A study finds human agency fades fast, unless renewed at every GenAI prompt. #
Posted July 26, 2026 [ Reviewed by Jessica Schrader
](/us/docs/editorial-process)
Key points
- Agency in AI collaboration must be sustained throughout a task, not just established at the start.
- Students who thought first before using AI still drifted toward following AI by the third prompt.
- Requiring an original idea at every prompt kept students directing the AI, not the other way around.
Co-authored by Ava Gilmartin, Paul Surlis, and Michael Hogan.
A university student opens ChatGPT minutes after receiving a creative problem-solving task. Within seconds, the AI generates several polished solutions. The student selects one, makes a few small changes, and submits the final response. Although the work appears sophisticated, something vital has been lost: the student’s independent idea generation. The AI completes much of the cognitive work before the learning process has even begun. While generative artificial intelligence (GenAI) has enormous potential to support learning, this scenario highlights an important challenge facing higher education. The question is no longer whether students should use AI, but rather how AI can support learning without replacing human reasoning.
Guo et al.'s (2025) study, Student–AI Creative Problem-Solving: The Role of Human Agency, addresses this challenge by examining how the integration of human ideas can affect the sustained creativity of student-AI
collaboration. Rather than treating AI as an answer generator, the researchers investigated whether requiring students to contribute their own ideas throughout the collaboration produced better educational outcomes. Their findings suggest that AI enhances learning only when instructional design keeps students cognitively engaged at every stage of the problem-solving process.
The subtle nature of pedagogical design #
Guo et al.'s study is particularly valuable because it does not simply compare AI users with non-AI users. Instead, it investigates how different patterns of collaboration shape students' creative thinking. Across two quantitative experiments involving 346 university students, participants completed the same creative task under different AI collaboration conditions: developing innovative ways to improve a toy bunny to increase sales. In the first experiment including 170 participants, one group generated their own ideas before consulting ChatGPT (pre-SHT), another group consulted AI first and then added their own ideas afterwards (post-SHT), while a control group relied on AI throughout the task without any instruction to engage in independent thinking (without-SHT). Students who generated their own ideas before using AI reported greater initiative, stronger ownership of their work, and invested more mental effort than those who immediately turned to AI. This pre-SHT group also produced more cognitively sophisticated prompts that reflected higher-order thinking rather than simply repeating information from the task. However, this advantage gradually disappeared as students continued collaborating with AI. By the third prompt, many participants had begun following the AI's suggestions instead of directing the collaboration themselves. This finding is very interesting: simply delaying AI use encouraged independent thinking initially, but it was insufficient to sustain human agency throughout the task.
Recognising this limitation, Guo et al. redesigned the collaboration in a second experiment involving 176 students. Instead of asking participants to think independently only at the beginning, they introduced a deep idea integration (DII) condition, which required students to integrate their own ideas into every interaction with ChatGPT. Rather than asking broad questions and accepting AI-generated solutions, participants first generated an original idea before using AI to refine, challenge, or extend it. For example, instead of asking ChatGPT how to improve the toy bunny, a student who had already proposed making the toy glow safely in the dark might ask which child-safe materials could achieve this or whether similar products already existed. In this condition, students, not the AI, determined the direction of the collaboration.
The learning benefits of the DII constraint in Experiment 2 are notable in comparison to the effects observed in Experiment 1. Students maintained higher levels of perceived agency throughout the task, produced prompts that remained cognitively sophisticated across all stages of the collaboration, and generated final solutions that were significantly more novel and useful. Their final design solutions were also less similar to one another, suggesting that students were developing genuinely original ideas rather than converging on the AI's preferred solutions. Importantly, participants also reported greater perceived cognitive improvement, indicating that the collaboration enhanced their own learning experience rather than simply improving the quality of the final output.
Designing for agency #
Similar findings can be observed elsewhere in the literature. For example, in relation to the homogeneity of idea generation in working groups, Xia et al. (2026) found that, while GenAI-supported group work assisted learning and metacognitive engagement in working groups, the groups working *without *GenAI in their experimental study produced more original creative thinking solutions. Students also report these consequences directly. Notably, Perifanou and Economides (2025) found that students valued GenAI most when it was used to clarify difficult concepts rather than replace independent learning. Together, these findings reinforce Guo et al.'s central conclusion: AI supports learning most effectively when educational design requires students to remain active contributors rather than passive recipients throughout the learning process.
Guo et al.’s findings also highlight the importance of understanding what is meant by human agency in AI-supported learning. In this context, agency refers to students’ ability to intentionally direct their own learning by contributing to and maintaining ownership over the problem-solving process. However, it is important to recognise that Guo and colleagues measured agency primarily through students’ self-reported perceptions, which may not capture the complexity of how agency operates and develops during learning. Despite this limitation, the findings align with broader educational theories that position agency as central to effective learning. The implication for education is therefore not to completely suppress the use of AI in learning environments, but to design interactions that preserve students’ intellectual independence, curiosity, engagement, and understanding of how AI can influence learning. Generative AI can act as a powerful learning partner when students use it to refine and expand their own ideas rather than outsource the learning process. As AI becomes increasingly embedded within higher education, educators will need to develop assessment and teaching strategies that reward the learning process as much as the final product. Encouraging students to critically question and build upon AI-generated suggestions will help ensure that AI can strengthen, rather than diminish, the intellectual skills and the human agency of students.