Where Assessment Breaks
The graded artifact stopped being evidence of learning. Why detection and bans are losing strategies, and the shift from grading the product to grading the thinking that produced it.
The Broken Assumption
Almost every assessment ever built rests on one quiet assumption:
The artifact the student hands in is evidence of the thinking the student did.
A good essay meant a mind had wrestled with the argument. A correct problem set meant someone had worked the method. The product was a reliable proxy for the process. Generative AI severed that link. A polished artifact is now equally consistent with “a student thought hard” and “a student typed one sentence into a model.” The proxy broke, and grading practices that depend on it broke with it.
There are three common responses. Two of them lose.
The Two Losing Responses
Detection. Run submissions through an AI detector and penalize the flagged ones. This fails on every axis: detectors produce false positives that wrongly accuse real students, false negatives that miss lightly-edited AI, and they create an adversarial arms race you can’t win against models that improve monthly. Worse, it’s a strategy fought below the line — it tries to preserve lower-order assignments by policing them, instead of asking whether those assignments still measure anything.
Banning. Forbid AI, move everything to blue-book exams and locked-down rooms. This can restore the product-process link by brute force, but at a steep cost: it walls students off from the tool they’ll use in every job for the rest of their lives, and it usually pushes the assessment down the ladder (handwritten recall under time pressure), which is the opposite of the evolutionary leap. You’ve protected the proxy by shrinking the ambition.
Both responses share a hidden goal: keep the old assignments and defend them against AI. That’s the losing frame. The old assignments were proxies, and the proxy is gone. You don’t defend a broken proxy — you replace what you’re measuring.
Detection and honor codes are integrity tools — they’re about catching rule-breaking. They are not design tools, and they can’t substitute for design. If your only response to AI is a stricter integrity policy, you’ve answered “how do I stop cheating?” while leaving “what should this course now measure?” completely untouched. This course is about the second question. A well-designed higher-order assessment reduces cheating as a side effect — because outsourcing the thinking produces visibly worse work — but that’s a consequence of good design, not the goal of it.
The Winning Response: Grade the Thinking
The third response is the one this module builds: stop grading the artifact as your evidence, and start grading the thinking — made visible through the process, the judgments, and the student’s ability to defend and extend the work.
This is a genuine shift in what a submission is. Instead of “here is my finished essay,” the submission becomes some combination of:
- The reasoning trail — the choices the student made and why, including where and how they used AI and what they did with its output.
- The judgment — a defended position the student owns and is accountable for, not a neutral summary.
- The transfer — the same skill applied to a fresh context where a canned answer doesn’t fit.
- The critique — the student catching what the AI (or a source) got wrong.
Each of these lives above the line, and each is hard to fake precisely because faking it requires doing the very thinking you’re trying to elicit. That’s the design property you’re hunting for: the cheat path and the learning path converge. When the fastest way to get a good grade is to actually think, you’ve stopped fighting AI and started using the assessment to build cognitive agency.
Reframing the Goal
Hold this contrast as you move into the next lesson:
| Old goal | New goal |
|---|---|
| Detect who used AI | Design so AI-use produces worse work if the student didn’t think |
| Grade the finished product | Grade the reasoning, judgment, and transfer |
| Protect the assignment from AI | Replace the proxy with evidence of thinking |
| Prevent cheating with policy | Reduce cheating as a side effect of good design |
The next lesson gives you a concrete, repeatable pattern — a five-step loop — for turning any AI-vulnerable assignment into one where the graded evidence lives in the student’s thinking.
Two instructors respond to AI. One buys a detection service; the other rewrites the final so students must critique the AI's own answer and defend a correction. Why is the second response structurally stronger?
Reveal Answer & Explanation
The first defends a broken proxy from the outside and loses an arms race; the second changes what’s measured so the cheat path and the learning path converge.
Detection tries to preserve the assumption that the artifact equals the thinking — the exact assumption AI severed — and it does so with a tool that has false positives, false negatives, and a monthly-improving adversary. The rewrite abandons the broken proxy: to get a good grade, a student now has to actually understand the material well enough to find the model’s error and defend a fix, which is the learning objective. It reduces cheating not by catching it but by making the thinking the only path to the grade.
Up Next
The five-step loop: a step-by-step pattern — attempt, question, critique, transfer, reflect — for converting any assignment into one that grades thinking instead of artifacts.