Will generative AI enhance human creativity? As tools such as ChatGPT, Claude, and Midjourney become part of everyday life, this question is no longer a distant debate about the future. AI is already entering creative processes in writing, design, advertising, music, video, and education. Yet an important question remains: does human creativity automatically increase simply because AI can produce creative outputs?
McGuire, De Cremer, and Van de Cruys’ 2024 paper, Establishing the Importance of Co-Creation and Self-Efficacy in Creative Collaboration With Artificial Intelligence, offers a nuanced answer. Its conclusion is straightforward. Creating with AI does not always enhance creativity. In fact, when AI first presents an output and humans are left merely to revise it as “editors,” creativity may decrease. By contrast, when humans take the lead and create with AI in an alternating, interactive process, the creative benefits of AI collaboration become more visible.
The importance of this paper lies in how it shifts the focus of the generative AI debate. Much of the discussion so far has centered on how creative AI models are. Can AI write poems like humans? Can it create stories? Can it generate images? This study changes the question. What matters is not how creative AI is on its own, but what kind of relationship humans have with AI during the creative process.
The researchers focused on the role humans play in human-AI creative collaboration. Does the human become an “editor” who receives and revises an AI-generated draft? Or does the human become a “co-creator” who builds the work step by step with AI? According to the paper, this difference changes the outcome. It is not merely an interface difference. It directly affects how humans perceive themselves during creation and how much they believe in their own creative ability — in other words, their creative self-efficacy.
The creative task in the experiment was poetry writing. Poetry is one of the representative forms of human creativity. Across two experiments, the researchers compared human-only creation, human-AI collaboration, AI-only creation, and co-creative human-AI collaboration. Importantly, the evaluation of creativity did not rely only on the participants’ self-assessments. Ten professional poets anonymously evaluated the creativity of the poems, using the Consensual Assessment Technique, a widely used method in creativity research. This is a major methodological strength of the study.
The first study recruited 101 participants. After excluding those who failed attention checks, 96 participants were included in the final analysis. Participants were assigned to one of two conditions. One group wrote an eight-line poem without assistance. The other group received a poem generated by AI and was allowed to freely revise it before submitting the final version. The researchers also added 50 AI-generated poems as an AI-only comparison group.
The AI collaboration interface was not overly simplistic. Participants could directly edit each line of the AI-generated poem and could also select alternative lines from a drop-down menu. When a specific line was changed, other candidate lines were updated accordingly. In other words, the researchers did not design a weak or superficial human-AI collaboration condition. Participants were given a fairly high level of editing freedom.
Even so, the results were unexpected. Poems written by humans alone were rated as more creative than poems produced in the human-AI collaboration condition. In the professional poets’ creativity ratings, the human-only condition received an average score of 18.23, the human-AI condition received 12.55, and the AI-only condition received 9.45. The human-AI condition performed better than AI alone, but it fell far short of human-only creation. Participants’ self-ratings showed a similar pattern: those in the human-only condition rated their poems as more creative than those in the human-AI condition.
This finding offers an important warning. The mere presence of AI in the creative process does not automatically augment human creativity. When AI provides the first output and humans revise it afterward, humans may become constrained by the AI’s initial draft. Psychologically, this can be understood as an anchoring effect. The first draft becomes the reference point, and people may fail to move far enough beyond it. As a result, the human becomes not an active creator, but a passive reviser.
The researchers interpret this as a limitation of the “editor role.” An editor is essentially someone who evaluates and modifies something someone else has created. When AI produces the creative work first and humans revise it, AI captures the initial direction of creation. The human is placed in a reactive position. In that situation, people may feel less like the true agents of creation. Creativity often emerges from free exploration and self-confidence, but the editor role can narrow those possibilities.
The second study examined whether this limitation could be overcome. It included 152 participants and divided them into three conditions. The first was a human-only creation condition. The second was a human-AI editing condition similar to the first study, in which AI generated a poem first and humans edited it. The third was a newly designed co-creative human-AI condition.
The key feature of the co-creative condition was that the human began the creative process. Participants first selected a set of three words to serve as the theme of the poem, such as “sorrow, longing, admiration.” Then the participant wrote the first line, AI generated the second line, the participant wrote the third line, and AI generated the fourth. In this way, the poem was completed through alternating turns between the human and the AI. Participants could also revise AI-generated lines during the process.
This structure differed from the editor model used in the first study. The human was not merely revising a completed AI output after the fact. The human selected the theme, wrote the opening line, responded to the AI’s contribution, and continued the work. Creation did not begin from a fixed draft. It unfolded more like a dialogue between the human and AI. The researchers expected this structure to increase the human’s creative self-efficacy.
The results aligned with that expectation. Creative self-efficacy was lowest in the human-AI editing condition. The human-only condition had an average score of 4.71, the co-creative human-AI condition scored 4.62, and the standard human-AI editing condition scored 3.74. The co-creative condition did not differ statistically from the human-only condition, but it was significantly higher than the editor-style human-AI condition.
Self-rated creativity showed a similar pattern. The human-only condition averaged 23.29, the co-creative human-AI condition averaged 21.16, and the editor-style human-AI condition averaged 16.96. Again, the co-creative condition did not significantly differ from the human-only condition, but it scored higher than the editor-style condition.
