Ethical Framework and Academic Integrity

Ethics and AI: Asking the Right Questions

Integrating AI into education challenges concepts we once considered stable. What counts as original work when a student uses ChatGPT? Where do you draw the line between legitimate assistance and cheating? There is no universal answer, which is precisely why you need an explicit framework: anything you don’t define will be interpreted differently by every colleague and every student, and you won’t discover those interpretations until a conflict arises. In schools, part of this framework already exists: the Ministry of Education’s usage framework considers any use of generative AI for an assignment without explicit authorization or personal appropriation work to be fraud. This lesson helps you build the rest for your institution.

The Spectrum of AI Use

Student use of AI isn’t binary. Pitting total permission against total prohibition creates unenforceable rules: a blanket ban runs into the fact that some students have already violated it, while blanket permission empties certain assessments of their meaning. A spectrum exists, and five levels cover nearly every teaching situation:

• Level 0 — No AI use allowed (100% independent work)
• Level 1 — AI for idea generation and brainstorming only
• Level 2 — AI for structuring and drafting; student writes the final version
• Level 3 — AI as co-author, with attribution and critical reflection
• Level 4 — AI as an integrated tool; assessment focuses on the ability to steer the AI

These levels only make sense when applied to specific activities. Level 0 applies to proctored exams, where you measure what the student has memorized and can mobilize alone. Level 1 fits research tasks where the learner broadens their angles before choosing one, with no line of the final output coming from the tool. Level 2 matches the classic take-home assignment: AI helps build the outline, but the writing remains personal, and that’s what gets graded. Level 3 belongs in long-term projects, where you require students to attach their conversation logs with the tool as an appendix and comment on what they kept or discarded. Finally, Level 4 shifts the object of assessment: in a professional monitoring task, it’s no longer the produced text that matters, but the relevance of the prompts and the rigor of the verification.

Each level is legitimate depending on the context. The key is to state which one applies, activity by activity, and to write it on the assignment sheet itself rather than announcing it orally at the start of the semester: an instruction repeated on the distributed handout cannot be contested.

The Plagiarism Question

AI detectors are not reliable. OpenAI states in its help center that its own detector classified Shakespeare and the U.S. Declaration of Independence as AI-generated text, and these tools disproportionately penalize students learning the language or those whose writing is concise and regular. Moreover, minor tweaks are enough to evade them, and ChatGPT itself cannot tell if a text was generated: when asked “did you write this?”, it guesses randomly. The CNIL (France’s data protection authority) also notes that the Ministry advises against using these tools to identify AI-written copies. Penalizing on this basis alone exposes you to legal disputes and destroys trust.

Instead of playing detective, shift the problem to assignment design. Ask students to document their creation process on an appendix page where they describe their dead ends and choices: a thought process is much harder to fabricate than a text. Integrate visible intermediate steps—draft, approved outline, corrected version—that make full outsourcing more costly than doing the work itself. Supplement this with an oral assessment of critical thinking and argumentation: five minutes of questioning is enough to determine if the author masters their submission, and this practice accuses no one since it applies to everyone. Finally, design prompts that require personal or local knowledge—a case encountered during an internship, data collected in class—that nothing allows you to fabricate remotely. When AI is permitted, OpenAI suggests an additional path: ask students to include a share link to their conversation, which shows their approach and makes them accountable for their use.

One rule remains for you: never delegate grading or decision-making to the tool. The CNIL reminds us of this, and OpenAI’s usage policies prohibit relying on the model for an assessment decision without a human in the loop.

Bias, Diversity, and Confidentiality

ChatGPT can reproduce biases present in its training data, and your vigilance should focus on three areas. First, cultural stereotypes: verify that generated examples do not convey biased representations, such as a series of statements where engineers are systematically men. Second, diversity of perspectives: the model tends to present one viewpoint as universal, so on a controversial issue you must explicitly ask for multiple documented readings; otherwise, you will distribute an opinion mistaking it for a synthesis. Third, linguistic accessibility: non-native French-speaking students obtain different results because an imprecisely formulated request produces a less useful response—the tool then widens the gap instead of closing it.

Confidentiality demands the same rigor. Never share student personal data with ChatGPT in a personal account: no named copies, no class lists, no identifiable individual situations. ChatGPT Edu offers specific guarantees in this regard: conversations are not used to train models, the administrator sets retention periods, and the institution has access to administration and compliance tools. These properties distinguish a defensible institutional use from a merely tolerated private use. In schools, the CNIL adds one more rule: a consumer tool must not force a student to create an individual account with their personal data.

Building Your Ethical Charter

A useful charter fits on one page and answers five questions. Transparency: when and how must students declare AI use? Authorization levels: which degree of assistance applies to each type of work, according to the spectrum described above? Attribution: in what form do you cite ChatGPT or any other tool, and where in the document? Responsibility: the non-negotiable principle is that the student remains responsible for the accuracy and quality of the content they submit, including any reference invented by the tool. Finally, sanctions: what are the consequences of non-compliance, and what is the procedure?

Do not write this charter in the abstract, where all formulations seem solid: start from a real assignment.

Your Turn

Your First AI Clause.

1. Take your next assignment
2. Write the clause that will accompany it, five lines maximum, specifying the authorized level and the expected form of declaration
3. Submit it to ChatGPT and ask for ambiguities a bad-faith student could exploit
4. Correct it, then print it on the assignment sheet itself
5. Plan a five-minute oral exchange for all submissions, rather than using a detector

Official Sources: How can educators respond to students presenting AI-generated content as their own? · Getting started with ChatGPT as an educator (OpenAI Help Center) · Enseignant : comment utiliser un système d’IA dans le cadre de vos missions ? (CNIL) · Publication du cadre d’usage de l’intelligence artificielle en éducation (Ministry of National Education)

Key Takeaways

• Define a clear usage spectrum rather than a binary prohibition
• AI detectors are not reliable, and the Ministry advises against using them: prioritize pedagogical approaches
• Never delegate grading or decision-making to the tool
• Never share student personal data in a personal ChatGPT account
• An explicit usage charter protects both teachers and students
• The student remains responsible for the accuracy of any content they submit, regardless of its origin