Chapter 4. Academic Integrity and AI Use Policies

Learning objectives

• Select an AI permission level based on the ability the assignment aims to assess, and state it in the instructions so students know in advance.
• Have students disclose AI use by distinguishing among an AI Usage Log, citations of original sources, and quotation marks for AI-generated sentences.
• Explain with official sources why AI detectors cannot serve as the basis for judging students.
• Select assignment designs that verify authorship and understanding without relying on detectors.
• Design a fair handling procedure from verifying standards to notification when misconduct is suspected.

Introduction

At a high school teacher training session, the facilitator posed a question: "If a student received a report outline from an AI and wrote the report themselves, is it misconduct?" The answers of the teachers who raised their hands were split exactly in half. To the follow-up question, "Then what about using a spell checker?", everyone answered no. At "What if they revised their conclusion after looking at counterarguments received from an AI?", opinions were divided once again.

What this scene demonstrates is not that the teachers' thinking was wrong, but the fact that without established standards, different people judge the same behavior differently. The same is true for students. Students who do not know what is permitted can violate standards even in good faith, and conversely, end up hiding even permitted assistance. In a survey introduced by UNESCO when it published its generative AI guidance in 2023, fewer than 10% of more than 450 schools and universities had institutional policies or official guidance on the use of generative AI.

This chapter begins with establishing standards. It determines AI permission levels according to assignment objectives and establishes rules for disclosing and citing AI use. Next, it examines the limitations of AI detectors using official sources and looks at assignment designs that verify authorship without detectors. Finally, it addresses handling procedures when misconduct is suspected and creating school-wide standards.

Key concepts

The questions of academic integrity have changed

Before AI, academic integrity mainly asked, "Did you copy someone else's writing?" In the AI era, the question divides into three:

1. Authorship: Can the parts written directly by the student and the parts generated by AI in this writing be distinguished? Did the student disclose this honestly?
2. Process: Did the student actually go through the abilities the assignment aims to cultivate (research, argumentation, expression)?
3. Accuracy and responsibility: Were the facts provided by the AI verified? Can the student take responsibility for the content of the output?

These three questions are more sophisticated than "Did you use AI?" A student can use AI and still provide good answers to all three questions, and can fail in accuracy even without using AI. Therefore, a good AI use policy addresses how it is used and disclosed, not whether it is used.

From assignment objectives to permission levels

AI permission levels stem not from the subject, but from what the assignment aims to assess. This book uses four levels.

Even within the same subject, levels can vary by assignment. A beginning-of-semester writing diagnostic corresponds to Level 0, an inquiry report to Level 1, and an "AI response verification" activity to Level 3. The notion that "the level must be identical across the same subject to be fair" is mistaken. Fairness comes from applying the same criteria to the same assignment, and differing levels across assignments can prevent confusion through clear labeling in the instructions.

Conversely, avoid the approach of "leaving the decision of AI use to student judgment, and if used, having them merely write 'Assistance Received'." Because it is impossible to know who received what and how much, assessment becomes blurred, and students who did not use it conscientiously may actually be put at a disadvantage.

Writing permission level instructions

Write the permission level in the assignment instructions so that students can read and understand it immediately.

[Example of AI use criteria in assignment instructions]
AI permission level: 1 (Exploration Only)
- Permitted: Finding topic candidates, finding reference materials, asking for definitions of terms
- Prohibited: Receiving body sentences, receiving conclusions, polishing sentences
- What to submit: AI conversation logs (screenshots or exported files via the school-designated method), 'AI Usage Log' at the end of the report
- Facts and statistics provided by AI: Check against original sources and cite the original sources
- If suspected of violating criteria: Confirm evidence according to school academic integrity regulations and provide an opportunity for explanation

It is important to state specifically in the instructions "what you must not do." Vague sentences like "Do not rely on AI" are interpreted differently by each student.

AI Usage Log

The most fundamental method of disclosing AI use is the usage log. Have students record the following four items at the end of the report:

OpenAI's educator guidance also introduces methods for having students log AI use, cite sources, and share conversations they had with AI. However, because conversation share links can often be viewed by anyone who has the link, if personal information is contained in the conversation, it is safer to have students submit screenshots or exported files using a school-designated method. Sharing conversations allows teachers to see the student's level of questioning and critical thinking, enables students to reflect on how their questions improved, and verifies whether they used AI responsibly.

