University teaching after generative AI: beyond lectures and inquiry to a certification-and-consultation paradigm in which learners become evaluators

Synthesis: Lectures and inquiry-based teaching both rest on the premise that what a student hands in reflects what the student learned, and generative AI breaks that premise for both. Deng (2026) argues that credible certification of ability now requires sampled, face-to-face probing — more human time than any current staffing model supplies — and that the posts providing that time can rebuild the junior rung of the evaluator career ladder that AI is eroding. The evidence he assembles is secondary: unrestricted GPT-4 access raised Turkish high-school practice scores by 48% while lowering later unassisted exam scores by 17%, and in a national survey of 18,664 Chinese doctoral graduates 35.3% had served as teaching assistants, more than 60% of those grading homework. The proposal is a four-level ladder on which certified students hold supervised certification and consultation posts, so learners move from being judged to judging. Deng presents this as a third paradigm of university teaching organized around judging understanding, with lectures and inquiry demoted to ways of learning.

Key Findings

1. A correct submission no longer evidences learning, so certification must happen in person. Deng's instrument is sampled probing: the student explains the reasoning, reproduces a central step, says why a plausible alternative fails and applies the method to an unfamiliar variant.
2. The arrangement is a four-level ladder staffed by students. Lower-year students are supported and certified; senior undergraduates probe within their certified areas; graduate students take cross-field consultation; professors hold standards and train their successors.
3. Consultation is distributed because AI-assisted work is broad rather than deep. A task spanning several fields exceeds any single judge, so Deng conjectures judging becomes routine, citing a blinded review by 27 clinicians across 13 specialties that found substandard answers and fabricated medical information in model outputs.
4. Existing teaching-assistant posts are grading posts, and grading is what AI made least informative. In a national survey of 18,664 Chinese doctoral graduates, 35.3% had served as teaching assistants and more than 60% of those graded homework; a smaller study found 95.65% graded assignments against 13.04% teaching exercise sessions.
5. The ladder is offered as a replacement for the rung AI is removing. Judgment was learned through supervised junior work, and Deng cites a Princeton computer scientist who can no longer use the starter projects he gave first-year graduate students.
6. Six propositions state what would refute the argument. The paradigm claim is not tested directly; it stands on P1–P6, each paired with named contrary evidence, such as supervised students' certification decisions diverging from independent faculty review.
7. Credentials would attest to explanation and defense rather than completion. Deng contrasts this with a baseline in which HarvardX and MITx completion was 3.13% among participants who viewed course content in 2017–18 and 46.02% within the paid verified-certificate track.

From submitted work to sampled probing

The argument's starting move is a validity claim rather than a cheating claim, aligning Deng with work that treats validity as the central problem of AI-era assessment. A finished product shows that an output was produced, not that its author learned; grading answers was cheap and has become uninformative, while probing is informative and expensive. The proposed instrument is a competence framework stating what every graduate must be able to do, with decisions made by sampled probing at each checkpoint. Two lanes are certified separately: AI-free probes establish what the student can do unaided, and AI-permitted tasks establish whether the student can direct an AI system and catch its errors.

Four conditions make the instrument credible. Decisions combine several probes by different probers, sessions are recorded and audited, and students can appeal. Probes are scored on reasoning rather than fluency, with adjusted formats for second-language students, and students know the probe types and criteria in advance, so the randomness lies in which work is drawn. A person who consults for a student does not certify that student, which removes the incentive to hide confusion and makes consultation compatible with oral assessment at scale.

Why judgment has to be distributed

The second problem is supply. AI tutoring can teach well, Deng concedes, but students cannot rely on it to recognize its own errors or theirs, and an AI-assisted task embeds choices from several fields at once — the methods used, the data relied on, the claims made, the decisions implied. The difficulty is breadth rather than depth, so he conjectures that judgment of such work becomes distributed: no one person judges the whole, and consulting people with depth in each field becomes routine, as second opinions are in medicine. The clinical review by 27 clinicians across 13 specialties is his small-scale demonstration that model outputs fail in ways only domain experts catch.

Consultation is also a competence students must acquire. Knowing when a question exceeds one's own depth, whom to ask, how to frame it and how to weigh the answer is a skill in its own right. This is the paper's connection to evaluative judgment as a recognized aim of higher education, and the reason the same posts can be argued to serve learners and train them at once — the earlier finding that students who give reviews gain more in writing than those who only receive them is Deng's precedent.

A ladder that supplies its own evaluators

The institutional form already exists: paid, trained and supervised teaching-assistant posts. Deng's redesign changes their work from grading to probing and consultation, opens them only in framework areas where the holder has been certified, and organizes them across year levels so responsibility follows demonstrated ability rather than year of study. Depth is matched to the case — a misread formula needs a certified senior student, while a forest-carbon model combining ecological data with economic assumptions needs a graduate student or professor from each field. Professors sit inside the ladder, because without them there is no standard to maintain and no one to train the trainers.

The training claim rests on supervised practice: probing an explanation, diagnosing understanding, checking AI output for errors and deciding when to pass a case upward are the judgments AI-assisted work demands. Deng's contrast is explicit — AI support alone serves students but trains no one, and grading occupies the same hours while cultivating less judgment and offering no cross-field consultation. He is equally explicit about refutation. P1 fails if probe outcomes do not predict independent performance on the same competences; P4 fails if assistants randomized to service diagnose understanding no better than those randomized to grading; P5 fails if the pool does not grow with enrollment. Against the objection that a university could use ladder language to replace paid staff with students, his conditions are competence-based entry, payment or degree credit, supervision, feedback and progressively harder responsibility.

What this means for practice

• Move teaching-assistant hours from grading to probing. The transition needs no new posts and can begin inside existing courses, with assistants running short sampled probes on reasoning and faculty reviewing consequential decisions.
• Write the competence framework before the probes. Certification depends on a statement of what every graduate must do and which competences are certified without AI and which with it; without it, probing has no standard to score against.
• Separate the consultant from the certifier. Assign consultation and certification to different people, and record and audit sampled sessions so decisions can be reviewed.
• Open supervised posts by demonstrated competence, not seniority. Scope each post to framework areas the holder has been certified in, pay or credit it, and stage harder cases after feedback on easier ones.
• Treat personal paths as certifiable, not exempt. A task agreed with an adviser can be probed against the common framework, which is what lets personalized learning sit inside a degree without lowering the standard.

Limitations

• The paper is conceptual: it reports no study of its own, its declarations state that no data were generated or analyzed and that ethics approval was not applicable, and the paradigm claim stands or falls with six propositions none of which Deng tests.
• The 48% practice gain and 17% unassisted loss come from Turkish high-school mathematics with GPT-4, not from undergraduate teaching.
• The teaching-assistant evidence is descriptive: the 18,664-respondent survey covers Chinese doctoral graduates, and neither study measures what grading or probing develops.
• The staffing question is unresolved: probing costs minutes of a qualified person's time for every student at every checkpoint, and P5 fails if supervision consumes the hours released from grading and lecturing.
• The argument is scoped to undergraduate study in knowledge and problem-solving disciplines, and it offers no cost model, staffing ratio or implementation timeline, so a reader cannot estimate what the ladder would require at a given enrollment.

Citation

Deng, L. (2026). University teaching after generative AI: beyond lectures and inquiry to a certification-and-consultation paradigm in which learners become evaluators. EdArXiv preprint. https://doi.org/10.35542/osf.io/q9gry_v1