If You're Thinking About Banning AI, Read This First
A practical guide to clearer expectations, better assessment design, and honest conversations about AI in education.
When a Ban Actually Works
A ban on AI is legitimate and defensible when you control the assessment environment. That means you can directly observe what students are doing. Controlled environments include:
- In-person proctored exams
- Live oral examinations or viva voce
- In-class presentations with Q&A
- Lab demonstrations or practical skills assessments
- In-class written tasks with no device access
In these settings the environment removes the opportunity, not just the permission. The ban works because students simply do not have access.
Note for instructors in controlled environments
If students cannot access AI during the assessment, there is nothing to disclose. The category-by-category configurator is designed for uncontrolled environments where students have a choice. It may not be what you need here.
Why Bans Don't Work in Uncontrolled Environments
If your assessment is submitted remotely (essays, problem sets, reports, case studies, take-home exams), a ban is a request, not a guarantee. Two findings in the research are clear.
AI detection tools are not reliable
They produce high false positive rates, flagging human-written work as AI-generated. They are especially unreliable for non-native English speakers, neurodivergent writers, and highly fluent writers. Using them as enforcement tools undermines trust and can cause serious harm to students who have done nothing wrong.
(Perkins et al., 2024; Corbin, Dawson & Liu, 2025)
Bans without enforcement push use underground
Students who use AI in a banned context don't stop. They learn to hide it. Students who comply are disadvantaged. Disclosure is replaced by concealment. The conversation about responsible AI use never happens.
(Amlani & Davis, 2025)
Some responses change the rules. Better responses change the task.
Research from Corbin, Dawson and Liu (2025) draws a useful line between two kinds of responses to AI. The first is to add rules, guidelines, and policies that ask students to comply. The second is to change what the assessment actually measures and how.
Rules still matter. Clear communication still matters. But if the task itself can be easily outsourced to AI, clearer rules alone will not fix that.
The strongest approach combines both: make expectations explicit and design tasks where genuine thinking is hard to fake.
(Corbin, Dawson & Liu, 2025)
Why This Tool Still Matters
This tool is not a detection system. It will not catch cheaters, flag suspicious work, or enforce compliance. That is not its job.
Its job is to help you think clearly about how AI fits into your assessment, say that out loud to your students, and create conditions where honesty and reflection are easier than concealment.
Clearer expectations
In uncontrolled environments, students need more than a vague warning or a blanket ban. They need to know, category by category, what you expect and why. This tool helps you make those expectations specific and visible.
Honest communication
A policy that says "Discouraged" in an uncontrolled setting is more honest than one that says "Banned" but cannot be enforced. Honest framing invites honest responses.
Reflection and transparency
When students are asked to disclose their AI use, the goal is not surveillance. It is to prompt them to think about their own process: what they did, why, and what they learned. That reflection has real pedagogical value even when it is not graded.
Better conversations
When expectations are explicit, conversations about AI use actually happen. Students ask questions. Instructors learn how students are working. The topic moves out of the shadows and into the classroom.
A practical starting point
Not every instructor is ready to redesign every assessment. This tool gives you a way to start: set expectations clearly now, learn from what students tell you, and strengthen your approach over time.
This tool works best when paired with stronger assessment design, not used as a substitute for it. Clear expectations are necessary. But the most effective approach also rethinks what the task measures and how. The strategies below can help.
Assessment Redesign: What Actually Works
These strategies reduce over-reliance on AI. They are also, independently, better assessments. They test higher-order thinking more effectively and reward genuine engagement with course material.
Require personal experience and local context
AI has no access to a student's own life, classroom discussions, or course-specific content. Replace generic prompts with ones that draw on personal experience, in-class decisions, or locally specific scenarios.
Old
"What are the most promising trends in sustainable finance?"
New
"In our Week 6 discussion, we identified three trends most relevant to BC-based credit unions. Evaluate them using the framework from the assigned reading."
Stage your submissions
Progressive deliverables (topic proposal, annotated bibliography, draft, final submission) create a trail that makes wholesale AI outsourcing visible. They also give you opportunities to talk with students at each stage and assess the authenticity of their engagement.
(Amlani & Davis, 2025)
Use assessment twins
An assessment twin pairs an AI-vulnerable task (essay, report, problem set) with a verification task in a controlled or semi-controlled setting: an oral defence, a viva voce, a short in-class follow-up, or a group discussion with rotating facilitators.
The key principle is interdependence: the second task verifies the first. A student who outsourced the essay cannot defend it in conversation. The twin does not need to be elaborate. Even a 10-minute follow-up discussion strengthens validity considerably.
(Roe, Perkins, Giray et al., forthcoming. arXiv: 2510.02929)
Ask for reasoning, not just answers
AI produces plausible outputs but struggles to narrate authentic reasoning processes. Ask students to explain how they arrived at their conclusions, why they chose a particular approach, or what they would do differently given new information.
Design scenario-based questions with judgment calls
Questions that blend two domains, use ambiguous data, or require culturally contextualised judgment challenge AI tools while assessing higher-order thinking more effectively.
(Amlani & Davis, 2025)
Use presentations and live events
Presentations with Q&A, simulations, role plays, and debates require demonstration of understanding in real time. AI can help students prepare. It cannot perform for them in the moment.