The wording of the criteria determines what should be recorded in the usage log. If the standard is "Record parts where AI was involved in content generation; do not record cases where only orthography was corrected, such as spell checks," then reflecting AI translation results in a summary or rewriting a section with counterarguments presented by AI is recorded, while spell checking is not. If the standard is "Record all AI use," then spell checking is also recorded.

Rules for disclosure and citation

Disclosing AI use does not end with the usage log alone. Rules differ depending on how the AI was used.

1. Only ideas were obtained → A usage log is sufficient.
2. Using facts or statistics provided by AI → Do not cite the AI as the source. Find and verify the original source, and cite the original source. If the original source cannot be found, do not use that fact as evidence.
3. Including sentences written by AI verbatim (in permitted assignments) → Distinguish them from student sentences with quotation marks, and disclose which tool generated the sentences.
4. AI processed student writing, such as translation, summarization, or sentence polishing → If within the scope where criteria require logging, record in the usage log what was processed and how. If the processing result became content (e.g., reflecting the content of translated material in a summary), it is regarded as use involved in content generation.

The second rule is identical to the principle seen in Chapter 1 ("Facts provided by AI are clues, not sources"). The third rule is to protect authorship. If an assignment asks to "Compare a summary written by AI with a summary written by yourself," the AI paragraph must be included, but that paragraph must not appear as if it were the student's writing.

Listing AI as a co-author is not appropriate. An author is a person who takes responsibility for the content, and AI cannot bear responsibility. Responsibility for the content of the report lies with the student who submitted it.

Why AI detectors cannot serve as evidence

AI detectors are tools that show the probability of a text being written by AI as a score. While they appear convenient, OpenAI explains detectors in its educator guidance as follows:

• In its own research, detectors were not reliable enough for educators to make long-lasting decisions about students.
• Detectors trained by the company judged human-written text as AI-written, such as the works of Shakespeare or the Declaration of Independence.
• There were indications that they could disproportionately impact students who learned or are learning English as a second language, or students whose writing is particularly formulaic or concise.
• Even if a detector is accurate, students can evade detection with minor edits.
• ChatGPT does not know what text was generated by AI, and when asked questions such as "Did you write this text?", it can fabricate groundless answers.

OpenAI also notes that this explanation is not an evaluation of detectors from other companies. However, considering that the same limitations can arise from the principles of detection, and given the weight of the consequences left on students, using detector scores as evidence for judging misconduct is dangerous. The same holds true even if results from multiple detectors coincide. Agreement does not guarantee accuracy.

Designing to verify authorship without detectors

Instead of detectors, place verification mechanisms within assignment design.

When choosing mechanisms, consider the number of students, teacher time, and the ability the assignment seeks to evaluate together. Having 180 students record their entire writing process on screen makes it impossible for teachers to review and imposes a heavy privacy burden on students. Changing an essay to be written by hand during an exam prevents AI intervention, but loses the objective of evaluating research skills concurrently. In-class drafting and short oral verification often allow direct verification of process and understanding with minimal time.

Procedures when suspicion arises

Cases suspected of violating announced criteria are handled in the following order. The results of each step serve as material for the next step.

1. Confirm criteria: Check the AI use criteria and instructions announced for the corresponding assignment. These criteria determine what constitutes a violation.
2. Collect evidence: Gather evidence in light of the criteria, such as submissions, conversation logs, in-class drafts, and staged submissions.
3. Notice and explanation: Inform the student of the grounds for suspicion and the evidence, and hear their explanation.
4. Decision: Review the explanation together with the evidence, and determine whether a violation occurred according to school procedures.
5. Notification and appeal: Notify the student of the result and reasons, and provide guidance on how to file an appeal.

Even if evidence is clear, the opportunity for explanation is not skipped. An explanation is a mechanism to correct misunderstandings of factual relations and protect the legitimacy of the handling outcome. A method where a teacher deducts points alone and ends the matter bypasses due process.

