The Detection Problem
2 min readAI detection tools have emerged that offer to identify AI-generated content in student submissions. These tools feel like a straightforward solution.
But here’s the problem: they don’t work reliably.
Detection tools produce false positives (flagging human writing as AI-generated) and false negatives (missing AI-generated content). More importantly, there are serious equity concerns. Research shows these tools disproportionately flag multilingual writers and students whose writing doesn’t match dominant linguistic norms.
Even when detection tools work as intended, they still miss the deeper issue: a student who submits AI-generated work might be caught, but the assessment itself hasn’t changed. They’ve simply demonstrated that they can get away with (or not get away with) a shortcut.
Echoing the assertions of Corbin, Dawson, Liu, and others the 2026 Assessment Redesign Framework makes a bold claim: detection is not the answer. Redesign is.
Rather than trying to police student behaviour, we should redesign assessment so that learning is visible and shortcuts become more difficult.
False Positives
Human-written work incorrectly flagged as AI-generated, causing unwarranted misconduct investigations.
False Negatives
AI-generated content that passes undetected, giving a false sense of security.
Equity Concerns
Tools disproportionately flag multilingual writers and non-dominant writing styles.
Doesn’t Address Root Problem
Detection polices behaviour but doesn’t change the task itself. The shortcut remains available.
Two Approaches
2 min readThere are two fundamentally different ways to respond to GenAI in assessment.
Approach 1: Discursive Control
This means creating rules about what students can and can’t do. For example:
- ‘You may not use ChatGPT for essays’
- ‘AI use must be disclosed’
- ‘Breaches will be addressed through formal academic misconduct procedures.’
These rules communicate clear expectations. But here’s the catch: rules rely on compliance. Students remain free to follow or ignore them. You might catch some with detection software, but you can’t reliably enforce an expectation that students remain free to ignore.
Approach 2: Structural Redesign
This means changing the actual mechanics of the assessment task itself. For example:
Instead of: ‘Write a 2,000-word essay analysing a research topic’ (easily AI-generated)
Redesign to: ‘Submit a proposal (week 5) → annotated draft (week 8) → final essay (week 10) + one 15-minute oral defence where you explain your key decisions and respond to challenging questions’
In this redesigned task, learning is visible at every stage. The student must show how their thinking evolved. They must articulate their reasoning in real-time, face-to-face. AI can’t show up to an oral defence.
This is structural redesign: the task itself now makes it harder to submit work that isn’t genuinely their own.
Why This Matters
1 min readThe shift from detection to redesign has a profound implication: assessment redesign is not primarily about fighting cheating. It’s about improving learning.
When assessments are redesigned to make learning visible — through staged tasks, oral interactions, reflective elements, and authentic problems — something shifts. Students can’t just submit a final product; they must show their thinking. That visibility strengthens the validity of the assessment AND makes it harder to fake understanding.
This framework rests on five core principles that guide this kind of redesign:
Throughout this course, you’ll return to these principles as you redesign your own assessments.
Activity: The Scenario
3 minScenario: Your institution has decided to disable AI detection tools. In a faculty meeting, your colleague suggests three responses:
‘We should add a strict rule: students who use AI will face misconduct proceedings. We’ll monitor submissions carefully.’
‘We should require students to disclose any AI use in a signed statement.’
‘We should redesign our assessments. Instead of one final essay, we’ll have students submit a proposal, an annotated draft, and a reflection on feedback before the final submission. We’ll also add a brief oral component where they explain their key choices.’
Select an option to see feedback. There’s no wrong answer — this is about exploring different approaches.
Approach A: Discursive Control
This approach relies on rules and monitoring — discursive control. The problem: if a student is determined to use AI, the rule alone won’t stop them. Detection tools might catch some, but as we discussed, they’re unreliable and have equity concerns. This approach addresses the symptom, not the root issue.
Approach B: Self-Reporting
This approach asks students to self-report. It’s better than A because it builds some responsibility. But it still relies on compliance. A student who wants to misuse AI can simply not disclose it. Without changing the task itself, you’re still hoping rules will work.
Approach C: Structural Redesign
This approach is structural redesign. By changing the task mechanics, you’ve made learning visible at every stage. The student must show how they thought through the problem (proposal), how they refined their thinking (draft + reflection), and how they can defend their choices (oral). This makes it much harder to submit work that isn’t genuinely theirs — not because of a rule, but because the task itself requires authentic engagement.
Found something interesting? Feel free to pause and discuss this scenario with your module team before continuing.
Knowledge Check
Based on what we’ve discussed, what’s the key limitation of using AI detection tools as a primary strategy for academic integrity in the age of GenAI?
✓ Correct
Detection tools have significant technical and equity limitations. But equally important: even when they work, they only police behaviour. They don’t change what students are being asked to do. The framework argues for a different approach: redesign the task itself so that authentic learning is required.
Not quite
The real issue is deeper than cost, time, or cleverness. The core problem is that detection tools can’t reliably work (false positives and negatives), and more fundamentally, they don’t address the root issue: the task itself might encourage shortcuts. What if, instead of trying to catch students, we redesigned what we’re asking them to do? That’s what this framework focuses on.
Now that you understand why detection alone isn’t the answer, the next step is to figure out where to start. Which of your assessments are most vulnerable to inappropriate AI use?
That’s what we’ll explore in Module 2.
Next: Assess Your Risk →