Sam's take
Two years ago, the weekly request was "Can we have dynamic questions?" "Can we have more questions?" Students were lifting answers off Chegg, and faculty wanted variation to stay ahead of it. I haven't heard that request in six months.
The threat moved. Students are using AI to complete assessments, and the tools keep getting better. The Chegg arms race is over because it's no longer the main problem.
What I'm seeing professors actually do: pull assessment back into the room. Starting late last year, we kept hearing the same thing from Finance and Real Estate faculty, that they'd gone back to paper exams. We built a free auto-grader for paper exams because of those conversations, and it spread faster than anything else we've shipped. That tells me the shift is real, not anecdotal.
I don't think paper is the whole answer. It's the fastest answer. The more interesting work is further out: oral defenses of submitted work, in-class presentations, grading the process rather than the artifact. Some of this is going to feel like going backwards. Some of it probably is. I'd rather a student stumble through explaining their own model than submit a polished one they can't defend.
So the question I'm sitting with, and the one worth using the summer on: what does an assessment look like that a motivated student can't outsource to a model, but that still respects your time to grade? I don't think anyone has a clean answer yet.
Teaching notes
Is assessment broken?
Most of the conversations I have with finance and real estate faculty right now circle back to the same problem: the take-home assignment is broken, and nobody is sure what replaces it. Multiple-choice gets gamed. Essays get generated. Even the "explain in your own words" prompts come back in a voice that sounds suspiciously like GPT's.
So when I came across this post from Danil Fedorovykh, an economics professor sharing a concrete set of AI-resistant assignments, I read it twice and then sent it to three faculty I'd been talking to about exactly this. It is the most practical thing I have seen on the topic in months, and I think it generalises cleanly to finance and real estate.
A few things stood out as worth pulling apart.
The shift is from "produce an answer" to "produce a trace of thinking." This is the move. If you ask a student to value a commercial property or build a DCF, ChatGPT will hand it to them. If you ask a student to value a specific commercial property in their city, walk through three assumptions they had to make, defend why they chose one cap rate over another, and reflect on what changed in their thinking between draft one and draft two, you have asked for something the model cannot fabricate on its own. The student still has to do the work. The AI becomes a tool, not a substitute.
Danil's examples lean heavily on this principle. He builds assignments around process artefacts: drafts, revisions, in-class follow-ups, oral defences. Each one is cheap to add and expensive to fake.
Local and personal beats general and abstract. One of his ideas I particularly liked: anchor the assignment to data or context the student has to gather themselves. For a finance class, this might be analysing a public company headquartered in the student's home town, or interviewing a family member about a real financial decision they made. For real estate, it is almost too easy: pick a property within five miles of campus, pull the actual comps, talk to a broker. The friction of going out and getting the inputs is the assignment. The model can help with the analysis, but the student owns the raw material.
Oral components do more than you'd expect. Danil notes that even a five-minute follow-up conversation about a submitted assignment dramatically changes student behaviour. They know it is coming, so they actually engage with the work. You do not have to grade the oral; the existence of it is the deterrent. I have heard variations of this from three finance professors this semester, and one of them has gone further: every major submission triggers a randomised 10% sample of students for a brief Zoom conversation about their work. The grade weight is small. The signal it sends is large.
In-class anchoring is doing a lot of quiet work. Several of Danil's assignments are constructed so the take-home portion only makes sense if the student was present and engaged during a specific in-class exercise. The take-home builds on a dataset the class generated together, or extends a problem the professor walked through. This is a clever way to use class time as a kind of authentication layer. It also rewards attendance without making attendance the explicit point.
What he is not doing is policing. No detector software. No "did you use AI" affidavits. No adversarial framing. The assignments are designed so that the AI question becomes less interesting: either the student did the underlying thinking or they did not, and the structure of the assignment makes that visible. That is the right posture, in my view. Trying to out-detect the models is a losing game. Designing around them is not.
The thing I would encourage you to take from this is not a specific assignment template but the underlying move: every assessment now has two layers, the deliverable and the evidence of thinking that produced it. If you can capture even a little of that second layer, cheaply, the integrity problem softens considerably.
Worth a look
An interesting essay on how technology has influenced students' ability and indeed willingness to read. Worth a read, if you're willing!
Upload your own videos into Blended Teaching
It's now possible to upload your own videos into Blended Teaching. The platform will automatically split it up into bite-sized video pieces, generate notes for them, and you can even create question banks from your own content.
The Future of Real Estate Education
Routledge has published an exciting new book from a phenomenal group of authors. Hit the link to check out the latest pedagogical and technological advancements in the world of real estate education.
Sam