With and without standards

Even for the same suspicion, the response differs completely depending on whether standards existed.

If an assignment that lacked standards is penalized retroactively under new criteria, students suffer disadvantages due to rules they could not have known. The claim that "academic integrity should naturally be observed even without standards" may apply to longstanding rules like copying someone else's writing verbatim, but it does not apply to areas where standards differ across schools, such as the extent to which AI may be used.

Types of risks arising from student AI use

Problems with student AI use are not all of the same type. Prevention methods vary depending on the type.

Data manipulation, in particular, is an issue of a different dimension from citation mistakes. This is because it involves fabricating non-existent research results (fabrication) or arbitrarily altering actual results (falsification). It can be prevented by requiring the submission of raw materials.

Creating school-wide standards

Assignment-specific instructions are difficult to keep consistent through the efforts of individual teachers alone. Having school-wide standards makes it possible to transparently manage differences among teachers. Good school standards encompass five elements:

1. Definition of permission levels: Levels such as 0–3 and what can and cannot be done at each level
2. Mandatory labeling: Indicate the level on all assignment instructions
3. Logging and disclosure methods: Usage log templates, methods for submitting conversation logs, citation rules
4. Verification methods: The principle of not using detectors as evidence, and verification mechanisms to use instead
5. Procedures: Sequence for handling suspected cases, opportunity for explanation, methods for filing an appeal

Approaches such as banning AI school-wide and having teachers permit ad-hoc exceptions, having teachers provide only verbal guidance without school standards, setting a threshold for detector scores, or merely collecting pledges fail to satisfy the condition that "what is permitted for each assignment must be known in advance."

Beyond rules to literacy

The purpose of an AI use policy is learning, not policing. OpenAI's educator guidance explains that having students share conversations with AI allows them to demonstrate their ability to understand AI's limitations and critically verify information, and prepares them for responsible use in a future where AI use will become commonplace. UNESCO also emphasizes a human-centered approach and age-appropriate use regarding generative AI.

Therefore, policies should serve as guidance showing "use it like this and disclose it like this" rather than a list of "do nots." In Level 3 assignments, activities that identify errors in AI responses and correct them with evidence teach students both how to use AI and how to doubt AI.

Going deeper

How to read "98% accuracy": The base rate trap

Let us suppose a vendor advertises a detector's accuracy as 98%. What this number actually means cannot be known from the number alone. This is because it does not reveal what texts were tested, whether texts by English learners were included, or whether texts revised by students were also tested.

Let us do a simple calculation for the sake of explanation. This is not vendor data, but an assumption to aid understanding. Suppose that in a certain school, among texts submitted by 500 students, 20 texts are actually written with AI, and suppose that the detector's rate of falsely classifying human-written texts as AI is 2%. Then, out of the 480 human-written texts, approximately 10 texts are falsely classified. Even if the detector catches all 20 AI texts, among the 30 texts marked as "AI-written," one-third are texts by falsely accused students. And if those falsely accused students are concentrated among English learners who write formulaic texts, the problem becomes even more serious.

What this calculation shows is that the more honest students there are, the more false accusations even a small error rate creates. In statistics, this is called the base rate trap. Even if the overall accuracy is high, if texts actually written with AI are rare, the probability that a verdict of "AI-written" is correct drops significantly. The accuracy stated by the vendor and "the probability that a text receiving this verdict is actually an AI text" are different numbers. The reason a detector cannot serve as the basis for judging a student is not only because the accuracy figure is low, but because the cost of a false judgment is too high compared to the weight of a judgment rendered on an individual student.

The cost of a ban policy

Imposing a blanket ban on AI seems simple and fair. However, a ban entails three costs. First, it is difficult to enforce on take-home assignments, and there is no reliable way to single out students who did not comply. In the end, only honest students may be disadvantaged. Second, students lose the opportunity to learn how to use AI responsibly. Third, as students use it secretly, the best evidence—the record of use—disappears.

Of course, there are assignments for which a ban is appropriate. Assignments that diagnose one's current independent ability, and basic skills practice conducted during class, are such examples. The key is to prohibit it not across an entire subject or school, but at the assignment level, and in conjunction with verifiable methods (such as in-class writing).

