How Students Use ChatGPT for Essays, and What Cut the Outsourcing

Researchers recorded every request 77 college students made to ChatGPT while writing an essay. The most common request was the essay itself. What cut those requests later was not a detector.

The study comes from a Virginia Tech team led by Sang Won Lee, an associate professor of computer science, and it was published in the Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. Virginia Tech described the work on October 8 and called it a rare attempt to capture student and AI collaboration in real time. It is about as close as anyone has come to seeing how students use ChatGPT for essays without guessing.

Most of what faculty know about student AI use comes from two places: what students tell us, and what a detector says. This study suggests both are weaker evidence than we thought, and it points at something better.

How students use ChatGPT: what 77 of them actually asked for

The team built an online writing environment with its own ChatGPT tab. It logged every question a student asked, every answer that came back, and every paste into the essay, down to the keystroke. Students wrote an argumentative essay in response to a prompt that laid out three perspectives on an issue, in about 30 minutes, and were told to write it as if it were a class assignment submitted for a grade. Nobody told them to use ChatGPT or to avoid it.

Here is what the record shows.

Six of the 77 students never used ChatGPT at all. Twenty, about one in four, simply asked it to generate the entire essay, and the paper reports that this was the single most common request in the study. In most of those cases they pasted the essay prompt with little of their own thinking added, and it was often their first request, which means many of them started with a finished draft and worked backward. Count requests for part of the essay too, a paragraph or a conclusion, and the paper finds that more than half of the 77 asked ChatGPT to write some of their essay at least once.

Others used it more narrowly. Planning help came next, with 35 of the 77 asking for help choosing a position or generating ideas, and 34 asked for help reviewing, mostly proofreading and polishing a draft they had written.

The researchers also named a pattern they call vibe writing. A student steers the AI at a high level, asks for changes, asks again, and assembles an essay the student never quite writes. The paper notes that this work appeared original and was far harder for instructors to recognize as AI assisted than a pasted prompt. That is the student a detector is least likely to catch, and the one whose learning is most at risk.

The essay requests are not the most uncomfortable finding. How the students felt afterward is.

The ownership problem nobody can see

After writing, students rated how much the essay felt like theirs. Students who asked ChatGPT to write the whole thing reported significantly lower ownership. No surprise there.

The surprise was everyone else. Students who leaned on ChatGPT for planning or for reviewing reported a sense of ownership comparable to students who never used it at all. A student who let the AI choose the argument still felt like the author.

“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?”

That finding sits beside one we keep coming back to. When a Berkeley researcher found the share of A grades growing more in writing and coding courses than in other classes since ChatGPT arrived, his read was that students had relied on generative AI to do better in their studies, not that they were learning more. A grade can rise and a feeling of authorship can hold steady while the thinking quietly moves somewhere else.

That is the part of how students use ChatGPT that no policy statement reaches. Ask a student who planned with ChatGPT whether the essay is their own work, and you will often get an honest yes. It just does not answer the question you meant to ask.

If self report fails and detectors miss the students most at risk, what is left?

What changed behavior was visibility

The team asked a second question in a separate classroom deployment, which Virginia Tech reports is not part of the published paper. Students completed assignments in a system the researchers built, called Student Writing with Accountable Generative AI. They were told their AI interactions would be recorded and could be seen by their instructor.

According to Lee, that single change shifted how students approached the work. Requests for the AI to write the essay dropped significantly, and the drop held even when submissions were anonymized, so the instructor could not tell which student wrote which essay.

“This suggests that the students were internalizing the accountability norm,” Lee said.

Read the next line carefully, because it is the whole argument. “The drop in outsourcing behavior points to a shift in how students saw the task itself, not just a fear of consequences,” Lee said. The same system also let instructors set guardrails for a specific assignment, so the AI could decline to write the whole essay, or ask questions instead of handing over an answer.

One caution. The classroom result comes from the university’s account of work still in progress, and the published study measured what students did, not what visibility does. Treat it as a strong signal, not settled science.

It is also the same argument this site made in September, when mathematics settled its own AI dispute with receipts instead of detectors. A detector asks a question after the work is done. A record changes the work while it is happening.

The AI fluency framework we keep returning to describes managing AI as delegating tasks responsibly while a person sets the parameters and reviews what comes back. You do not need the Virginia Tech software to borrow the principle. You need to say, before the essay is written, which kinds of help are open, and what record a student hands in with the work.

Make the draft visible in your own course this week

Here is the method, small enough for one assignment.

One. Name the skill the assignment exists to build. Not the topic, the skill. If the point is forming an argument, then planning is the assignment, and help with planning is the help that costs the most learning. If the point is research and synthesis, the risk sits somewhere else. The paper’s own caution is that even planning or reviewing help can shape learning, depending on what the assignment is for.

Two. Mark each kind of help. Use the study’s four categories: planning, which means choosing a position and generating ideas; drafting, which the paper calls translating, which means turning ideas into sentences; reviewing, which means revising and proofreading; and the whole draft. For each one, decide whether it is open, open with a record, or closed, and write one line on why. Students follow a rule better when they can see the reason for it.

Three. Ask for the record, and grade it lightly. The record can be simple: a link to the chat or a pasted transcript, plus three sentences on what the student asked, what they kept and what they changed. Put one line on the rubric for the process, worth enough to matter and not so much that it becomes the assignment. Tell students on day one that you will read it, the same way the classroom deployment told its students.

This is not surveillance, and it should not feel like it. In the classroom deployment, students knew their AI use was on the record and used AI anyway. They just leaned on it differently.

A detector asks whether a student cheated. A visible draft shows you what the student actually did, and that is the only one of the two you can teach from.

Your next step

Take one writing assignment through The Visible Draft Map, by hand, in about 25 minutes.

When you finish you will hold a stage by stage AI map for that assignment, the record students hand in with it, a process line for the rubric and a paragraph for the assignment sheet. Then the next thought arrives on its own. That is one assignment. Every writing assignment in the course needs the same map, and so does every section you teach.

That is the job the Assessment Builder does. It builds assignments, rubrics and tagged question banks grounded in your Course Record, the five minute record of your course, your outcomes, your policies and your voice that every tool in the toolkit uses. Put your AI map in your Course Record once, and every assignment and rubric you build carries it.

If you would rather think it through with other instructors first, the Harvest Kernel Learning Community is free, and this question is the conversation there this week.

Sources

• Tonia Moxley, “Research reveals the ways students really use AI to write,” Virginia Tech News, October 8, 2026. news.vt.edu
• Andrew Jelson, Daniel Manesh, Alice Jang, Daniel Dunlap, Young-Ho Kim and Sang Won Lee, “An Empirical Study to Understand How Students Use ChatGPT for Writing Essays,” Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. doi.org, full text at arxiv.org
• “‘A’ Grades Are Suddenly Everywhere Since the Arrival of ChatGPT,” The Wall Street Journal. wsj.com