“Did you use AI?” is the wrong question.
AI use has already become routine in college life: 57% of college students use AI tools in their coursework daily or weekly, even as many campuses still lack clear guidance.1 Institutional policy is stuck in a detect-and-punish loop. The actual problem we have is one of trust calibration, not integrity. When we treat AI as a binary yes/no threat, we break the loop between teaching, feedback, and evidence.
Don’t ask if they used AI. Ask why they didn’t use it well.
Many universities still rely on a punitive model of academic integrity: they set standards, deploy detection software, and route suspected cases into misconduct processes.2 The problem is that AI detectors are unreliable — they produce false positives, flag non-native speakers at disproportionately higher rates, and cannot distinguish between a student who cheated and one who just didn’t know where the line was.3 You are building enforcement infrastructure around a tool that doesn’t work, for a problem it can’t solve.
AI use in academic work is many things: it’s 1) a cheating tool, 2) a legitimate support, and 3) an accessibility resource — sometimes all three in the same assignment, for the same student.4 Detection software collapses it into one verdict.
And students who end up in misconduct referrals are often the ones who needed the most support and received the least.2 The students who knew the rules — or knew someone who knew the rules — managed just fine.
If your integrity strategy starts at detection, you've skipped past the learning itself.
A student who submits AI-generated work they can't defend has trusted an output they were not equipped to evaluate.
AI's surface confidence is exactly what makes it hard to evaluate — it states a wrong answer with the same steadiness as a right one. A student who can’t accurately judge whether an AI output is correct, appropriate, or sufficient isn’t necessarily trying to cheat. They may simply not know. They're confident in a product they don't understand.
It’s a failure of calibration — how confident you are versus how correct you actually are. Researchers call this metacognitive sensitivity:5 a skill that can be taught, practiced, and assessed.
Can the student explain and defend the work — and extend it — without AI?
Your assessments should be designed to answer this. Your integrity policy tells students what's off-limits, but it can't give them the judgment to evaluate what the AI handed them. That has to be built into the task.
Good learning design has always worked as a sequence: learners receive feedback on their developing work, they then use that feedback to improve, and ultimately produce something that demonstrates understanding. The loop works when the learner is actually inside it — processing, adjusting, thinking.
Image alt. text: Flowchart of the classroom feedback loop: Draft (D2L) → Draft (Instructor Feedback) → Final Submission (D2L) → Final Submission (Feedback and Grade).
AI breaks that sequence quietly. A student uses AI to generate a draft, receives instructor feedback on it, submits a revised AI draft. They completed every step and learned nothing. The loop ran, but there was nothing inside it.
Research confirms what most faculty already sense: students using AI for writing tasks can produce stronger final products while retaining significantly less of the underlying knowledge — and students report reducing their own cognitive effort in the process.6,7 The work gets better while the learner's grasp of it doesn't move.
Stricter assessment requirements don't address the problem. We need to design for the evaluation step that AI skips — and make that step visible, gradable, and part of what the learning experience is actually for. An assessment that accepts output without process ends up measuring the tool.
Students who struggle likely had the least institutional scaffolding going in — never modeled good academic practice, never shown where the lines were, and were never given a safe space to get it wrong before the stakes were high.2
Meanwhile, male students, STEM students, and students from more advantaged socioeconomic backgrounds use AI at higher rates than their peers.⁹ They arrive already familiar with the tools, and familiarity tends to bring confidence about where the lines are. Applied to an uneven playing field, a zero-tolerance policy produces inequity rather than integrity.
Research on scaffolded learning and self-determination theory is useful at this point: educate first ➡ enable practice ➡ then expect accountability. In that order. Reversed, the sequence mostly catches the learners who were let down before they ever opened the tool.2
In most institutions, the people who set the academic integrity policy are not the same people who design the assignments. Policies get written without the assignments in view, and the failures collect there. Closing that gap requires both sides — and a shared language for what responsible AI use looks like in practice.
The good news: you don’t need a new course, a new policy, or a new technology stack. You need three design moves.
None of these are surveillance tools; they're scaffoldings that build cognitive trust before the stakes are high.
In a classroom built on cognitive trust, responsible use becomes the path of least resistance, and policing largely takes care of itself.
Most rubrics assess what a learner produced. This one assesses how well they evaluated what they used to produce it. The three dimensions below map onto constructs from existing integrity and assessment frameworks.8
|
Dimension |
What You’re Actually Grading |
|
Transparency |
Did the learner accurately disclose what AI contributed? |
|
Verification |
Can they identify what they checked, corrected, or changed? |
|
Defensibility |
Can they explain and extend the work without the AI present? |
Start with one assignment. Add the process note requirement. Score the three dimensions alongside your content criteria. Defensibility is the load-bearing dimension — if those scores don’t track with Transparency and Verification, the process note has become a form, not a learning tool.
That’s the outcome integrity policy was always trying to protect. Now it’s designed in.
➡ Download the full AI Trust Calibration Rubric
The academic integrity conversation almost always arrives after the fact — a policy update, a misconduct spike, a faculty member asking what to do about “the ChatGPT problem.” By then, the design window has already closed.
The better move is to build for cognitive trust before the first assignment goes out. When the learning strategy connects feedback to evidence through explicit process requirements, responsible AI use turns from a rule students follow into a skill they actually build.
That’s the difference between designing an integrity policy and designing an integrity environment. This is part of a continuing series exploring how AI changes what learning design must make visible:
Subscribe to the LX Studio newsletter for research-grounded insights on designing learning environments that perform in the age of AI.
➡Explore our recent launch of AI Learning Services
View our resource library that inspired The Human Side of AI in Learning Series.