Teaching with AI

Course design, materials and feedback, for faculty and TAs. About 45 min, beginner level.

You'll be able to

• Decide which teaching tasks to hand to AI and which need you, using the four Ds of AI fluency.
• Draft practice questions, examples and study aids with AI, and review every item before students see it.
• Explain to students how you use AI in your course, and keep their data out of tools that aren't approved for it.

Where AI helps in teaching

AI is most useful for the parts of teaching prep that need volume and variety:

• Worked examples at several levels of difficulty, or set in different contexts.
• Practice questions and alternate versions of problems.
• Study aids from your own materials: glossaries, summaries and review guides. A grounded tool like NotebookLM answers from the slides and readings you upload, with links back to the passage.
• Analogies and case scenarios to test in class.
• First drafts of rubrics, announcements and activity instructions.
• Transcripts and captions of recorded lectures, which help students who need them, and everyone reviewing later.

What stays with you: judging whether content is correct, knowing your students, deciding grades, and the conversations where learning actually happens. Used deliberately, AI takes over some drafting so you have more time for those.

Delegation and description

The AI Fluency Framework by Rick Dakan and Joseph Feller describes four competencies, the four Ds: Delegation, Description, Discernment and Diligence. They map neatly onto teaching.

Delegation is deciding whether and how to involve AI, based on your goal, the task and what the tool can do well. Drafting ten practice questions on a topic you know deeply is a good fit: you can spot every error. Writing an explanation of a topic at the edge of your expertise is riskier, because you're less able to catch mistakes.

Description is communicating clearly what you want. For practice questions, that means the learning objective, the level, the misconceptions you want to target, the format, and what the answer key should include. “Ten questions on photosynthesis” gets generic recall items. A description of what students must be able to do gets questions that test it.

Discernment: review everything

Discernment, the third D, is evaluating what the AI produced. With teaching materials, review every item, because students will trust what you give them.

Common problems in AI-drafted questions:

• Wrong answer keys, especially in math, chemistry and anything with calculation steps.
• More than one defensible answer, or a correct option that's worded more carefully than the rest.
• Weak distractors that no student would choose.
• The wrong skill: recall questions when you wanted application.
• Content you didn't teach, or terms defined differently from your course.
• Narrow examples: names, contexts and assumptions that don't reflect your students.

Solve each question yourself. Grounded tools reduce invented content but don't remove errors, so follow the citations. A second reviewer, like a TA, catches what you miss. Expect to edit a meaningful share of what's generated; the time saved is in drafting, not in skipping review.

Diligence: transparency and privacy

Diligence, the fourth D, is taking responsibility for how AI is used and what it produces.

Tell students how you use it. A line such as “Practice questions were drafted with AI and reviewed by me and the TA” models the disclosure you'll probably ask of them, and invites them to flag errors.

Protect student data. Student work and grades are education records, which FERPA protects in the US. Don't paste identifiable student work, grades or messages into a personal AI account. USC provides ChatGPT Edu, which keeps chats in a university workspace, but even there, check your school's guidance on which kinds of data it's approved for.

Feedback and grading. If you use AI to help draft feedback, check your school's guidance first, tell students, and review every comment. Grades remain your decision.

Recording. Tell students before you record or transcribe a class.

For current guidance, start with USC's Center for Excellence in Teaching and your school's own policies.

A practice set from one lecture

An instructor uploads one week's lecture slides to NotebookLM.

Description: “Using only these slides, write ten multiple-choice questions for second-year students. Each should test applying a concept to a new situation, not recalling a definition. Four options each, one correct, with distractors based on common misconceptions. Give the answer, a one-sentence explanation and the slide it relies on.”

Discernment: the instructor solves every question. Of the ten:

• Two have answer keys that don't match the slides.
• One has two defensible answers.
• Three are recall questions despite the instruction, so they're rewritten as short scenarios.
• Several distractors are too obviously wrong and get replaced with real misconceptions from past exams.

Second review: the TA takes the quiz cold and flags one question as ambiguous. It's reworded.

Diligence: the set goes out with a note: “Drafted with AI from our slides, then checked and edited by the teaching team. Spot an error? Tell us.”

The draft took minutes. The review took most of an hour, and the result is a set the instructor stands behind.

Try it

Ten practice questions built from one lecture's material, edited by you and reviewed by a TA or colleague, with a short note on what you changed and a one-line disclosure for students.