Who is an oral check fair to?

A short oral check is a much better means of verification than a detector, but it is not completely fair in itself. Students who lack confidence in speaking, students who are learning Korean, and students with high anxiety may not be able to explain well in speech even text that they wrote themselves.

Therefore, an oral check is designed not as an exam for punishment, but as a conversation to verify understanding. The scope of questions is announced in advance (e.g., "explain the core claim and one most important piece of evidence"), and if necessary, written alternatives or interpretation support are provided. Misconduct is not determined solely by the fact that a student could not explain well orally; it is examined alongside other evidence (such as in-class drafts and milestone submissions).

At work

Green Hill High School: Standards agreed upon by 52 teachers

As AI-related complaints arose repeatedly during the semester, Green Hill High School decided to establish student AI use standards. In a teacher survey, opinions were split between "blanket ban" (31%), "subject discretion" (46%), and "open use" (23%).

The standards drafting committee reviewed three drafts. The blanket-ban draft conflicted with AI hands-on practice in the informatics curriculum, and because there was no way to enforce it on take-home assignments, they discarded it. The subject-discretion draft respected teachers' opinions, but had the problem that students and guardians could not know in advance what was permitted for each assignment. The detector-centered draft proposed by a vendor was discarded after they read OpenAI's educator guidance together. This was because human-written texts could be flagged as AI, and it could disadvantage approximately 40 multicultural students at the school.

The committee defined four permission levels (0: No use, 1: Exploration only, 2: Drafting assistance, 3: Unrestricted use) and required the level to be indicated on all assignment guidelines. Subject discretion was preserved in "which level to choose." For assignments at Level 1 or higher, submitting chat logs and records of use was made mandatory, and rules for disclosure and citation were created as a one-page student guide. When a violation was suspected, they required going through five stages: checking standards, gathering evidence, hearing the explanation, making a decision, and providing notification.

One semester later, the committee inspected a sample of 40 assignment guidelines, identified 6 guidelines that lacked level indications, and notified the relevant subjects. In a student survey, responses stating "I know how far I am allowed to use AI for each assignment" increased significantly compared to before the standards were introduced.

Cloudstream Middle School: From suspicion to conclusion

A second-year Korean language performance assessment at Cloudstream Middle School was announced as a Level 1 (Exploration only) assignment, and the submission of chat logs and records of use was required. While grading, the teacher noticed that the writing style of the concluding paragraph in one student's report was distinctly different from the preceding sections.

The first thought that occurred to the teacher was to run it through a detector, but the teacher refrained because school standards stipulated that detectors were not to be used as evidence. Instead, the teacher opened the chat log submitted by the student. In the log, there was a conversation where the student had requested, "Write an impressive conclusion," and the AI had written a concluding paragraph, and the report's conclusion was almost identical to that paragraph. This was direct evidence showing a violation of the announced standard (prohibition on receiving body sentences).

The teacher did not deduct points immediately. In accordance with school procedures, the teacher showed the student the wording of the standard and the relevant portion of the chat log, explained the reason for suspicion, and heard the student's explanation. The student said, "I thought it was okay to receive help for just the conclusion." The Academic Integrity Committee reviewed both the log and the explanation and determined it to be a violation of the standards, but taking into account that the guidelines had not explicitly specified the "conclusion," the committee decided on a measure requiring the student to rewrite the concluding paragraph within the scope set by school regulations. The outcome and the reasons were communicated to the student in writing, along with instructions on how to appeal.

After this case, the subject department explicitly specified "receiving conclusions" under "What not to do" in the guidelines. The inclusion of "receiving conclusions" in the guideline example in the preceding section also reflected this experience. The options the teacher discarded were running a detector test, immediate point deduction without hearing an explanation, and "overlooking it on the grounds that the student was honest because they submitted the log voluntarily." This was because submitting records was merely part of the standard and did not erase the violation that the record demonstrated.