A Better Signal Than the Final Product Alone?
Here is an idea worth considering: the final product may not be the only, or even the best, evidence of learning.
When students work with AI, how they interact with it can also reveal understanding. A student who asks sharp follow-up questions, challenges a weak AI response, spots an error in the output, or revises a draft with clear reasoning is demonstrating real thinking. A student who copies and pastes without engagement is not.
One emerging framework, called DRIVE (Oliveira et al., 2025), looks at two things: how well students steer and question the AI, and how much genuine expertise and subject knowledge they show in that interaction. In simple terms, the quality of a student's questions, critiques, revisions, and judgment may tell you more than the polished output alone.
What this means in practice
- This does not mean "grade prompt logs instead of assignments."
- It does mean that paying attention to process, reasoning, critique, and explanation can strengthen your picture of what a student actually understands.
- Asking students to reflect on their AI use, explain their choices, and describe how they evaluated AI output gives you a richer signal than the final product alone.
Where this tool fits: Our configurator helps you make expectations visible. The disclosure step asks students to reflect on what they used and why. Together they open the door to more thoughtful, process-based assessment.
The AI Assessment Scale: Choosing an Intentional Stance
The AI Assessment Scale (Furze, Perkins, Roe & MacVaugh, 2024) is not a compliance checklist. It is a practical tool for assessment design — a way to make a deliberate choice about what a given task is built to measure and where, if anywhere, AI use fits within that design.
It helps you think about three things: what you want students to learn, how AI might support or undermine that learning, and what the task is actually built to assess. Choosing a level is an architectural decision: it describes how the task was designed, not what students are permitted to do. An instructor choosing Level 2 (AI Planning) is saying “I designed this task around planning and ideation” — not “students are allowed to brainstorm with AI.” That framing changes how you build the task and how students understand the expectations.
If you wanted "No AI" but found "Discouraged" as the lowest option in the configurator, here is the distinction:
No AI / Controlled
AIAS Level 1 is for controlled environments. If your assessment is in-person and proctored, that level applies. In the configurator, setting all 15 categories to N/A reflects that reality. The category-by-category framework is designed for uncontrolled settings where students have a choice.
Discouraged / Uncontrolled
"Discouraged" in an uncontrolled setting is honest. It tells students: we would prefer you do not use AI here, here is why, and if you do anyway, disclose it. That is more useful than "banned" in a setting where you cannot enforce it.
The full scale maps the spectrum from No AI through to AI Exploration. It helps educators think through what each task is built to assess and whether that holds up when students have AI access. Clearer conversations with students about expectations follow from that design clarity.
Explore the full AIAS frameworkFurther Reading
Assessing Students' DRIVE: A Framework to Evaluate Learning Through GenAI Interactions
Oliveira, M. et al. (2025).
An emerging framework that examines how students question, steer, critique, and refine AI outputs as possible evidence of learning. Useful for instructors interested in assessing process, judgment, and visible reasoning, not just final products.
Assessment Twins: A Protocol for AI-Vulnerable Summative Assessment
Roe, J., Perkins, M., Giray, L. et al. (forthcoming). arXiv: 2510.02929
A framework for pairing AI-vulnerable tasks with verification tasks to strengthen validity without abandoning established assessment formats.
arxiv.org/abs/2510.02929 →Note on misconduct: Assessment twins strengthen validity, but they do not eliminate academic integrity risk. If a student uses AI on a banned take-home task and cannot defend their work in the follow-up, that is still a misconduct situation — one your institution's academic integrity processes are better equipped to handle than any assessment design. Structural changes narrow the gap; they do not close it.
Talk is Cheap: Why Structural Assessment Changes Are Needed for a Time of GenAI
Corbin, T., Dawson, P. & Liu, D. (2025). Assessment & Evaluation in Higher Education, 50(7), 1087-1097.
Draws the line between rule-based responses and structural changes to assessment design. Explains why policy alone cannot restore validity when the task itself is outsourceable.
doi.org/10.1080/02602938.2025.2503964 →The AI Assessment Scale Revisited: A Framework for Educational Assessment in the Age of Generative AI
Perkins, M., Roe, J., & Furze, L. (2025). Journal of University Teaching & Learning Practice.
An updated framework for redesigning assessments in light of generative AI. Each level describes how a task was designed to be assessed — from No AI to AI Exploration — with practical guidance for applying the scale to task construction.
open-publishing.org →ChatGPT Ate My Homework: What Educators Need to Know About Generative AI
Amlani, A. & Davis, P. (2025). Canadian Edition.
The disclosure framework this tool is built on. Chapters 8 and 9 cover assessment redesign and reimagining education in depth.
GenAI Detection Tools, Adversarial Techniques and Implications for Inclusivity in Higher Education
Perkins, M., Roe, J., et al. (2024). arXiv: 2403.19148
The evidence base for why AI detection tools are not reliable enforcement mechanisms, with particular attention to equity impacts.
arxiv.org/abs/2403.19148 →Ready to design your AI expectations?
The configurator helps you set clear, category-by-category expectations for your students. It is a starting point for better conversations, better assessment design, and more honest engagement with AI.
Build Your Policy