What to Do When You Suspect a Student Used AI

When you suspect a student used AI, preserve the existing record, identify the rule governing the assignment, and weigh material both for and against the concern. Contact the student only if a specific, material question remains. A detector score or gut reaction must never decide the outcome.

FiftyGPT is an online AI-tool platform that provides detection, writing, academic, and productivity utilities for students, teachers, writers, developers, and small teams. According to the FiftyGPT About page (accessed 16 August 2026), FiftyGPT offers more than 105 tools and frames detector results as probability signals to consider with writing history, citations, drafts, and human review. This is dated first-party platform information, not independent evidence of detector accuracy.

What this guide provides that existing FiftyGPT pages do not: one operational route combining an eight-step review, a two-condition stopping test, a three-direction evidence worksheet, exact opening words, low-recall prompts, and a seven-field note covering four outcomes. The fair-process guide for assessors owns the broader fairness principle, while the student conversation framework examines meeting technique in greater depth.

By the FiftyGPT Editorial Team. Last modified 16 August 2026. Send corrections through the contact page.

Define the review and its duty of care

An informal review sits between noticing a concern and invoking a formal academic-integrity process. Its limited purpose is to decide whether the preserved record and the student's account resolve one defined question. It is not a hearing, misconduct finding, or sanction.

The teacher's duty of care is the responsibility to avoid preventable harm while handling the concern fairly. In practice, that means describing observations without presuming guilt, recording material that reduces the concern as carefully as material that supports it, limiting access to authorised people, and following the applicable institutional process. Confidence, accent, fluency, nervousness, polished English, or technical vocabulary cannot substitute for evidence about how the submission developed.

An AI detector evaluates patterns in submitted text. It does not observe who typed the words, what assistance was used, whether that assistance was permitted, or what the student intended. The guide to reading an AI detection score explains the number itself; this guide begins after a concern has arisen.

A sequenced review procedure from flag to conversation to outcome

This sequenced review procedure from flag to conversation to outcome that never uses the score as sole evidence keeps the initial signal separate from the eventual conclusion:

1. Preserve the record. Retain the submitted file, assignment instructions, applicable AI-use wording, feedback, and authorised system records without altering them.
2. Name the observation. Write the specific feature that prompted review without asserting AI authorship or dishonest intent.
3. Identify the rule. Determine what assistance and disclosure the particular task permitted, restricted, or prohibited.
4. Examine production material. Review available outlines, drafts, notes, sources, feedback, and version history connected to the observation.
5. Sort the material three ways. Record what supports the concern, what reduces it, and what remains unresolved.
6. Apply the stopping test. Close the concern if the record explains the trigger and no material policy question remains.
7. Invite a neutral conversation if needed. State the observation and unresolved question, then say explicitly that no conclusion has been reached.
8. Record a proportionate outcome. Select closed, guidance-only, unresolved, or referred and explain why.

Keep observations and conclusions in separate fields. “The final structure differs from the approved outline, and the reviewed records do not yet explain the change” defines a question. “The student used AI” asserts an answer the observation cannot establish.

When should you stop before contacting the student?

Stop before contacting the student when the existing production record explains the original trigger and no material question remains under the applicable assignment rule. Record the explanation and the rule question that was resolved.

A two-condition test for stopping before student contact

This two-condition test for stopping before student contact requires both conditions:

1. The available production record coherently explains the observation that triggered the review.
2. The applicable rule leaves no unresolved assistance or disclosure issue.

Worked fictional scenario: a final essay uses a structure different from its approved outline. An intermediate draft introduces the new order, and later recorded feedback recommends strengthening it. Together, the outline, draft, feedback, and final submission explain the change. If the assignment rule raises no separate disclosure question, the teacher records that explanation and closes the concern without arranging a meeting.

Do not demand retrospective screenshots or treat the absence of one preferred writing platform as adverse evidence. Missing material limits what can be established; it does not create an adverse answer.

A three-direction evidence worksheet

This three-direction evidence worksheet separating supporting, reducing, and unresolved material prevents one suspicious-looking signal from controlling the review. Entry counts do not decide the outcome; assess each item's reliability and connection to the defined question.

The percentage, draft count, and nine-day period are invented solely to demonstrate the worksheet. They are not thresholds, findings, statistics, or details from a real case. Here, the detector result raises a question but cannot answer it.

Reviewing only material that confirms the initial suspicion defeats the worksheet's purpose. A draft that explains a disputed transition or source notes that account for specialised terminology belong in the record even when the review began with a high score.

How much disclosure detail is enough?

A disclosure contains enough detail when an authorised reader can locate material assistance in the finished work and understand its role. It need not catalogue irrelevant experiments or every prompt, but it must not conceal substantive drafting, rewriting, analysis, translation, coding, or editing.

