Rubrics for AI-permitted assignments

Editable rubrics for the assignment categories in Part II of Learning with AI. The core rubric below is the worked example printed in the book. The variants adapt it to writing and coding courses without changing what it measures.

Everything here is MIT-licensed. Copy it, cut it, rename the rows, and make it yours.

Why these three rows

An AI as Collaborator (AIC) assignment permits AI for brainstorming, drafting, debugging, revision, critique, outlining, coding, or comparison. The student still has to verify the result, own the work, and disclose the help.

That means the three things worth grading are not the polish of the product. They are verification, ownership, and disclosure.

Grade the rows separately. Do not average them into a single score. A strong analysis with absent verification is not a strong AIC submission, and averaging hides exactly the failure the category exists to catch.

Core rubric (the book’s version)

Criterion Strong Adequate Needs revision
Verification discipline Key claims and code tested against known cases; checks documented. Spot checks on the main results; some documentation. Output accepted as returned; no independent check visible.
Integration and ownership Student can explain and extend every part; AI material rebuilt in the student’s own structure. Student can explain the main line; some passages remain tool-shaped. Student cannot explain or vary substantial parts.
Disclosure quality States tools, roles, and what was verified, proportionate to use. States tools and roles. Missing, vague, or inconsistent with the work.

Use it as printed for a data-analysis report, a modeling assignment, or any task where AI may draft both code and prose.

Variant: writing courses

The rows are the same. What counts as evidence changes.

Criterion Strong Adequate Needs revision
Verification discipline Every source exists, was read, and supports the claim it is attached to; quotations and figures checked against the original. Main sources checked; a few citations taken on trust. Citations unverified; at least one source does not say what the paper claims.
Integration and ownership The argument is the student’s; AI-suggested phrasing has been rewritten to fit the student’s line of thought. The argument holds, but passages read as inserted rather than written. The student cannot say why the argument moves the way it does.
Disclosure quality Names the tool, the stage it was used at (outline, draft, revision), and what was checked afterward. Names the tool and the stage. Missing, vague, or contradicted by the draft history.

The first row carries the most weight here. AI-generated citations that look correct but reference the wrong year, journal, or paper are the characteristic failure of AI-assisted writing.

Variant: coding courses

Criterion Strong Adequate Needs revision
Verification discipline Tests cover the normal case, the edge cases, and at least one case with a known answer; failures are shown, not hidden. The happy path is tested; edge cases are thin. Code runs; correctness is asserted rather than demonstrated.
Integration and ownership Student can explain every function, justify the design, and modify it under a changed requirement. Student can explain the main flow; some blocks are opaque to them. Student cannot explain or safely change substantial parts.
Disclosure quality States which parts were AI-drafted, what was rewritten, and how the result was tested. States that AI was used and roughly where. Missing, or inconsistent with the commit history.

For the ownership row, a live modification request is the fastest check: change one requirement and ask the student to adapt the code in front of you.

Adapting this rubric

Keep three rows. Two collapses verification into ownership. Four starts grading polish again.

Say in the assignment what each row means for this task. Students calibrate to the rubric they are given, which is the point.

Set a floor, not just a weighted sum. Many courses require at least Adequate on verification to pass the assignment, regardless of the other rows. This is the same principle the book applies to Core Competence: some things do not average away.

Match the disclosure row to your policy. If your course requires a disclosure statement, grade it. If it does not, delete the row and say so, rather than leaving an ungraded requirement on the page.

From Learning with AI: A Framework for Students, Instructors, and Universities (Part II). Repository materials are MIT-licensed.