The take-home essay had a good run. For a century it was the workhorse of higher education — cheap to assign, rich to grade, and, most importantly, a reasonable proxy for whether a student understood something. To write a competent essay on a topic, you generally had to understand the topic. The artifact stood in for the cognition. We graded the artifact and inferred the thinking, and the inference was usually close enough.
That inference is now broken, and everyone teaching knows it. A student can produce a fluent, well-organized, correctly-cited essay on almost any prompt without engaging the material at all. The artifact still exists. The cognition it was supposed to prove no longer has to. This isn't a story about dishonest students — it's a story about an assessment format whose core assumption, artifact implies understanding, quietly stopped being true. The completion looks identical. The learning behind it is unverifiable.
The reflexive responses don't hold up. Detection is an arms race you lose by entering — the detectors are unreliable, they flag non-native writers, and the models improve faster than the classifiers. Lockdown proctoring turns every student into a suspect and still only tells you whether they can pass a test under surveillance, not whether they understand. There's a better instrument, and it's an old one: the oral defense. Ask a student to explain their own work, justify a choice, and adapt their reasoning to a question they didn't see coming, and you are measuring understanding directly — not a proxy for it. A student who outsourced the work can hand you the paper. They cannot defend a judgment they never made.
The objection to oral defense has always been the same single word: scale. You cannot personally sit across a table from ninety students for fifteen minutes each. That constraint — and only that constraint — is what kept authentic defense a boutique instrument, reserved for the dissertation and the honors thesis. It's also the constraint that's finally lifting. But the format only works if the questions are right, so let's start there.
The principle: assess the reasoning, not the artifact
The move that makes an assessment AI-proof is counterintuitive. You don't try to prevent students from using AI on the work. You let them use whatever they'd use in real life — notes, sources, an AI assistant, the whole toolkit — and you assess the layer underneath the artifact: the decisions they made, the alternatives they rejected, and their ability to reason live about their own work.
This reframes the entire problem. You're no longer policing the process that produced the paper. You're asking the only question that ever mattered — does this student actually understand what they turned in? — and you're building a conversation only someone who does can navigate. AI can write the essay. It cannot sit inside the student's head and defend the essay's choices under adaptive follow-up, because those choices, if the student didn't make them, aren't there to defend.
Designing a defense that measures understanding
A good defense is short and built from a few high-signal question types. You don't need many questions. You need the right ones — questions that a student who did the thinking answers naturally and a student who didn't can't fake under follow-up.
- Justify a specific choice in the work. Point at something in what they submitted and ask why. "You framed the argument around this theory rather than the obvious alternative — walk me through that decision." A student who made the choice has a reason. A student who didn't will describe the theory rather than defend the choice, and the difference surfaces immediately.
- Defend against the alternative. Name a different reasonable approach and make them argue it down. "Why not analyze this dataset the other way?" Real understanding includes knowing what you ruled out and why. The rejected path is precisely the thing a generated artifact never contains, because generation produces the answer, not the deliberation behind it.
- Apply it to a case they haven't seen. Introduce a novel wrinkle on the spot. "What would change in your conclusion if the sample had been drawn differently?" Transfer to an unfamiliar situation is the understanding you're after; reproducing the submitted work is not.
- Locate the boundary. "Where does your argument stop holding? Under what conditions would it fail?" Knowing the limits of your own claim is a hallmark of having actually reasoned through it — and it is very hard to fake under a follow-up question.
Notice what isn't on the list: definitions, recall, anything a student could simply look up. Those test the cheapest, most automatable layer of knowing. Every question here targets judgment the student either built or didn't — and the follow-up is where the truth lives. A first answer can be rehearsed. The third one, chasing the student's own reasoning down a path they didn't prepare, cannot.
A defensible defense also needs a defensible score. If a mark is going to stand up to the student who didn't pass — and to an appeals committee — it has to be grounded, not impressionistic:
- Score against the understanding, not the polish. A nervous student who reasons well passes; a fluent talker who can't defend a choice doesn't. Grade the thinking, not the delivery, or you've built a confidence test.
- Ground every judgment in something the student said. "Couldn't justify the framing choice; described the theory instead of defending its selection" is a mark that survives challenge. "Seemed shaky" is not.
How the tool fits
Here's the turn. Everything above describes an excellent assessment that, run by hand, you can afford to give perhaps five students. The labor is the whole barrier — and the labor is exactly what's now automatable, without automating the judgment.
TeachingsByDesign builds authentic defense in through its sibling product, Crucible. From an assignment, you generate a no-account link — one per student — that opens an AI oral-defense examiner. The student clicks it, no login, and defends their work in conversation. The examiner poses questions in the spirit of the four types above, keeps a poker face so a confident tone earns nothing, and asks adaptive follow-ups — pressing where an answer is thin, following the student's reasoning rather than running a fixed script. It's the defense you'd run yourself, run for every student on the roster instead of the five you had time for.
Then it hands you back the seat that matters. You don't outsource the grade. Crucible returns a grounded, evidence-quoted score — every judgment tied to a specific thing the student actually said — for you to review and confirm. The machine runs the conversation at scale; you make the call. What gets automated is the part that made authentic defense impractical — sitting through ninety fifteen-minute conversations — not the pedagogical judgment, which stays exactly where it belongs, with you.
Used this way, AI flips from the thing that hollowed out your take-home essay into the thing that lets you finally verify understanding — for the whole class, not just the honors section. The instrument AI supposedly broke turns out to be the one instrument AI can't complete on the student's behalf.
The bottom line
The take-home essay assumed the artifact proved the understanding, and generative AI severed that link for good. You can't out-detect it and you can't proctor your way to they actually get it. The instrument that survives is the oldest one — the oral defense — because it measures reasoning live: the student justifies a choice, argues down the alternative, applies the idea to a new case, and names its limits, scored against understanding and grounded in what they actually said. It was always the most valid check you had. The only thing that ever made it rare was time, and that is precisely the constraint now lifting. Design the defense so the questions demand real judgment, and let the scale problem be the part you hand off.
Crucible, built into TeachingsByDesign, runs an authentic oral defense for every student — a no-account link, an adaptive AI examiner that holds its poker face, and a grounded, evidence-quoted score you review and confirm. Verify understanding, not the artifact. See how it works.