A 2026 Berkeley study of 500,000 grades found that grade inflation from AI tracked how heavily a course weighted homework. Courses weighting homework at 40 percent inflated more than courses weighting it at 10 percent. Weighting is the mechanism, so weighting is where the cheapest fix lives.
If you're teaching finance or real estate this fall, the semester is close enough that the goal is no longer to redesign your assessments. It is to re-sort what you already have into three lanes, and shift weight toward the lanes where AI is either visible, mostly irrelevant, or explicitly required.
We ran a webinar this week walking finance and real estate faculty through the exact set of moves. The full recording is above. The printable handout gives you five changes to make before your students arrive, each with a worksheet and a checkbox, plus a one-page summary you can pin above your desk. About three hours of work end-to-end.
What's in the handout
Five moves, and roughly what each one costs you:
- Re-weight before you redesign (30 min). One table in your syllabus. No new assignments; different point values. This item alone closes most of your exposure, and the handout has the specific rule of thumb for when it applies.
- Write the AI clause per assessment, not per course (45 min). Three sentence templates. Pick one per assignment. The handout has the exact wording for restricted, open-with-disclosure, and required.
- Convert one assignment to "Required" (60 min). One assignment where the critique of AI's output is what's graded, not the output. The strongest version is the planted-error audit; the handout has the three-step protocol.
- Add one process artifact to your biggest take-home (15 min). Either the working file, or a short build note answering two specific questions. Grade it 0 / 1 / 2. Unexpectedly diagnostic.
- Book your supervised checkpoints now (10 min). Rooms and slots disappear by week three, and this is the item that quietly fails if you defer it. The handout has starter viva questions for finance and real estate, and a 30-second model check rubric worth keeping on hand all semester.
Plus the one-page summary at the back. Three lanes for sorting what you already teach, and a balance rule for how many of each you need per outcome you claim.
If you only have one hour, do items 1 and 4. Together they address most of your exposure.
The honest version: no out-of-class assessment in 2026 is AI-proof. Assessment can be AI-visible, AI-irrelevant, or AI-integrated. Those are achievable in three weeks. "AI-proof" is not.
