact to edit and participating as a co-creator produced different creative outcomes.
That distinction matches my own creative process.
My creative writing is iterative. I build an idea, remove parts, rearrange it, ask questions, look for evidence, and test whether two pieces really belong together.
A sounding board can help with that work. I might bring an idea for an article and ask for several ways to organize it. I may reject all of them and keep one question the exchange surfaced.
Sometimes I ask what perspective is missing. I may ask whether research supports an idea I am developing. I may test three transitions and realize that the real problem sits earlier in the argument.
The value comes from continuing to create inside the exchange. I bring purpose and context. I choose what deserves another look. I decide which possibilities change the work and which ones disappear.
What happens after assistance matters too
A 2026 study, Think First, ChatGPT Later: Guiding Human-AI Collaboration for Learning Gains in Independent Human Creativity, by Sarah Shi Hui Wong and Sophia Xuefei Qiu, in Educational Psychology Review, adds another useful distinction.
The researchers worked with 196 university students on creative product tasks. One group worked without AI. One group used ChatGPT freely. A third group first generated its own ideas, then used ChatGPT to improve, develop, and evaluate those ideas.
On the first task, the unrestricted ChatGPT group produced the most creative solutions.
Then every group completed a new creative task without AI.
On that later independent task, the guided group produced more creative solutions than both the unrestricted-AI group and the human-only group. The unrestricted group returned to a level similar to the human-only group.
The study used university students and two bounded product-creation tasks. It gives us one concrete example of how the structure of an AI interaction can shape creative performance.
It also shows why assisted performance and later independent performance deserve separate attention. A tool can help produce a strong result during one activity. We can also ask what creative work the person is practising while that result is being produced.
That second question matters when creative work is part of your professional craft and you will return to it again.
What this can look like in K-12 communication work
The studies above examined university students, adult writers, poetry, short fiction, product ideas, and other bounded creative tasks. The examples that follow apply their underlying distinctions to professional K-12 communication.
Suppose you need a fresh way to explain a complicated change.
You can begin by deciding what people need to understand and which facts carry the explanation. Then you can use AI to explore several structures. You compare them, combine useful parts, and choose the sequence that fits the situation.
Or suppose a draft feels predictable.
You can ask which perspective the draft may be missing. You can ask for questions from a recipient who sees the issue differently. You can ask for several possible openings, then decide that the strongest opening is one you write yourself after seeing what the options exposed.
The same pattern can apply when an idea is still forming. You may ask for research related to the idea, read the source, and discover that your first claim was too broad. You revise the idea because the evidence changed your thinking.
You can also use AI to test connections. If two sections feel separate, ask what relationship the tool sees between them. The suggestion may reveal a useful bridge, or it may show that separating the sections produces a clearer piece.
In each case, AI adds material to examine. You remain active in generating, selecting, combining, rejecting, checking, and deciding.
That is a different creative experience from receiving a finished artifact and polishing it.
Look at the creative work you kept doing
Look at one recent piece of creative work where AI helped.
Ask yourself three questions:
- What idea did I bring?
- What did the exchange add?
- What did I reject, combine, rearrange, or decide?
The answers show where you were still creating inside the process.
Across these studies, creative results changed with the task, the participant, and the design of the AI interaction. One recurring question is what role the person keeps in producing and developing ideas.
AI can give a creative process more material to work with. You give that material direction, meaning, and final shape.
For my own practice, that is the standard I want: more possibilities to think with, with the creative work still mine.
Sources
- Wong & Qiu (2026), Think First, ChatGPT Later: Guiding Human-AI Collaboration for Learning Gains in Independent Human Creativity
- Doshi & Hauser (2024), Generative AI enhances individual creativity but reduces the collective diversity of novel content
- McGuire, De Cremer & Van de Cruys (2024), Establishing the importance of co-creation and self-efficacy in creative collaboration with artificial intelligence
- Zhu & Zou (2026), Generative AI adoption in human creative tasks: Experimental evidence
- Adam Grant, Originals discussion, Knowledge at Wharton