A new pilot study from Kennesaw State University challenges the assumption that college students simply hand over their writing assignments to artificial intelligence. Rather than analyzing finished papers or relying on what students report about their habits, researchers employed think-aloud protocols to observe the actual writing process as it happened. This approach captured something earlier studies may have missed: the moment-by-moment decisions students make when working alongside AI tools.
The research involved 20 undergraduates who completed a 20-minute writing task arguing a position on the value of striving for perfection. Participants were explicitly told to use generative AI exactly as they normally would and that there were no correct or incorrect ways to incorporate these tools. To minimize the Hawthorne effect, where observation changes behavior, researchers set a timer and left the room during writing sessions. Students' screen activity and verbal commentary were recorded for later analysis, allowing the research team to examine their choices without interfering with the writing process itself.
Several consistent patterns emerged from analyzing the transcripts and screen recordings. Many students initiated their writing by using AI to generate preliminary ideas or draft thesis statements. However, this represented the beginning rather than the end of their engagement. One student characterized the AI output as a starting point rather than a final product. The generated text helped overcome initial paralysis when facing a blank page, functioning more as a brainstorming companion than as an author substitute.
The study also revealed that students rarely accepted AI-generated text verbatim. Instead, they actively edited and reshaped the language. One participant described a cyclical process where AI rewrites the student's prompt, then the student rewrites the AI's output, ultimately claiming authorship of the result. Another student explicitly redirected the AI when its response failed to address the assignment requirements. These interactions suggest students view AI as something closer to a collaborative sparring partner than a ghostwriter.
Perhaps most tellingly, some students rejected AI suggestions entirely. Multiple participants articulated decisions not to use AI for particular aspects of their work, especially research. Others abandoned AI-generated text when it felt too generic or misaligned with their developing arguments. This selectivity demonstrates that students establish boundaries around where AI fits into their writing process.
Students frequently turned to AI during moments of struggle or uncertainty, with one acknowledging heavy AI use when experiencing difficulty. Even in these cases, however, the pattern showed students using AI as supportive scaffolding rather than directly copying its output.
The findings suggest AI enters student writing as part of a negotiated collaboration rather than replacing human authorship entirely. Students appear to maintain control over argument selection, voice, and final language while using AI primarily for idea generation, revision, and overcoming blocks.
The researchers appropriately caution against overinterpreting these preliminary results given the small sample of 20 participants. The team is expanding to 100 participants to test whether these patterns hold at larger scale. The expanded study will also examine how neurodivergent writers interact with generative AI, an area currently lacking research attention.
These findings have practical implications for educators designing assignments and policies. Understanding that AI use involves active negotiation during the writing process itself, rather than simple substitution, may help keep human writers central to the composition process.