As generative AI becomes part of everyday college writing, the real question is no longer whether students are using it. It is how they use it, where they rely on it, and what that means for learning.
A new tool called DraftMarks is designed to make that hidden process visible by showing how a document changes over time and where AI entered the writing.
DraftMarks was developed by researchers at Georgia Tech and Stanford University.
The tool arrives at a moment when students are already deeply weaving AI into their work. A 2025 AI in Education trend report found that 90 percent of college students use AI in their coursework, and nearly half use it while drafting.
That has left many instructors in a difficult position. Old ways of evaluating writing no longer seem enough, but simply catching AI use does not really explain anything either.
Instead of acting like a detector that spits out a percentage or tries to guess how much of a final paper came from a chatbot, DraftMarks focuses on the writing process itself.
It shows when a student revises with AI, when they accept AI-generated material, when they delete it, and when they reject AI output altogether.
The idea is that AI use in writing is not a single event, but a series of choices. A student may use AI to test an argument, brainstorm a paragraph, strengthen a sentence, or reshape their tone.
Another student may lean on it more heavily and let it do much of the work. Those are not the same thing, and DraftMarks is built around that difference.
Rather than flattening all of that into a suspicion score, the tool tries to tell the story of how a piece of writing came together.
That shift matters because AI has become ordinary enough that simply asking whether someone used it is no longer very useful. If most student writers are already turning to AI in some form, then the more meaningful question is what role it plays in the actual work.
DraftMarks works like an augmented reading layer placed on top of a document. As someone reads, they can see visual signals that correspond to different kinds of AI involvement.
Eraser crumbs show passages that were heavily revised. Smudges point to AI-generated changes that altered the strength of an argument without fully changing the content.
Furthermore, masking tape marks passages that AI first wrote, and glue residue shows where writers later removed AI-generated text.
Finally, ghost text signals that a writer prompted AI but decided not to use the result and different fonts separate human-written lines from AI-generated ones. The result is something closer to a living record than a clean final draft.
Instead of looking polished and sealed off from its own history, the document begins to show the traces of how it was built.
"By making the invisible parts of the process tangible, it forces writers to confront whether they are truly engaging with AI or just passively accepting it," said lead author Momin Siddiqui, a master's student at Georgia Tech.
"Ultimately, it helps writers make more intentional judgment calls about how they want to collaborate with AI in the future."
The researchers did not begin by building a detector and then asking whether teachers found it useful. They started with educators instead.
In an early study involving 21 people, the team watched how instructors read student writing and what kinds of clues they naturally looked for when judging revision, originality, and evidence of learning.
Thus, DraftMarks borrowed from physical signs people already associate with writing in progress: eraser dust, tape, smudges, traces of removal.
"These marks are meant to emulate the writing process in ways we're already familiar with," said Adam Coscia, a Ph.D. student at Georgia Tech.
"They help students and teachers see the effort behind the writing, and whether students actually met the learning objective."
Behind the interface, DraftMarks tracks draft history and classifies edits and AI interactions as they happen, so those visual cues can appear almost in real time.
To see how the tool worked outside a tightly controlled setting, the researchers later tested it with 70 participants. The group included students, teachers, journalists, and general readers.
Teachers showed particular interest in the path the writing had taken. They wanted to see how ideas changed, how much AI had shaped the work, and where students were still making real decisions of their own.
Other readers, though, often focused on something a little harder to measure: trust.
For them, DraftMarks offered clues about authorial intent. It helped them judge whether a writer relied on AI carelessly or used it deliberately.
If AI increasingly becomes part of journalism, public writing, and everyday communication, people may want more than a yes-or-no answer about whether it was involved. They may want to know how it changed the voice and the choices behind the final piece.
What makes DraftMarks stand out is that it tries to move the discussion away from detection and toward judgment.
AI detectors tend to imply that the most important thing is catching misuse. DraftMarks suggests the more valuable task may be helping writers and readers think about collaboration more clearly.
"DraftMarks completely changed how I think about my own writing," Coscia said.
"I was surprised by how much I cared about authorial intent once I could actually see how AI affected my tone. It made me realize small AI choices can subtly reshape what I'm trying to say."
The researchers hope tools like this can make the conversation around AI and education more useful.
Instead of asking only whether students used AI, they may help teachers and students think more honestly about what learning looks like when humans and AI are writing together.
The project was presented at the Association for Computing Machinery Conference on Human Factors in Computing Systems in Barcelona.
-- -
Like what you read? Subscribe to our newsletter for engaging articles, exclusive content, and the latest updates.
