STRAND C · ENHANCE | STRAND B · DEEPEN

Grade the AI against your rubric

Take one assignment you already give. Hand a machine the exact instructions your students get, grade what comes back on your existing rubric, and use the number. That single score tells you what your task actually measures now — and it is the difference between guessing at an AI position and deriving one from evidence.

A STARTING POINT — AND IT NEEDS YOUR JUDGMENT

This page is a draft, and I’m still iterating on it. It wasn’t written by an expert in your field — it is here to start a pathway you then shape to fit you: your subject, your level, your students, your department. Your expertise runs far deeper than anything here, and none of this is finished until your judgment has been on it.

That is the whole point. Judgment is the most important part of this — yours, and the judgment you are building in your students. These pages exist to help you develop and exercise it, never to replace it. Take what is useful, change what is not, and tell me where I have gotten it wrong — that is how these get better.

THE MOVE

Run your own real assignment through an AI, grade the output blind against your own rubric, and mark which criteria it earned and which it missed. Fifteen minutes, one task.

WHAT IT GIVES YOU

A defensible read on whether the task measures the artefact or the thinking — and a position for your policy that describes what your assessment actually does rather than what you hoped it did.

Why this rung sits in two strands at once. As a piece of task design (Strand C), grading the AI shows you what your assessment still requires when the student has every tool on the desk. As a piece of policy (Strand B), it is how the position you write into your syllabus stops being a hopeful guess and becomes a claim you have tested. Same fifteen minutes; two things it earns you.

WHY THE SCORE IS WORTH HAVING

A grade is a claim about a student: this score means she can do this thing. But you can only trust that claim if the task measures the thing you think it does — and right now the fastest way to find out is to let a machine sit the assignment and see what your own rubric gives it.

If the machine earns full marks, the task was measuring the artefact, not the thinking. That is not a disaster and it is not an accusation — it is information, and you now have it before your students do rather than after. A task that a first-draft AI answer earns a 95 on is a task whose grade no longer separates the student who reasoned from the student who pasted. You would rather learn that on a Tuesday in August than in a stack of April essays.

This is also the honest way a policy position gets set. The hub's Strand B, Deepen rung is exactly this: the position came from evidence — you ran your own task against AI, graded the result on your own rubric, and set the position knowing what it scored. Before you have that number you are guessing which of the four positions a task should carry. After it, you are deciding. And deciding is the whole of Strand C, Enhance: knowing what your task produces under AI, then choosing what to do about it.

HOW TO DO IT — FIFTEEN MINUTES, ONE ASSIGNMENT

Nothing here requires a new tool, a new account, or telling anyone. Pick a real assignment and work through these in order.

Pick one real assignment. Something you actually assign and actually grade — not the hardest one and not the one you already suspect is fine. The one whose result you can't predict is the one worth the fifteen minutes.
Give the AI the exact prompt a student receives. The instructions, the prompt, the source text, the word count — verbatim, nothing added, nothing clarified. If you improve the instructions for the machine, you are testing a task your students never see.
Grade the output blind against your existing rubric. Your rubric, unchanged. Score it the way you would score a paper with no name on it — and if you can, score it before you remind yourself it came from a machine, because knowing tends to make us harsher than the rubric says.
Mark which criteria it earned and which it missed. This is where the real information lives. Not the total — the pattern. Full marks on structure and evidence-handling, nothing on the criterion that names your actual construct? That gap is the map of what your task still protects and what it no longer does.
Then decide — one of three.Accept the task (the artefact really is the point, or the AI didn't earn the grade). Change the evidence you collect so the shortcut can't produce it. Or add a step the machine can't take for the student. The decision is yours; the number just makes it an informed one.

Where each decision goes next.

THREE HONEST LIMITS

This move is genuinely useful and it is easy to oversell. Three cautions, so you read the number for exactly what it is.

IT VARIES RUN TO RUN

The same prompt produces different output each time. One run is an anecdote. Do it two or three times and grade each — you are looking for the range, not a single verdict, and the range is itself worth knowing.

THERE IS NO VALIDATED RUBRIC FOR THIS

You are scoring AI output on a rubric built for students, and no research offers a tested instrument for appraising a machine's work. This is professional judgment, not measurement — which is fine, because judgment is the expertise. Just don't hand the number a precision it doesn't have.

A HIGH SCORE DOESN'T CONDEMN A TASK

Sometimes the artefact is the point — a student needs to produce the clean lab write-up, the correct proof, the finished translation, and it doesn't matter that a machine can too. A high AI score is a prompt to ask “is the artefact what I'm grading?”, not an automatic order to redesign.

And the standing caveat. Everything above is a starting point — a way to generate evidence, not a verdict handed down. Your rubric and your judgment are the authority here; the machine's score is an instrument you point at your own task, and what it measures depends entirely on the rubric you already trusted. If the number and your read of the task disagree, that disagreement is the finding worth chasing, not a reason to overrule yourself.

Bring one assignment to me and we'll grade it together — fifteen minutes. You pick the task, we give the AI the exact instructions your students get, and we score it on your rubric while you watch what it earns and what it misses. You'll leave knowing what that task measures now, and with a position you can actually defend.

Jason Beyer · Educational Technology · Regis Jesuit High School · A.M.D.G.

HelpDesk@regisjesuit.com · 303.269.8080 · LAST VERIFIED — AUGUST 2026