Evidence attribution, not AI detection · CoRAL Lab, ASU

See how it
was written.

DraftTrace records the writing process and attributes every character to its origin — typed, AI-assisted, cited, or uncited. Evidence an instructor can defend and a student can stand behind.

Native Canvas integration. A research project of the CoRAL Lab at Arizona State University.

attention_essay.md — recorded Recording
Typed AI-assisted Cited paste Uncited paste
The problem

AI detectors ask one question — and answer it badly.

Every major detector guesses “Is this AI?” from the finished text alone. The guess is unreliable, and it falls hardest on the writers least able to contest it.

61.3%

false-positive rate of AI detectors on essays by non-native English writers — and it dropped once those writers used AI to polish their vocabulary.Liang et al., Stanford · Patterns (2023)

Three structural flaws follow: detectors infer from a finished artifact (trivially gamed by paraphrasers), they output a probability, not an audit trail, and they penalize the constrained while rewarding hidden AI.

The turn

DraftTrace inverts the premise.

Instead of inferring AI from the finished text, DraftTrace records the writing process — keystrokes, pastes, edits, and in-app AI assistance — and attributes every character to its real origin. The evidence is recorded, not inferred: an auditable trail, grounded in what the student actually did.

Proof, not probability. Evidence, not accusation.
The instructor report

Four views. One recorded truth.

One submission, presented four ways — from a glance to the full replay. Everything an instructor sees is a reading of the same provenance record.

View 01

Global

Composition at a glance — how much was typed, AI-assisted, cited, and uncited. The clean read you grade from.

Smartphones fragment attention in ways that are easy to underestimate, as my own week-long log showed. Sana et al. (2013) found multitasking lowers comprehension. Working memory is severely limited.
View 02

Type Flow

The full text, colored character by character. You see exactly which sentences came from where — in the text itself.

View 03

Time Flow

Watch the draft get built, keystroke by keystroke. The moment a 400-word block appears at once is visible — not inferred.

Typed
Pauses
Pastes
Class med.
View 04

Comparisons

Where this draft’s process sits against the rest of the class, per category. No competitor surfaces class-level process comparison.

The advisory verdict

The verdict points.
The provenance proves.

DraftTrace summarizes each submission into a four-category advisory signal so an instructor can triage a class fast. Unlike a detector’s flag, it’s never the end of the conversation — it opens directly onto the evidence that justifies it.

Instructor-only and advisory: a reading of the record, never a standalone score a decision rests on. The evidence is.

Submission · J. RiveraBorderline · advisory
Typed54%
AI-assisted (disclosed)12%
Cited paste18%
Uncited paste16%
Positive by design

“Fully-human” is not the only good outcome.

A draft built largely with the sanctioned in-app assistant — and properly attributed — is a legitimate, well-disclosed submission. DraftTrace is the only system that can represent AI use as legitimate and attributed, distinct from hidden uncited paste.

AI on the record
Sources cited
Self-disclosed
The research behind it

Built in a lab, measured like one.

DraftTrace is a research project of the CoRAL Lab at ASU, presented at the ACL Birds-of-a-Feather session “NLP for Education and Workforce Readiness,” July 2026. Two studies are underway: a ground-truth provenance benchmark against seven baseline detectors, and a comparative trust study with graduate-student graders.

Benchmark · provenance vs 7 detectors ESL gap · expected ≈ 0 Trust study · graders, 5 tools ACL BOF · July 2026

Honest scope. DraftTrace reads only what happens in its own editor — text from an unmonitored source is captured as uncited paste, but its true origin is unknown. It does no web-scale matching, and the verdict layer is research-stage. The point isn’t to stop every workaround; it’s to raise the bar and make the record defensible.

See it in two minutes

Evidence an instructor can defend.

DraftTrace is being evaluated with instructors in real courses. Watch the walkthrough, and if you’d like to follow the research or discuss a pilot, tell us.

Native Canvas integration (LTI). Presenting at ACL BOF — July 2026.