Common mistakes

1. "If it is the same subject, having the same AI standard for all assignments is fair." While it appears consistent, the skills intended to be assessed vary by assignment. Applying the same standard to the same assignment is fairness, and differences across assignments are addressed through explicit labeling.
2. "There is no need to collect records for assignments that permit AI." Since it was permitted, there seems to be nothing to check, but records are necessary to verify whether it was used within the permitted scope and to evaluate the process.
3. "Statistics provided by AI can simply cite AI as the source." While citing the source seems honest, AI is not a source, but a clue to be verified. Find and cite the primary source.
4. "If multiple detectors agree, it is certain." Consensus among multiple tools appears objective, but agreement does not guarantee accuracy.
5. "You can verify by asking the AI directly." Since the AI wrote the text, it seems like the AI would know, but OpenAI explains that ChatGPT fabricates groundless answers to such questions.
6. "Even without standards, writing with AI constitutes misconduct." While it seems like common sense, in a domain where permitted scopes vary by school, punishing without standards penalizes students under rules they could not have known.
7. "If the evidence is clear, hearing an explanation can be skipped." While it appears efficient, hearing an explanation is a mechanism that upholds procedural fairness independently of the accuracy of the conclusion.
8. "Since the student submitted the records voluntarily, do not take issue with it." While it seems to respect honesty, submitting records is part of the standard. If the records show a violation, it must be handled according to procedure.

Summary at a glance

• Permission levels are determined not by the subject, but by the skill the assignment intends to assess: diagnosis → 0, research and writing → 1, revision practice → 2, AI literacy → 3.
• The guidelines specify the level, what can be done, what must not be done, what to submit, and the procedure in case of suspicion.
• Rules of disclosure: use records for ideas, cite primary sources for facts provided by AI, use quotation marks for AI sentences, and use records for processing such as translation or summarization according to standards. AI is not an author.
• Detectors: they can classify human writing as AI, may disadvantage second-language learners and formulaic writing, and can be evaded with minor edits. Asking ChatGPT also fails to verify authorship.
• Verifying authorship through assignment design: in-class drafts, milestone submissions, short oral checks, and chat logs.
• Procedure: Check standards → Collect evidence → Notification and explanation → Decision → Notification and appeal.
• Assignments that lacked standards are not penalized retroactively; instead, standards are established and applied starting from the next assignment.
• Prevention methods differ for each type of risk (factual errors, personal information, authorship, data manipulation, copyright).
• Five elements of school standards: level definitions, labeling requirement, recording and disclosure methods, verification methods, and procedures.

Check the official docs

• OpenAI Help — How can educators respond to students presenting AI-generated content as their own? — Read the "Do AI detectors work?" section and the benefits of sharing conversations in the original text.
• UNESCO — Guidance for generative AI in education and research — Read the explanations regarding human-centered, age-appropriate approaches, and the lack of institutional readiness for validation.
• UNESCO — Governments must quickly regulate Generative AI in schools — Check the survey findings showing that less than 10% of schools and universities had institutional policies.
• Anthropic — Reduce hallucinations — Reread why facts provided by AI must be verified against primary sources.
• National Law Information Center — Article 28 of the Copyright Act — Requirements when quoting the work of others in student writing.

Self-check

4-1. Write the three questions that academic integrity asks in the era of AI, and give one example of failing those questions even without having used AI.

4-2. Write the four permission levels of this book, and write one line for each level describing the nature of an assignment well-suited to it.

4-3. Write the disclosure method required for each case: when a student used AI only for ideas, when they used facts provided by AI, and when they included AI sentences verbatim.

4-4. Write four limitations of detectors stated by OpenAI in its educator guidance.

4-5. Explain why, even if the rate of incorrectly classifying human-written text as AI-written appears low, many wrongful cases arise in a group with many honest students.

4-6. Write three mechanisms for verifying authorship instead of detectors, and state which one you would choose when student numbers are large and teacher time is limited, along with your reasons.

4-7. Write the five steps in the procedure for handling suspected cases of improper use, and identify the step that must not be omitted even if the evidence is clear.

4-8. (Design task) Choose one assignment you will assign next month, and based on the examples in this chapter, write an AI use policy notice in eight lines or fewer.

Hands-on tasks