The professional poets’ evaluations confirmed the same direction. The human-only condition received an average score of 15.74, the co-creative human-AI condition received 14.70, and the editor-style human-AI condition received 12.53. The co-creative human-AI condition did not significantly differ from human-only creation, but it was significantly higher than the editor-style human-AI condition. In other words, the creativity deficit seen in AI collaboration disappeared when humans were positioned as co-creators.
The core concept of the paper is creative self-efficacy. Creative self-efficacy refers to the belief that one can produce creative outcomes. When creators have this belief, they are more likely to generate ideas actively, experiment, tolerate failure, and attempt unfamiliar forms of expression. When people feel that they are not creative, they are more likely to remain with safe choices and struggle to move beyond existing frames.
Through mediation analysis, the researchers also found that the effect of the co-creator role on creativity ratings was explained through creative self-efficacy. Compared with the editor-style human-AI condition, the co-creative condition increased creative self-efficacy, and that increased self-efficacy led to higher expert creativity ratings. This means the results were not simply due to a more convenient interface. The crucial factor was the psychological process through which humans felt themselves to be creative agents.
At this point, the paper challenges a common assumption about generative AI use. Many companies and technology leaders argue that productivity will increase if AI creates a draft and humans make the final judgment. Many AI tools are designed in exactly this way. AI first produces text, images, code, or plans, and humans select or revise them. But this study shows that for creative work, this may not be the best approach.
Creativity is not simply the ability to choose a good output. Creativity emerges through setting direction, opening possibilities, going through trial and error, and assigning meaning oneself. If AI presents a completed output from the start, humans may think within that output’s frame. By contrast, when humans define the direction first and AI responds to that flow, humans can maintain a stronger sense of creative control.
The paper has important implications for human-AI interface design. Generative AI tools should not focus only on providing better drafts faster. They should be designed so that users can set the creative direction, intervene during the process, interact with AI, and maintain the feeling that they are creating. A good AI creative tool should not replace the human. It should help the human believe more strongly in their own creative ability.
The implications for education are also significant. If students use AI only by receiving and editing AI-generated text, they may struggle to become the true subjects of writing. But if students choose the topic, write the first sentence, critically respond to AI suggestions, and continue in their own words, AI can become a tool that strengthens creative self-efficacy. The central issue in AI education is not simply whether AI use should be banned or allowed. It is what kind of role structure is designed.
The same applies to creative work in companies. In advertising, design, content planning, and brand storytelling, asking AI to produce the first complete draft and having humans revise it may be efficient, but it can weaken creativity. The stronger the initial draft is, the more likely humans may become trapped within it. If creative performance is the goal, it may be better to design systems in which humans first establish the direction and AI responds conversationally, rather than systems in which AI provides the answer from the beginning.
The paper also offers a different way to approach the common question of whether AI will replace humans. The key issue is not replacement itself, but role allocation. If humans are pushed into the role of editors, AI occupies the center of creation. But if humans remain co-creators, AI can become a partner that expands human thought. The same AI can suppress or support human creativity depending on the interface and task structure.
The paper has limitations. First, the study was conducted in a laboratory environment and cannot fully reflect the complexity of real creative work. In reality, creativity is shaped by deadlines, teamwork, organizational culture, market demands, and client feedback. Second, the creative task was limited to poetry writing. Poetry is a suitable genre for creativity research, but more studies are needed to determine whether the same pattern appears in design, video, music, coding, and planning documents. Third, the study did not sufficiently distinguish participants’ expertise levels. Beginners and professional creators may respond differently to AI-generated drafts.
Another limitation concerns the AI system used in the study. The poetry generation system was professionally designed, but it may differ from current general-purpose generative AI systems such as GPT-4o or Claude 3.5. Newer AI systems can produce more fluent and more diverse outputs. However, the researchers suggest that even as AI systems improve, the importance of the co-creator role is likely to remain. In fact, the stronger AI becomes, the more important it may be to design systems that prevent humans from being pushed into passive editing roles.
The paper makes several important theoretical contributions to AI co-creation research. First, it explains the effects of AI collaboration not only through output quality, but through the psychological mechanism of human self-efficacy. Second, it does not treat human-AI collaboration as a single form, but distinguishes between editor-style collaboration and co-creator-style collaboration. Third, it improves evaluation validity by using professional poets to assess creativity. Fourth, it empirically demonstrates that the human role experience in AI tool design can determine creative outcomes.
Ultimately, the message of the paper is clear. Generative AI does not automatically augment human creativity. A structure in which AI first produces an output and humans revise it may be efficient, but it can lower creative self-efficacy and weaken creative results. By contrast, a structure in which humans set the direction and build step by step with AI can preserve and strengthen human creativity.
In AI-era creativity, the most important question is not only how intelligent AI is. It is where humans are placed in the process. Will humans remain editors, or will they stand as co-creators? That difference may determine the difference in creativity.
The conditions under which AI supports creativity are becoming clearer. Humans must speak first, AI must respond, and humans must take the direction again. Creation is not merely the act of selecting an output. It is the process of opening possibilities. If generative AI is to become a true tool for creativity, it must return creative agency to humans.
