Assessment models for a course with AI

Low-stakes ways to see what a student can do unaided

The redesign and the assignment templates both depend on one thing: some way to see a student’s judgment directly, without the tool, or you end up grading the AI. This page describes how to do that. None of it requires a return to high-pressure closed exams. The aim is the opposite, a set of small, frequent, low-pressure checks that show you where each student’s unaided competence actually is, early enough to help.

A useful way to think about these is that they measure competence directly rather than policing how work was produced. That makes them fairer than detection. A student is judged on what they can do, not on a guess about whether a tool was involved, which is exactly the standard the book argues for when it warns against leaning on unreliable detectors.

Model 1: Short in-class judgment checks

What it is. A brief exercise, five to ten minutes, where students judge a provided solution or piece of output without tools. You show a finished analysis, a short piece of code, or a worked solution, and ask whether it is right, where it fails, and what they would check. It is the audit assignment in miniature, done in the room.

What it reveals. Whether the critical judgment the course is meant to build is the student’s own. A student who can catch the planted flaw on paper has the competence; one who cannot has told you what to work on next.

How to run it. Any class size works, on paper or with whatever response system you use. Keep each one short and low-stakes, and run them often, so they feel like practice rather than examination. Vary whether the solution you show is sound or flawed, so students actually judge rather than assume a flaw is always present.

Model 2: Brief oral defenses of submitted work

What it is. A short conversation, two to five minutes, about something a student has already handed in. You ask them to walk you through why they judged an analysis sound, or to explain a part of their own solution, or to say how they know a result is right.

What it reveals. This is the most direct check of all, because it shows whether the understanding behind a submission is actually the student’s. Work done with heavy AI help that the student cannot explain shows immediately, and work the student genuinely understands holds up under a simple question.

How to run it. In a small class you can reach everyone over a term. In a large one, sample: a rotating subset each week, chosen so every student expects to be asked eventually. A few minutes in office hours, in lab, or at the edges of class is enough. The standard is simple, can the student explain the key judgment behind their own work, and you are listening for understanding rather than polish.

Model 3: Explain-your-reasoning writing

What it is. A short piece of writing in which a student explains why they chose an approach and how they know the answer is right, produced in conditions where the tool is not the author, for instance written in class or as a quick reflection attached to a submission. It is brief by design, a few sentences of reasoning rather than an essay.

What it reveals. The sanity-checking and method-selection that survive even when a tool does the mechanics. A student who can say why an approach fits and what makes the answer trustworthy has the judgment the course is for, whatever produced the calculation.

How to run it. Attach it to assignments as a required short note, or run it occasionally in class. Grade it lightly and for substance, so it stays a low-pressure habit rather than another thing to perform.

A note on closed problems

You may still want an occasional short, unaided problem to confirm a foundational skill that has to be automatic. That is fine, kept low-stakes and infrequent. The point of all of these is not to recreate the high-pressure exam, but to see competence often enough and gently enough that both you and the student can act on what you find.

Fitting this in

Choose one or two of these to start, rather than all of them, and run them regularly. Tie what you check to the unaided core you named in the redesign, so the checks measure the thing that matters. When a check reveals a gap, treat it as information rather than a verdict: it tells you and the student where the understanding is thin while there is still time to build it. Used this way, these checks do more than grade. They keep students honest with themselves about what they can actually do, which is the habit the whole course is trying to form.

The assignment patterns in instructors/assignment-templates.md pair naturally with these checks, and the study sessions in students/study-sessions/ show the kind of judgment the checks are looking for.