How teachers can spot and handle suspected AI-written student submissions

What to look for

Concrete signals teachers can review

Rather than relying on a single detector score, evaluate multiple explainable signals across the submission, student history, and system metadata. Use these signals as prompts for human review, not as an automatic verdict.

• Abrupt style or register shifts: compare sentence length, vocabulary level, and common errors to prior student work; flag passages with noticeably different phrasing.
• Citation and reference anomalies: inconsistent formats, invented sources, or mismatch between claims and cited evidence.
• Metadata discrepancies: submission timestamps, file creation vs. upload times, and unexpected revision histories.
• Repetitive or high-density content patterns: repeated phrases, unusually uniform sentence structures, or excessive use of generalizations.
• Task mismatch: answers that satisfy rubric language but miss scaffolded elements or specific instructor prompts.

Collect reviewable artifacts

Source ecosystem: where to gather evidence

A reliable review uses multiple sources. Preserve copies and document your collection process before discussing findings with students or escalating.

• Submitted assignment text (LMS files and pasted text)
• Drafts and version history (timestamps, prior uploads, comments)
• Student writing baselines — earlier essays, low-stakes journal entries, discussion posts
• LMS metadata: submission times, file names, late-submission flags
• Plagiarism engine overlap reports and matched passages
• Code repositories, project artifacts, and execution logs for programming tasks
• Synchronous assessment records: in-class quiz responses, oral exams, proctoring logs where applicable

A repeatable process

Teacher-centered review workflow

Use a short, defensible workflow that balances inquiry with privacy. The workflow below is classroom-ready and adaptable to institutional policy.

• 1) Capture and preserve: download the submitted file(s), save metadata, and take screenshots of LMS views.
• 2) Snapshot baseline comparison: pull two earlier samples from the student and note stylistic differences with examples.
• 3) Mark passages of interest: extract 3–6 passages that triggered review and annotate specific concerns (style, citation, content mismatch).
• 4) Generate an evidence summary: 3–5 neutral bullets summarizing why the submission prompted review (use direct quotes and timestamps).
• 5) Decide next step: low-confidence — assign a short verification task; moderate — have a neutral conversation; high-confidence — prepare a referral packet.
• 6) Record outcome and remediation: document the meeting, student response, and any follow-up assignments or appeals.

What to say and submit

Classroom-ready templates and scripts

Use neutral, non-accusatory language for initial conversations and referral packets. Below are ready-to-use templates you can adapt.

Prevention through design

Assessment redesign to reduce misuse

Well-designed assignments make it harder to misuse AI and easier for instructors to determine authorship. Small changes often create large differences in validity.

• Staged submissions: require multiple drafts with instructor feedback and documented revision history.
• Low-stakes, frequent writing: short in-class or timed reflections that build a writing baseline.
• Personalized prompts: ask students to apply course-specific examples, local data, or reflections on in-class activities.
• Oral or synchronous checkpoints: include short oral defenses, presentations, or live problem-solving tied to the assignment.
• Focus rubrics on process: grade research notes, annotated bibliographies, and incremental milestones.

Prompts instructors can run or adapt

Prompt clusters for practical review tasks

Use these prompt clusters to produce consistent reviewer outputs that can be attached to a referral or used in a meeting.

• Summarize suspicious features in this submission and list the specific passages that merit human review.
• Compare this submission to the student’s prior paper and highlight stylistic or vocabulary shifts with side-by-side examples.
• Produce a 3–5 bullet evidence summary suitable for a referral packet, with redaction guidance for privacy.
• Generate a neutral conversation script to discuss suspected AI use with the student, including suggested questions and next steps.
• Create a review checklist that maps detected signals (style change, citation anomalies, metadata issues) to recommended instructor actions.

Documentation best practices

Preserving evidence and protecting privacy

Collect only what you need, timestamp actions, and limit distribution of sensitive artifacts. Redact student identifiers for departmental reviews when possible.

• Preserve original files and LMS views in a secure folder with access limited to required reviewers.
• Log dates, times, and steps taken during the review; avoid editing or annotating originals in ways that obscure the chain of custody.
• When sharing with departments, provide redacted summaries that include quoted passages, timestamps, and a list of preserved artifacts rather than full files.
• Consult institutional policy and your FERPA officer before sharing student records beyond the integrity office.

Fair adjudication

Practical decision rules to avoid false positives

Combine multiple explainable signals, use neutral language, and prefer verification tasks when doubt remains. Treat stylistic differences as a prompt for inquiry, not proof.

• Require at least two independent, explainable signals before escalating (e.g., large stylistic shift + metadata anomaly).
• If signals conflict or are inconclusive, assign a short in-class verification task or staged draft before formal referral.
• Document student responses and offer remediation options that emphasize learning outcomes over punishment.