Generative artificial intelligence tools like ChatGPT can draft an essay for students in minutes. But students also use them to brainstorm, revise, proofread, and more, leaving educators wondering what was written by students and what by AI.

A Virginia Tech–led research team headed by computer science associate professor Sang Won Lee decided to find out what students actually do when given unrestricted access to ChatGPT while writing.

He and his team created an online writing environment that logged students' interactions with an AI assistant while writing an assigned essay: every question they asked the tool, every response they received, and every time text was copied and pasted into their essays. Students were told they could use AI freely or not at all. 

Among 77 students in the study, six wrote without using ChatGPT at all. The rest used a mix of approaches, from generating ideas to proofreading drafts.

  • In nearly half of all AI interactions, students asked the tool to write the essay for them.
  • Some students steered the AI by typing prompts, directing the tool to write and revise the content in a process called "vibe writing.”
  • Many used ChatGPT more narrowly to brainstorm topics or to proofread and polish a draft they had written themselves.

The study, published in the Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, found that students tended to feel ownership of their work even when they delegated much of the writing process to AI. In later research, the team also discovered that students depend less on AI in their writing when they know their interactions with ChatGPT are being recorded.

The results offer a window into how students approach writing with AI that could help instructors think more about which kinds of AI assistance align with their learning goals — and which might compromise student learning.

Using AI can still feel like writing your own essay

One of the team’s notable findings involved how students felt about the essays they submitted.

Students who primarily relied on ChatGPT for planning or revising reported feeling just as much ownership over their essays as students who never used AI at all. The only students that consistently reported significantly lower ownership were the ones who asked ChatGPT to generate the entire essay for them.

"Even if they used ChatGPT to give them all of their ideas, they still felt ownership of the essay," said Andrew Jelson, the lead computer science Ph.D. student on the research. "If you don't generate the ideas, how can you own it?"

This isn't just an academic integrity question, according to the researchers. It's a skills question. Students who graduate from college believing they can think through hard problems because they've "written" dozens of papers, but who have been quietly outsourcing that process to AI, may be underprepared for the workplace.

Above: Computer science associate professor Sang Won Lee demonstrates the artificial intelligence platform his team built for recording students' use of generative AI for essay writing. Video by Tonia Moxley for Virginia Tech.

Transparency changes student behavior

In subsequent research, the team asked a related question: Does making AI use visible to instructors change how students use it?

In a classroom deployment separate from the published study, students completed assignments using Student Writing with Accountable Generative AI (SWAG), a system developed by the research team. They were told that their AI interactions would be recorded and could be seen by their instructor. 

That single shift in accountability changed how students approached the work, Lee said. 

Requests for the large language model to write an essay on a student's behalf dropped significantly. This reduction persisted even when submissions were anonymized so instructors couldn't see which student produced a particular essay or project. 

“This suggests that the students were internalizing the accountability norm,” Lee said. “The drop in outsourcing behavior points to a shift in how students saw the task itself, not just a fear of consequences. This might be the first experience for a lot of students to use generative AI in an accountable manner. They're practicing something they're going to need in the real world."

That approach shifts the focus from simply detecting whether a student used AI to understand whether the student used the tool according to expectations set by instructors. Students were more likely to comply with AI policies when they knew their instructor could see a record of their behavior.

In fact, the version of ChatGPT used in SWAG can change its behavior based on guardrails set by instructors for specific assignments. For example, the instructor can set up the system to decline requests to write the whole essay or to ask questions to stimulate critical thinking rather than give an immediate answer.

Dan Dunlap, visiting associate professor in computer science and a collaborator on the project, said the goal is integrate AI use into education and to give instructors the ability to direct how students interact with it.

The team, which also includes graduate student collaborators Daniel Manesh and Alice Jang, along with Young-Ho Kim at NAVER AI Lab, is now building a version that integrates directly with classroom learning management platforms. Instructors will be able to specify acceptable types of AI assistance for an assignment and receive transparency reports outlining students' AI interactions alongside submitted work.

Making AI use more visible and educational

In their classes this fall, most college students will use AI tools in ways their teachers can't see. This research is a rare attempt to capture student-AI collaboration in real time.

The team's ultimate goal is not to ban AI tools from the classroom but to find ways that instructors can guide its use while making students' interactions with the technology more visible and, ultimately, more educational.

"AI is not going to go away," Jelson said. "So let's build it into the curriculum in a way that makes sense."

Original study: 0.1145/3772318.3791056

Share this story