You don't have an AI detector problem.
You have an assessment design problem.
In international schools, that shows up in a familiar scene.
A thoughtful student submits an eerily polished response.
Your gut says, "This doesn't sound like them."
You either start policing.
Or you start redesigning.
The schools that are staying sane are choosing redesign.
Not because they love change.
Because AI has made output cheap.
And cheap output breaks any assessment that only rewards the final product.
The real risk isn't cheating. It's invisible thinking.
When students use AI without structure, they don't just "get help."
They can offload the thinking.
The OECD names this risk directly: students may lean on AI tools, "offloading knowledge acquisition and problem-solving," which can lead to lower proficiency and dependency—especially for students who have less access to structured learning strategies at home.
That's the shift.
Your job is no longer to ask:
"Did AI write this?"
Your job is to ask:
"Where is the student's thinking, judgment, and voice visible?"
Because if you assess what AI can easily generate, you end up grading AI access.
Not learning.
A better stance: "AI allowed—if you can prove ownership."
One of the most practical moves I've seen is this:
Stop writing "AI prohibited" policies that nobody believes.
Replace them with a short "AI allowed" statement that makes students accountable for process.
Here's a clean version (adapt it to your school):
Students may use AI for brainstorming, planning, and feedback only if they:
- Disclose where and how AI was used (tools, prompts, purpose)
- Verify any factual claims with acceptable sources
- Demonstrate ownership through reasoning, revision logs, or a live performance
This aligns with UNESCO's direction: institutions must validate and guide GenAI use to protect human agency and ensure ethical, pedagogically appropriate practice.
That one change does something powerful.
It shifts AI from a secret shortcut into a visible learning tool.
We've been here before (calculators didn't "ruin math")
If this feels like a crisis, it's because we're in the messy middle of tool adoption.
Calculators triggered the same fears:
- Students will rely on machines
- They won't learn fundamentals
- They'll cheat
Then reality kicked in.
Once students had calculators at home, schools couldn't "ban" them out of existence.
So educators had to redesign instruction and assessment around what calculators changed.
Generative AI is the same story, with higher stakes.
The answer is not stronger policing.
It's smarter assessment.
The AI-Amplified Assessment Playbook (a simple 4-step method)
Here's the system that makes this actionable without overwhelming your department.
Step 1: Decide what you actually want evidence of
If your standards say "critical thinking," don't assess only essay fluency.
Choose one of these evidence targets:
- Reasoning under questioning
- Use of evidence
- Quality of revision decisions
- Critical evaluation of sources and claims
- Judgment under constraints
- Transfer (apply learning to a new context)
Once you name the target, you can design the assessment so AI can help prepare—but cannot perform the thinking.
Step 2: Add a "performance moment"
A performance moment is any time the student must show thinking live or semi-live:
- oral defense
- seminar contribution grounded in evidence
- teaching a mini-lesson with cold questions
- roleplay simulation
- debate rebuttal
This is where authenticity becomes obvious.
The OECD explicitly notes oral assessment practices can help uphold academic integrity in the age of AI because live questioning can't be outsourced to current AI tools.
Step 3: Make process visible (not painful)
You don't need 20-page portfolios.
You need lightweight artifacts that reveal thinking:
- prompt journals
- change logs
- AI error audits
- evidence ladders
This is how you prevent the "AI did it" black box.
Step 4: Grade with one shared lens: the quality of reasoning
A common staff fear is consistency:
"If we redesign, won't it be subjective?"
Use a reasoning quality frame like Paul–Elder's intellectual standards (clarity, accuracy, precision, relevance, depth, breadth, logic, fairness) as your shared language for rubrics and feedback.
It gives departments a common vocabulary without forcing a single assessment format.
10 AI-amplified assessment formats (practical, school-friendly)
Below are the formats you can rotate across a unit.
Pick two to start.
1) Micro-viva (short oral defense)
Students submit the product.
Then they defend the thinking.
AI can help them prepare.
AI can't answer your probing follow-ups.
Best for: deep understanding, authenticity, ownership.
2) Fishbowl seminar + evidence tickets
Students can use AI to generate questions or prep summaries.
But every claim in the discussion must be tied to a specific text/data reference ("evidence ticket"), or it doesn't count.
If you want a quick visual refresher on fishbowl discussion structure, Edutopia's "60-Second Strategy: Fishbowl Discussion" is a useful clip to share with students.
Best for: synthesis, listening, academic discussion norms.
3) Debate + AI opposition research log
Students use AI to build the strongest opposing case.
Then they document:
- prompts used
- best counterarguments AI produced
- their rebuttals and evidence
And then they debate live.
Best for: reasoning in real-time, communication, preparation discipline.
4) AI-assisted draft + human revision (graded)
You don't grade whether they used AI.
You grade the improvement decisions.
A simple weighting works well:
- final quality
- change log + reasoning
- reflection on revision choices
Best for: writing, process, metacognition.
5) AI error audit (critical thinking check)
Give students an AI-generated explanation.
Their job is to find what's wrong:
- factual errors
- logical gaps
- missing context
- fake citations
Then correct it with sources.
This pairs naturally with Paul–Elder standards ("accuracy," "logic," "depth") to structure critique.
Best for: media literacy, critical evaluation, epistemic humility.
6) Prompt journal (process portfolio, lightweight)
Weekly log:
- prompts used (exact text)
- what the AI said (summary)
- what the student accepted/rejected and why
- what they learned
Best for: preventing cognitive offloading, building AI literacy.
7) Roleplay simulation (AI as stakeholder)
AI plays a client/patient/parent/student.
The learner must ask questions, decide, respond.
Then reflect.
Best for: applied skills, decision-making, SEL, languages.
8) Micro-teach + cold questions
Students teach a concept for 5 minutes.
Then face 3 minutes of unscripted questions.
This reveals understanding fast.
Best for: conceptual clarity, transfer, communication.
9) Constraints-first design brief
Students design a solution under constraints (budget, time, users, accessibility, privacy).
They are graded on tradeoffs and justification, not "the perfect answer."
Best for: design thinking, business, science, ethics.
10) Evidence ladder (source credibility + nuanced claim)
Students rank sources (systematic reviews → peer-reviewed → primary → secondary → opinion).
Then state a position with appropriate certainty ("evidence suggests…").
Best for: research skills, TOK-style reasoning, academic writing.
This reduces conflict with families and students
When your policy is "AI is banned," you create a trap.
Because students know they can use it.
Parents know they can use it.
And teachers know they can't reliably prove anything.
So the whole community ends up in a low-trust game.
When your policy is "AI is allowed—with disclosure, verification, and ownership," you can have a clean conversation:
"We're not punishing tool use.
We're assessing learning."
That's healthier.
And it scales across cultures and curricula in international schools.
(And it lines up with UNESCO's call for institutional responsibility and human-centred deployment.)
Imagine 30 days from now
You're in a moderation meeting.
No one is arguing about who used AI.
You're looking at:
- defense answers
- evidence tickets
- revision reasoning
- error audits
- student reflections
You can see thinking.
You can coach it.
You can grade it fairly.
That's not fantasy.
That's what happens when assessment stops rewarding "polished output" and starts rewarding "visible reasoning."
Getting started
Pick one upcoming assessment.
Then do this:
- Add a 3-line "AI allowed if…" statement (disclosure, verification, ownership).
- Add one performance moment (micro-viva, cold questions, seminar evidence tickets).
- Add one process artifact (prompt journal entry OR change log OR error audit).
That's enough to shift the culture.