Teaching when every student has AI

The fix isn’t a better detector — it’s work whose thinking a chatbot can’t do for the student.

The idea

If a task can be finished by pasting the prompt into a chatbot, it now measures the chatbot. So you redesign the task around two questions: can I see the student’s thinking, and could a bot alone pass this. Moving work in-class, adding an oral defense, and requiring visible drafts all raise the first and lower the second. And instead of policing with unreliable detectors, you set disclosure norms — allow AI where it helps, ask students to say what they used — so honest use surfaces instead of going underground.

How it works

Run every task through two questions, then reach for the moves that fix a weak answer:

Q1  Can I SEE the student's thinking?   drafts, notes, an oral defense
Q2  Could a chatbot alone PASS it?      if yes, it grades the tool

Design moves, strongest first for "can't delegate":
  ORAL DEFENSE       explain your reasoning, answer two follow-ups
  IN-CLASS / PROCESS do it here; show outline, sources, versions
  DISCLOSURE NORMS   allow AI, require a line on what you used & how

Where AI helps vs. takes the thinking away:
  HELPS   brainstorming, feedback, formatting, a rough translation,
          scaffolding for a struggling student
  TAKES   when constructing the argument or doing the derivation
          IS the skill you're assessing

The no-proof suspicion talk. Detectors are unreliable, so you often can’t prove it. Don’t accuse — ask the student to walk you through their drafts and explain a paragraph. If the assessment already required process and defense, you rarely need this talk at all: understanding was demonstrated, not litigated.

When to use it

Move	Trade-off to weigh
Oral defense — highest signal on real understanding	Time-intensive; hard to scale to 150 students; keep questions equitable
In-class / process-visible — kills real-time outsourcing	Limits research depth; plan for accommodations and test anxiety
Visible drafts — shows the thinking develop	Trails can be faked; adds grading load
Disclosure norms — builds honesty over a gotcha culture	Only works if you genuinely allow AI somewhere and teach the line

Watch out for

• Grading on an AI detector. They’re unreliable and misfire on multilingual and neurodivergent writing. A detector score is not evidence; never fail a student on it.
• Accusing without proof. Turn a “gotcha” into a process conversation — drafts and an explanation — not a verdict you can’t support.
• A flat ban. Banning AI drives it underground, forfeits the chance to teach judgment, and can’t actually be enforced.
• Oral defenses for everyone, no plan. 150 five-minute defenses is unworkable. Sample it, stagger it, or keep them short and targeted.
• Treating disclosure as enough. “Allowed with disclosure” keeps people honest but doesn’t prove the student thought. Pair it with process or defense.
• Assessing the artifact, not the thinking. A polished essay proves nothing about who did the reasoning. Grade the reasoning you can see.

Worked example

In a curriculum-design interview: “Your take-home essay is now trivially AI-completable. Redesign the unit.” A strong answer doesn’t reach for a detector. It keeps the essay as a thinking artifact but requires an outline, annotated sources, and revision history alongside it; adds a five-minute oral defense on two of the student’s own paragraphs; permits AI for brainstorming and feedback with a one-line disclosure; and moves the final synthesis paragraph in-class. Watch the dials: evidence of thinking climbs into “clearly visible” and AI-fakeable drops toward zero — and because AI is allowed honestly, no student needs to hide it and no teacher needs a gotcha.

Check yourself