Assessment in the AI era

Designing assignments where AI use is clear, fair and visible. About 45 min, intermediate level.

You'll be able to

• Test one of your assignments against current AI assistants and grade the results with your own rubric.
• Decide, for each assignment, whether AI use is excluded, allowed with disclosure or required, and say so plainly.
• Redesign an assignment so students' thinking is visible through process evidence, authentic tasks or short conversations.

Why design beats detection

AI detectors are tempting because they promise a quick answer. They aren't reliable enough to carry that weight.

• False positives: they flag writing that a student wrote themselves. Some studies have found higher false-positive rates for writers whose first language isn't English.
• False negatives: light editing or paraphrasing often gets AI text past them.
• No evidence trail: a score doesn't show what happened, and students can't meaningfully rebut it.

OpenAI withdrew its own AI text classifier in 2023, citing its low accuracy. A detector score shouldn't be the sole basis for an academic misconduct case. At most it's a reason to look more closely and talk with the student, following your institution's process.

The better investment is in design: clear rules about what's allowed, and assignments where the student's own thinking is visible. That works whether or not AI was used.

AI-test your own assignment

Before you redesign anything, find out what current tools do with your assignment as it stands.

1. Paste the prompt, exactly as students see it, into two assistants, for example ChatGPT Edu and Claude.
2. Do what a student might: ask a follow-up or two to improve the answer.
3. Grade each result with your real rubric, as if a student had submitted it.
4. Note which criteria the AI met easily and which it missed.

Use only your own assignment text for this, never student submissions.

The result tells you what the assignment currently measures. If an AI answer earns a solid grade, the rubric may be rewarding things AI does well, like fluent structure and broad coverage, rather than the learning you care about. It's common to find AI weakest on course-specific material, accurate citations, recent class discussion and personal reasoning. Those weak spots point to where the redesign should go.

Three stances, assignment by assignment

There isn't one right AI policy for a whole course. Decide per assignment, based on what it's for:

• Exclude AI where the struggle is the learning, like foundational skills or first attempts at a method. Pair this with conditions that make it realistic, such as in-class work.
• Allow with disclosure where AI can support learning without replacing it: brainstorming, feedback on a draft, checking grammar. Say which uses are allowed and how to disclose them.
• Require AI where using it well is the skill: critiquing an AI answer, comparing tools, or practicing a workflow from the field.

Whatever you choose, be specific. “You may use AI to brainstorm topics and check grammar, but not to draft paragraphs” is followable; “use AI responsibly” isn't.

This is the AI Fluency Framework's Delegation applied to course design. If you require AI, make sure every student has access; USC provides ChatGPT Edu. Check your school's guidance for any required syllabus language.

Make the thinking visible

When the final product alone can't show who did the thinking, assess the process too:

• Staged work: a proposal, a draft with revision notes, then the final version.
• Reflections on specific decisions: why this source, why this approach, what changed.
• Short conversations: a five-minute check where students explain a choice in their work.
• In-class components: a first draft, a problem set or a presentation done in the room.
• Authentic tasks: local data, a real audience, course readings, or the student's own experience.

Be realistic about authentic tasks. AI can still help with them; they make generic output less useful rather than impossible, which is often enough.

Then move rubric weight toward reasoning, use of evidence and justification, and away from polish alone. Watch the workload for you and for students: redesign one element at a time. A single well-placed checkpoint often does more than a full overhaul.

Redesigning an essay assignment

Original: “Write a 1,500-word essay analyzing the causes of the 2008 financial crisis.”

AI test: two assistants produced well-organized essays that scored well on structure and coverage. One cited a paper that doesn't exist. Neither used the course readings or class discussion.

What the assignment should measure: students' ability to build an argument from primary sources in the course reader.

Redesign:

1. Proposal (in class): choose one primary source from the reader and state a claim about it in a paragraph.
2. Draft: with a short note on any AI use.
3. Final essay: must quote and analyze the chosen source and one class discussion.
4. Five-minute conversation in office hours or section: explain one choice in the essay.

Rubric change: use of the primary source and reasoning now carry most of the weight; prose style carries less.

Policy line: “You may use AI to brainstorm and to check grammar. Don't use it to draft paragraphs. Add a two-sentence note on any AI use: which tool, for what, and how you checked it.”

Try it

One of your assignment prompts, the two AI answers graded with your rubric, and a redesigned version with one element changed so the thinking is visible, plus a policy line for that assignment.