Assignments and Feedback

Two Different Problems

This page is for faculty, and it addresses two things that tend to get discussed together but need different answers: designing assignments that still measure something real now that students have AI, and using AI to help with the grading and feedback load.

On the first: the instinct after seeing an AI-written submission is to make assignments harder to complete with AI. That is worth some effort, but it is a losing race on its own — the tools improve faster than assignments can be redesigned around them.

The more durable move is to make the assignment measure something the tool cannot supply. An AI can produce a competent essay about a novel. It cannot produce your student's reaction to a specific class discussion, field observational data collection, a reflection on their own draft from three weeks ago, or an defense of their reasoning in conversation with you.

That framing also tends to produce better assignments regardless of AI, which is the useful part.

Designing Assignments

Tie the assignment to material that does not exist in a model's training data: a particular class session, a local site or organization, a dataset the student gathered, a text you assigned in an unusual pairing, an event from this semester.

An AI can write about food deserts in general. It cannot write about the specific intersection your class visited, what the three students in that group noticed, and how it complicated the reading from week six.

Collect outlines, drafts, annotated bibliographies, revision notes, or a short reflection on what changed and why. This shifts the assessment toward the work of getting there.

It also creates a natural conversation if something seems off. A student who cannot describe how their draft evolved has told you something, without any accusation being made.

Keep it proportionate — process artifacts are extra work for you as well as for them. One or two well-chosen checkpoints beat a portfolio requirement nobody reads.

A two-minute conversation about a submitted paper tells you more than any detection tool will. "Walk me through why you organized it this way." "Which source was most useful and why?"

This scales badly in a large section, but it does not have to be universal. Even a sampled version — a handful of students per assignment, announced in advance — changes what students expect.

Instead of trying to exclude it, assign it: have students generate something with AI, then critique it, correct it, and document what it got wrong.

This works particularly well where errors are checkable. Students find the fabricated citation, the flawed derivation, the missing perspective — and they learn the verification habit in a way no warning achieves. You are assessing evaluative judgment, which is a skill worth having and one AI cannot demonstrate on their behalf.

"Use AI responsibly" is not a usable rule. Students genuinely cannot tell where the line is, and most policy violations come from ambiguity rather than intent.

Specify by task: brainstorming permitted, generated prose not; grammar checking permitted, rewriting not; practice questions permitted, submitted answers not. State what disclosure looks like if you require it. And explain the reasoning — students follow rules they understand considerably better than rules they are handed.

Academic Integrity and AI has syllabus language you can adapt.

Using AI for Feedback

This is where care is needed, and the boundary is sharper than it first appears.

Student work is student data. Pasting a student's paper into a consumer AI tool means putting identifiable student work into a system the district does not control. Once student work is collected and maintained by an educational institution, that work is protected student data and would raise FERPA concerns and is not a risk to take casually, whatever a tool's privacy policy claims. Check Data Privacy and Security.

Grades are yours. A grade is an institutional judgment you are accountable for. AI does not assign it, and "the tool scored it" is not an answer you can give a student in a grade appeal.

  • Building rubrics. Describe the assignment and the outcomes and ask for a draft, then fix it. Fast, and it surfaces criteria you had in your head but had not written down.
  • Drafting comment banks. Generate feedback language for the eight problems you see every semester, then personalize when you use them.
  • Testing your own assignment. Run your prompt through a tool. If it produces a solid submission in thirty seconds, you have learned something useful before your students do.
  • Clarifying your instructions. "Read this assignment description as a first-semester student — what is ambiguous?" This catches confusion that is invisible to the person who wrote it.
  • Varying your explanations. Generating three ways to explain a concept students keep missing.

None of these involve student work. That is the pattern to keep.

Do not use AI to:

  • Assign or recommend a grade
  • Evaluate whether a student's work is honest
  • Draft communication about a specific student's performance or conduct
  • Make a decision about accommodation, referral, or intervention

These require someone who knows the student and is accountable for the outcome. They are also exactly the decisions where a confident wrong answer does the most damage.

AI detection tools are not reliable enough to support an integrity finding on their own. They produce false positives, and at documented higher rates for writing by non-native English speakers — which makes uncritical reliance an equity problem, not only an accuracy one.

Treat a detection score as a prompt to look more closely, never as evidence. Academic Integrity and AI covers the full process, including what to do if you have a genuine concern.