How to Design an AI Evaluation Rubric for University Courses

Updated 5 November 2026 · Standard formula, shown below

An AI evaluation rubric helps instructors assess student work consistently while distinguishing subject knowledge from appropriate, transparent use of AI tools. Before drafting one, identify the course learning outcomes and decide whether AI use is allowed, limited, or prohibited for each kind of assignment; the steps below turn those decisions into criteria students and markers can apply.

What You’ll Need

The course and assignment learning outcomes, plus the assignment instructions and marking scale.
Your institution’s current rules on academic integrity, accessibility, privacy, and permitted AI use. Consult the relevant teaching or academic-integrity office if a rule is unclear.
Examples of successful work, common errors, or anonymised past submissions, if available.
Input from colleagues, teaching assistants, students, or an instructional designer to check that the rubric is understandable and usable.

Step-by-Step

Start with the learning outcomes. List what students must demonstrate, not simply what they must produce. Outcomes might include evaluating evidence, applying a method, developing an argument, or communicating for a particular audience. Keep the rubric focused on outcomes the assignment can actually assess. If an outcome is absent from the assignment, revise the assignment or leave that outcome out of its grading criteria.
Set the AI-use boundary for this assignment. State which uses are permitted, restricted, or not permitted, and explain any required disclosure. For example, an instructor might allow AI for brainstorming but require students to verify sources and write the analysis themselves. Another assignment may assess unaided performance. Make the boundary specific to the task and consistent with institutional policy; do not assume a general statement such as “use AI responsibly” tells students what is allowed.
Choose a small set of assessable criteria. Use criteria that directly reflect the outcomes, such as accuracy of disciplinary knowledge, quality of reasoning, evidence use, process documentation, or communication. Include an AI-related criterion only when AI use is part of the learning goal or the assignment requires a disclosure or verification process. Do not grade students on whether their prose merely “sounds human,” or treat a detector score as proof of authorship.
Define what each criterion means. Write a short description of observable evidence for each one. “Critical thinking” is too broad on its own; specify what markers should look for, such as comparing claims, weighing relevant evidence, and acknowledging limitations. Avoid criteria that reward access to particular tools, paid features, or technical fluency unrelated to the course outcomes.
Build performance levels with distinct descriptions. Select a practical scale, such as four levels from advanced to beginning, and describe what work at each level demonstrates. Make the differences meaningful: for evidence use, levels might distinguish well-integrated and relevant evidence from limited, poorly connected, or unsupported evidence. Avoid vague labels such as “good” and “poor” without explaining what they mean. Include a clear description for work that is missing or cannot be assessed, if your grading system needs one.
Decide weights and scoring rules. Assign points or percentages to criteria according to their importance to the outcomes. Check that the weights add to the assignment total and that a weakness in one area does not unintentionally erase achievement in another. State how partial credit, late work, missing disclosures, or other policy matters are handled, but refer policy violations to the established institutional process rather than inventing penalties in the rubric.
Make expectations for AI evidence proportionate. If students must disclose use, say what to include—for example, the tool or type of tool, the purpose, and any substantial output incorporated or adapted. Ask only for information needed to assess the assignment, and do not request private account details or sensitive data. A disclosure can help explain a process, but it does not by itself establish the quality or originality of the submitted work.
Check accessibility and clarity. Use plain language, explain course-specific terms, and provide the rubric before students begin whenever possible. Ensure that criteria do not disadvantage students for disability-related communication differences, language background, or lack of access to optional tools unless a skill is an explicit learning outcome. Offer an appropriate way for students to ask questions or request clarification.
Test the rubric before using it. Apply it to a few varied sample submissions, ideally with another marker. Compare scores and discuss where descriptions lead to different interpretations. Revise overlapping criteria, unclear level boundaries, or requirements that cannot be judged consistently. If suitable samples do not exist, use hypothetical examples and review early submissions before finalising marks.
Explain and review after marking. Walk students through the criteria and permitted AI practices. During marking, record brief evidence for important judgments and use moderation where feasible. After the assignment, look for recurring questions, unexpectedly low scores, and criteria that did not distinguish levels well. Update the rubric for the next offering and tell students what changed.

Common Mistakes

Making the rubric a list of AI rules. A rubric should assess learning; put detailed conduct procedures in the relevant assignment guidance or policy.
Trying to grade AI use itself. Reward the relevant learning outcome—such as evaluating generated claims—not the mere presence or absence of a tool.
Using vague or overlapping criteria. If “analysis,” “critical thinking,” and “argument” all measure the same thing, markers may double-count it. Define each criterion and separate its purpose.
Relying on AI detectors as grading evidence. Detection results can be uncertain and should not replace review of the work, a fair conversation, or institutional procedures.
Making disclosure requirements unrealistic. Do not demand complete logs that are difficult to produce or do not improve assessment. Explain what is necessary and why.

Tips

Use student-facing language and show a short example of how one criterion maps to a performance level.
Separate process from product when both matter: assess a required reflection or verification record only if it supports a stated outcome.
Keep a marker version with space for evidence notes, and use the same rubric across sections unless assignments differ.
Invite students to identify confusing wording before submission; their questions often reveal hidden assumptions.
Revisit criteria whenever assignment design, course outcomes, or institutional AI guidance changes.

Troubleshooting

Markers give very different scores: Hold a short calibration session using the same sample work. Ask each marker to identify evidence for each criterion, then clarify descriptions where interpretations diverge. Keep a record of agreed examples, while avoiding a rigid checklist that overrides academic judgment.

Students are unsure what AI use is allowed: Replace broad wording with task-specific examples of permitted and prohibited uses, and clarify whether disclosure is required. Direct unresolved policy questions to the appropriate institutional office before grading begins.

The rubric rewards polished output more than learning: Check whether presentation is weighted too heavily relative to the outcomes. Add or strengthen criteria for reasoning, evidence, or disciplinary application where appropriate, and make sure the assignment gives students a fair opportunity to demonstrate those skills.

A required process record is missing or incomplete: Follow the assignment instructions and institutional policy consistently. Do not infer misconduct from an incomplete record alone; assess available evidence and use the established process if a concern needs review.

The rubric is too long to use: Remove duplicate criteria, shorten repeated level descriptions, and move general instructions outside the scoring table. Keep only distinctions that affect feedback or marks.

Frequently Asked Questions

Should every university assignment have an AI-use criterion?
No. Include one when AI practice is a learning outcome or a required part of the task. Otherwise, explain permitted use separately and assess the assignment’s stated outcomes.

How many criteria should a rubric include?
There is no fixed number. Use enough to represent the important outcomes, but avoid splitting one skill into several overlapping criteria. A short rubric is usually easier to understand and apply.

Can I use an AI detector to decide a student’s grade?
A detector result should not determine a grade or establish misconduct. Use the rubric to assess the submitted work and follow institutional procedures for any separate concern.

Should students disclose every AI interaction?
Only require the information needed for the assignment’s purpose and consistent with institutional guidance. State the expected level of detail in advance.

When should students see the rubric?
Ideally, provide it with the assignment. Early access lets students understand the learning targets and permitted AI use before they plan their work.

How often should I revise the rubric?
Review it after each use, especially when students or markers repeatedly misinterpret a criterion, scores do not reflect intended outcomes, or course guidance changes.

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.