Build your own project-based assessment

Claude, ChatGPT, and Microsoft Copilot all handle this well, and the free tiers are enough for a conversation this size. If your college already licenses one, use that one, since it likely comes with institutional data protections that consumer accounts do not.

One practical note: free tiers sometimes cut off long conversations. If that happens partway through, paste the last thing it produced into a fresh chat along with the original prompt, and pick up where you left off. The "stop and summarize" phrase above makes this painless.

You are the author of record, and you are responsible for everything you assign. That does not change here. What changes is who does the typing on the first draft.

Two things I would do anyway. Check your institution's AI policy, since some require disclosure for instructional materials. And be candid with colleagues about how you built it, because the specifics are useful to them and hiding it makes it look shadier than it is.

No. Keep student names, grades, submissions, and anything else from your gradebook out of consumer AI tools entirely. Under FERPA those are education records, and pasting them into a service your institution has not vetted is exactly the kind of disclosure the law is about.

You do not need any of it for this task. You are designing an assignment, not analyzing students. If you later want AI help with something involving real student data, that conversation goes through your institution's approved tools and your registrar.

Yes, and it is the most important limitation to know about. AI assistants sometimes state wrong things with complete confidence, and they are notorious for inventing citations, statistics, and quotations that look real and are not. The industry calls this "hallucination."

For this task it is mostly harmless, because you are asking for structures and drafts, not facts. But treat any factual claim, any citation, and any statistic it produces as unverified until you check it yourself. If it names a study or a source, assume it is made up unless you can find it. And never let an AI-drafted factual claim reach a student handout without your own verification. The rule of thumb: trust it with the typing, never with the truth.

The design does, and this turned out to be the unplanned benefit of the whole approach. Authentic tasks require presence and evidence: attending something, interviewing someone, visiting a place, building a physical thing, documenting a process with original photographs. A chatbot can write a competent paragraph about federalism. It cannot sit through a zoning hearing.

Require evidence that would be a real hassle to fabricate: a photo of the agenda with a timestamp, a recording of the interview, a signature from the office they visited. That is cheaper and more pleasant than escalating proctoring, and it makes the assignment better rather than more adversarial.

Come with three things. First, the alignment: because you built this backwards from your learning goals, every activity already names the outcome it serves. That mapping is the language departments and accreditors speak, and it is usually cleaner than anything a multiple-choice exam can show.

Second, the rubrics, because a published checklist is more defensible under scrutiny than an answer key nobody outside your course has seen. Third, run it as a pilot in one section and collect student feedback. Data from your own students ends most of these conversations.

Yes, and the translation is usually obvious once you look for it. Nursing students shadow a clinic and document a workflow. Business students interview a small-business owner about a real regulatory headache. Biology students survey a local habitat. Art history students visit a collection and write about one object. Composition students report a story that requires talking to a stranger.

The test is not "is it civic." The test is whether a student has to leave the browser tab and come back with something only they could have produced.

Yes, with one honest adjustment: "leave the browser tab" becomes "leave this tab." City councils and school boards stream their meetings, interviews happen over Zoom or a phone call, and plenty of real civic and professional interactions now run through official portals. The test stays the same: did the student interact with something real and come back with evidence only they could have produced?

For online sections, lean on evidence that is hard to fake remotely: a recording of the interview, a confirmation email from the office they contacted, a timestamped screenshot of the live stream with the agenda item visible. Say so in your constraints in Step One and the AI will build the menu online-first.

The menu is structurally friendlier than an exam, because multiple routes to the same points is exactly what universal design asks for. A student who cannot attend an evening meeting picks the interview; a student with a processing accommodation is not racing a clock. When you build the menu, make sure every in-person activity has a remote or asynchronous equivalent, and the prompt on this page already tells the AI to flag accessibility concerns.

Formal accommodations still run through your disability services office, same as always. The difference is that most of them now require no special arrangement, because flexibility is the design rather than the exception.

One assignment group, weighted at whatever the project is worth. In Canvas: create a group called, say, "Project" at 40%, put each activity inside it as an assignment worth its point value, and set the group to total against your point target. Checkpoints go in as small milestone assignments or zero-point reminders, whichever matches the policy you chose.

The pleasant surprise is transparency: students can see exactly where their points stand at any moment, which kills the "what's my grade" email almost entirely.

Complete/incomplete against a published checklist. That one decision is what makes the model survivable. There is no partial credit to calculate, no rubric row to agonize over, and no negotiation, because the criteria were public before the student started. Grade disputes mostly vanish for the same reason: the remedy is always "fix the missing criterion and resubmit," so there is nothing to argue and no discretion to appeal to.

It took me three rounds of rubric revision to get there. The first versions had criteria I could not actually verify from what students submitted, which is the mistake to watch for.

Parts of it will not, the first time. My checkpoint pacing was wrong for 8-week sections, verifying attendance took more of my time than I planned for, and a few activities drew almost no takers.

None of that is a reason to go back to the exam. Keep what worked, cut what did not, and run it again. That is not a flaw in the method. That is the method.