How to Assess AI-Assisted Student Work

Start with the role AI is allowed to play#

Students will use AI. The question for an assignment is what role it should have. Write one sentence in the assignment brief that places the work on a simple scale.

The level should follow the learning outcome. If the outcome is independent writing, the level is probably "no AI" or "AI for support," and your assessment needs evidence collected in conditions that show independent skill. If the outcome is judgment about AI output, the level is "AI required," and the grade should reward the quality of the evaluation.

Require a short disclosure#

A disclosure turns a hidden practice into something you can discuss. Keep it brief so students complete it honestly. A workable template:

• Which tools did you use?
• What did you ask them to do?
• What did you accept, change or reject, and why?
• Which parts of the work are entirely your own?

Do not penalize honest disclosure of permitted use. If students are marked down for admitting what the policy allows, they will stop admitting it. Disclosures are also easy to produce with a model, so treat them as a prompt for conversation, not as proof.

Assess the student's understanding as well as the artifact#

The artifact tells you what was submitted. It does not tell you what the student understands. Add a step in which the student explains the work and answers questions about it. Options include:

1. A short oral defense. Five to fifteen minutes, with questions drawn from the student's own submission. Questions can ask the student to justify a claim, explain a source, describe an alternative they rejected or apply the argument to a new case.
2. A written reflection under supervision. A short in-class response to a question about the submission.
3. A revision task. Ask the student to change the work in response to a new constraint, live.

Score the explanation with its own rubric. A strong artifact with a weak explanation is information about the student's understanding, and you can weigh it in the grade or follow up with a conversation. Our guide to the AI oral exam format explains one way to run the defense at scale, and the oral assessment rubric guide gives criteria you can adapt.

Add criteria that suit AI-assisted work#

Rubrics written before generative AI often reward surface qualities such as fluent prose and tidy structure, which a model produces easily. Shift weight toward criteria that require the student's own judgment:

• Accuracy of claims and sources. Did the student verify each claim and check each citation? Models can produce references that do not exist, so ask students to be ready to show sources.
• Quality of reasoning. Does the argument follow, and can the student defend each step?
• Evaluative judgment. Can the student explain what is strong and weak in their own work and in any AI output they used?
• Use of course material. Does the work engage with the concepts, data and cases from your course?
• Ownership. Can the student answer follow-up questions without notes?

What to do when you suspect misuse#

1. Do not use a detector score as evidence. Weber-Wulff and colleagues tested a range of detection tools and concluded they are neither accurate nor reliable. Liang and colleagues found detectors misclassified writing by non-native English speakers as AI-generated. Using a score as proof risks a wrongful accusation. See AI oral assessment compared with AI detection.
2. Talk to the student. Invite them to walk you through the work. A conversation about their sources, choices and reasoning tells you more than a score.
3. Compare against other evidence. In-class work, earlier drafts and the explanation give context.
4. Follow your institution's process. Document what you observed and what the student said. Misconduct decisions belong to the process your institution has, not to a tool.
5. Treat the outcome as information. If many students struggle to explain work they submitted, the assignment may need redesign.

Cotton, Cotton and Shipway discuss strategies for academic integrity in the era of ChatGPT, including redesigning assessment and educating students about expectations. Those strategies work better together than alone.

Fairness and accessibility#

Explanation-based methods can disadvantage students who are anxious, who speak English as an additional language or who have a disability. Plan the following before you start:

• A practice attempt that does not count.
• Extra time on request, and a text response option where speech is not essential to the outcome.
• Criteria that score reasoning rather than fluency.
• A clear appeal path.

Where Rocketproof.ai fits#

Rocketproof.ai runs the explanation step as a structured oral defense (see the oral defense platform page) after a student submits a paper, a slide deck, a reading response or other work. The AI asks questions drawn from the submission and follows up on the answers, and the instructor reviews the transcript and draft rubric scores before any grade is released. Students can respond by speaking, by live conversation or by typing, and approved accommodations such as extended time are supported. It assesses whether the student can explain and defend the work, which is separate from detecting AI-written text.