Virginia Tech researchers built an online writing environment that recorded every student keystroke and every interaction with an AI assistant while writing an assignment. Among 77 students with unlimited access to ChatGPT, they found that in nearly half of all interactions with the tool, students asked it to write the assignment for them.
Six students did not use ChatGPT at all. The others took a mix of approaches: some used it to generate ideas, while others used it to correct or edit a draft they had written themselves. Some directed the AI by typing prompts and asking it to write and revise content, in a process the researchers call “vibe writing.”
The study was published in the proceedings of the CHI 2026 conference on human factors in computing systems. One of the most notable findings was that students felt the work was their own even when they had delegated much of the process to AI. Those who relied mainly on ChatGPT for planning or revision reported just as strong a sense of ownership as those who did not use the tool at all. Only those who asked ChatGPT to produce the entire assignment consistently reported a significantly lower sense of ownership.
Andrew Jelson, the lead computer science doctoral student on the research, posed the question: “Even if they used ChatGPT to give them all the ideas, they still felt the work was their own. If you aren't creating the ideas yourself, how can you own it?” The researchers stress that this is a question of skills as well as academic integrity. Students who graduate believing they can think through difficult problems because they have “written” dozens of assignments, when they had actually quietly delegated much of the process to AI, may be poorly prepared for the workplace.
In subsequent research, the team examined whether making AI use visible to instructors changes how students use it. One class used the Student Writing with Accountable Generative AI (SWAG) system, developed by the research team. Students were told that their AI interactions would be recorded and could be viewed by instructors.
This change in accountability transformed behavior, according to associate professor of computer science Sang Won Lee. Requests for the language model to write the entire assignment fell significantly. The reduction persisted even when submissions were anonymized so instructors could not see which student had produced each assignment. Lee noted that students appeared to internalize the standard of accountability and that the drop in outsourcing suggested a change in how they viewed the task itself, rather than simply fear of consequences.
The approach shifts attention from simply detecting whether a student used AI to whether they used it in line with instructors' expectations. The version of ChatGPT used in SWAG can change its behavior based on rules instructors set for specific assignments. For example, an instructor can configure the system to refuse requests to write an entire assignment or to ask questions that prompt critical thinking instead of giving a direct answer.
Dan Dunlap, a visiting associate professor of computer science and collaborator on the project, said the goal is to integrate AI use into education and give instructors the ability to guide how students interact with it. The team, which also includes graduate collaborators Daniel Manesh and Alice Jang, as well as Young-Ho Kim of NAVER AI Lab, is now building a version that integrates directly into learning management platforms. Instructors will be able to define acceptable types of AI assistance and receive transparency reports describing students' interactions alongside their work.
The research represents a rare effort to record student-AI collaboration in real time. The ultimate goal is to find ways for instructors to guide the use of AI tools, making the interactions more visible and ultimately more educational, rather than to ban them from the classroom. As Jelson said: “AI isn't going away. So let's integrate it into the curriculum in a way that makes sense.”
Comments