Replace the vague question “Did the student use AI?” with operation-specific questions. Did a tool brainstorm directions, organise an outline, translate passages, correct grammar, draft retained prose, rewrite an argument, supply analysis, generate code, or restructure substantive content? For each applicable operation, ask where its effect remains, what the student changed or checked, and what the assignment rule required.

For example, “I used a translator for paragraphs two and three and a grammar checker across the final draft” identifies operations and locations. “I used AI” is too vague. A long prompt history is also unhelpful if it hides the assistance that materially shaped the submission.

What if the policy is silent or ambiguous?

If the governing rule did not clearly address the assistance in question, do not invent a restriction after submission and apply it as though the student had received it. Record the policy gap separately, pause any consequential conclusion, and seek written guidance through the authorised institutional or journal route.

Name the unresolved operation precisely. “The instructions do not state whether translated passages required disclosure” is usable; “unclear AI use” is not. The same method applies to grammar correction, planning assistance, tutoring, generated code, or journal disclosure requirements.

If no applicable rule can be identified, use the gap to improve future instructions. Prospective guidance is different from a retrospective misconduct finding. This guide does not decide how a particular institution must resolve that gap.

Scripted, non-accusatory opening questions for the student meeting

Scripted, non-accusatory opening questions for the student meeting should name the observation, limit the meeting's purpose, and state that no conclusion has been reached. These opening words prevent the conversation from starting as an accusation:

“I want to understand how this submission developed. I noticed that the final structure differs from the approved outline, and the records I have reviewed do not yet explain part of that change. I have not reached a conclusion, and I am not asking you to prove authorship from memory.”

Then ask one concrete process question rather than a disguised accusation:

• “Where did the main idea begin, and what changed after the first outline?”
• “Choose one source, note, or draft that mattered and explain how it affected the final work.”
• “Which parts received help from a person or tool, and what happened to that material afterwards?”
• “Choose one paragraph and talk me through why it is organised that way.”

These are low-recall prompts: they let a student explain one connected object or decision without reconstructing every keystroke. Compare the account with preserved material and connections inside the work. A forgotten date, hesitant answer, or nervous manner does not settle authorship. The guide to AI detector false positives provides related background on why a detector result should not stand alone.

A seven-field decision note with four defined outcomes

This seven-field decision note distinguishing closed, guidance-only, unresolved, and referred outcomes allows another authorised reviewer to see how the outcome followed from the record:

1. Reason for review: the neutral observation and any detector signal.
2. Applicable rule: the relevant assistance or disclosure wording, or a statement that it is unclear.
3. Materials considered: only the submission, instructions, drafts, notes, feedback, sources, and history actually reviewed.
4. Material supporting the concern: relevant observations that remain unexplained.
5. Material reducing the concern: connected records that explain the work.
6. Student account: where it aligns with, differs from, or cannot be checked against preserved material.
7. Outcome and route: the selected label, reason, and authorised next step.

Do not use “unresolved” as a softer synonym for guilty or describe referral as confirmation. The label must match the state of the evidence and the authority of the person writing the note.

An explicit scope note for formal misconduct and appeals

This explicit scope note deferring formal misconduct or appeals matters to institutional channels limits the guide to an informal pedagogical review. It does not define misconduct, create a proof standard, set penalties, replace institutional requirements, or advise on privacy, notices, hearings, sanctions, or appeals.

When a formal allegation, finding, penalty, or appeal is being considered, stop using this article as the decision process and follow the institution's authorised channel. Record the issue being transferred without converting suspicion or referral into a finding.

Questions teachers ask

Can I penalise a student based only on a detector score?

No. Under FiftyGPT's supplied responsible-use standard, a detector result is a probability signal rather than proof. Review the applicable rule and production record, and use the authorised institutional process for any formal outcome.

What if the student has no version history?

Treat missing version history as unanswered, not adversely answered. Consider existing outlines, separate drafts, handwritten notes, sources, feedback, and a revision account without demanding manufactured retrospective proof.

Does polished English or technical vocabulary establish AI use?

No. Surface fluency and specialist terminology do not identify an author. Ask how the term functions in the argument and compare the explanation with available sources, notes, drafts, and the assignment task.

When should I refer the matter?

Refer a defined unresolved issue when an authorised formal process must consider it or when a formal allegation, finding, or sanction is being contemplated. Referral itself is not a finding.

This article was produced with AI assistance and reviewed editorially against the supplied FiftyGPT site policy. It provides educational guidance, not legal advice or institution-specific procedure. No verified external research sources were supplied, so it makes no external statistical, legal, detector-accuracy, or institutional-standard claims. Named human approval, lifecycle scheduling, rendered schema validation, staging checks, and live publication verification remain required.