Grading AI-Permitted Assignments: How to Assess Work When Tools Are Allowed

You wrote it into the syllabus: students may use AI tools on this assignment. Then the first batch lands in your inbox and you realize you never decided what you’re actually grading. If the polish is free now, what’s left to assess?

Plenty, it turns out. But it isn’t the thing you used to circle in red pen.

Key takeaways

• When AI is permitted, what you grade shifts from the finished artifact to the student’s judgment, revision, and reasoning around it.
• A detector score is the wrong tool here. If you allowed the tool, a high AI reading isn’t a violation; it’s expected.
• Build rubrics that reward decisions you can see: what the student kept, cut, fact-checked, and disagreed with.
• Require a short process artifact, a disclosure note, a prompt log, or an annotated draft, so the work shows its seams.
• Use brief oral or in-class components to confirm the student can stand behind the ideas on the page.

Stop grading the artifact

For most of the history of written assignments, the artifact and the effort were the same thing. A clean five-paragraph essay meant someone sat down and produced five clean paragraphs. The document was the proof.

That link is broken now. A student can get a competent, grammatically spotless draft in thirty seconds. So if your rubric still rewards “clear thesis, correct grammar, logical structure,” you’re handing points to whoever pasted the best prompt. Those criteria haven’t stopped mattering, but they’ve stopped being scarce. They no longer separate the student who thought hard from the one who didn’t.

The fix isn’t to ban the tool you already allowed. It’s to move the grade toward the parts a model can’t do for the student: choosing what’s worth saying, deciding what’s wrong with the first draft, bringing in evidence the model never had, and defending a position when pushed. Those are still scarce. Grade those.

What to assess instead

Judgment and revision

The most gradable thing in an AI-permitted assignment is the gap between what the tool offered and what the student turned in. Did they accept a generic opening, or cut it for something specific? Did they catch the model’s confident-but-wrong claim, or let it ride?

You can see this if you ask for it. Require students to submit one earlier draft or a short note describing two changes they made to AI output and why. The “why” is what you grade. “I removed this paragraph because it repeated my second point” is a real decision. “I reworded it to sound better” is not.

Originality of framing and evidence

AI is fluent and generic. It writes the median answer. So reward what pulls away from the median: a local example, a primary source, an interview, a personal observation, a counterargument the obvious draft would skip. A student who connects the reading to something happening in your own town is doing work no model did for them.

This also tends to produce better writing on its own terms. Heavily AI-leaning work reads thin and oddly weightless, full of confident summary and short on anything concrete. You’ll notice it. Your rubric should let you grade it down for substance, not for tool use.

Accuracy and sourcing

Models invent citations and state plausible falsehoods with total confidence. Make verification part of the grade. Require that every factual claim and every source be checkable, and dock points when one isn’t. This rewards the student who actually read the article they cited, and it quietly catches the one who pasted output and never looked.

A mini-scenario: two history essays

Two students submit on the causes of the 1929 crash. Both used AI; your policy allowed it.

Essay A is smooth. Solid structure, no typos, accurate in a textbook way. It could be about any market crash. Every example is one the model would reach for first, and there’s a footnote to a book that, when you check, doesn’t say what the essay claims.

Essay B is a little rougher in places. But it opens by comparing margin buying in 1929 to a payday-loan ad the student saw last week, leans on two sources you recognize as real and correctly summarized, and includes a paragraph where the student writes, “The draft I started with blamed the Fed entirely; I cut that after reading Galbraith, who’s more skeptical.” Attached is a three-line note listing the tools used and one suggestion the student rejected.

Under an old rubric, A might edge out B on polish. Under a rubric built for AI-permitted work, B wins clearly: real evidence, visible judgment, a verifiable trail, and a fabricated citation in A that costs it. You graded the thinking, and the grades came out right.

Designing the assignment so it’s gradable

The assessment is only as good as the assignment. A few moves make AI-permitted work far easier to grade honestly.

• Ask for a process artifact. A prompt log, an annotated draft, or a short reflection. It doesn’t need to be long; it needs to show decisions.
• Anchor to something local or personal. Tie the prompt to class discussion, a field observation, the student’s own data, or this week’s specific events. Generic prompts get generic output.
• Build in a checkpoint. An outline due early, or a quick in-class write, gives you a baseline for the student’s voice and thinking before any tool touches it.
• Make sourcing load-bearing. If claims must be verifiable and tied to assigned or student-found materials, fabrication becomes a liability instead of a shortcut.

None of this requires surveillance software. It just makes the work show its seams, which is what you want when the surface is free.

Writing the rubric

Translate all of that into criteria you can mark. A workable AI-permitted rubric tends to weight four things:

1. Substance and originality (the largest share): Is the framing specific? Is the evidence real and well-chosen? Does it go past the median answer?
2. Judgment and revision: Do the process artifact and the draft show deliberate choices the student can defend?
3. Accuracy and sourcing: Are claims and citations verifiable and correctly used?
4. Clarity and mechanics (the smallest share now): Still counts, but it’s table stakes, not the headline.

Notice what’s missing: any line that scores “amount of AI used.” You’re not measuring input. A student who used AI heavily but made sharp editorial calls can score exactly as high as one who wrote it cold. The tool is allowed; the thinking is graded.

Why a detector shouldn’t drive the grade

It’s tempting to keep an AI detector in your back pocket as a check. Resist using it to set grades here. The logic doesn’t hold: you permitted the tool, so a high AI reading isn’t a violation. It’s your policy working.

Detectors also aren’t built for verdicts. They output probabilities, not proof, and they generate false positives, sometimes flagging non-native English writers more often, as a 2023 Stanford study by Liang and colleagues found. OpenAI retired its own AI text classifier in July 2023 for low accuracy. If you want to understand what those scores actually measure, the statistics behind AI detectors are worth a read, mostly so you know why not to lean on a number that wasn’t designed to bear weight.

The honest verification tool in an AI-permitted course isn’t software. It’s a short conversation. Ask one follow-up question about a choice the student made. Two minutes tells you whether they own the work, and there’s no false-positive rate on a student explaining their own reasoning.

Putting it together

Allowing AI doesn’t lower the ceiling on what students can produce; it raises the floor on the artifact and forces you to assess the part that was always the point. Your job moves from catching cheaters to coaching judgment, and your rubric should follow. Grade the decisions, the evidence, and the reasoning. Let the polish be free, because it is now.

If you teach, it helps to know firsthand what AI-generated text looks like before and after a human shapes it, and how detection scores behave on the same passage. You’ll find more articles for educators on our blog, and you can try a humanizer and detection checker yourself to see how the tools your students use actually work. The better you understand them, the easier they are to grade around.