Deutsch - English - Español - Français - Italiano
When anyone, or anything, may have made the paper handed in, evaluation moves back to the making. Third in the four-part series on what the machine is doing to school, and on who will decide what happens to the human.
The paper is in. Three days later, a voice starts asking about it.
Start in New York, autumn 2025, at NYU’s Stern School of Business. Professor Panos Ipeirotis noticed the preparatory assignments in one of his courses were coming in “suspiciously good”. In his phrase, they were worthy of a McKinsey memo that had survived three rounds of editing. He started cold-calling students in class. Many who had turned in thoughtful, well-structured work “could not explain basic choices in their own submission after two follow-up questions”. Some could not answer at all. The old system, in which the paper handed in measured understanding, he concludes, is “dead, gone, kaput”. His response, worked out with colleague Konstantinos Rizakos, was fighting fire with fire. A voice agent conducts personalised oral examinations. Three different language models grade the transcripts. The whole thing costs 42 cents per student. The proposal is on his blog and written up as a paper.
The pattern learning-design specialists now describe has two steps. The student hands in his written work, as he always has. Two or three days later, he defends it for ten to fifteen minutes to a voice agent that has read his paper and adapts its questions to the answers it gets. The grade no longer marks the paper. It marks the ability of the presumed author to answer for it.
The Center for Democracy and Technology survey already cited in this series has 71 % of USAmerican teachers reporting that when pupils use AI on their assignments, it is hard to tell whether the work handed in is really the pupil’s own. And the automatic detectors supposed to settle the matter turn out to be inadequate, with false-positive rates that rule them out as evidence. The paper handed in, whether essay, report or code, has stopped being a reliable signal. Anyone, or anything, may have made it, and nothing in the object itself tells you which. Yet all of modern assessment rested on that object. When the product proves nothing, only one direction is left for proof to move in: upstream, back to the making.
The oral makes a comeback as the watertight test. And the machine that killed the paper administers it.
The first route back upstream is the oldest in the academic world. The disputatio was the medieval university’s exercise in which the candidate defended his theses orally against objections, the only way of establishing that he actually possessed what he was advancing. The University of Pennsylvania now pairs oral defences with written work, in the name of academic integrity as much as critical thinking. And a three-year Australian study of 290 students across 25 subjects (Droulers, Krautloher & Shaeri, Teaching in Higher Education, 2026) finds interactive oral assessment beats written work on both validity and reliability, reduces fraud, and is preferred by students. The machine can write your paper for you. It cannot sit your oral for you.
The objection to the oral is old and known. It cannot be generalised: having every student defend every assignment costs a wage bill that written work had precisely saved. That is why written work had won in the first place.
The machine, though, lifts that objection. The voice agent that questions the student costs less than half a dollar per session, does not tire, and can question as many students in parallel as needed. The technology that broke the paper as evidence is now put to work administering the test that replaces it. The failures of the first New York version are documented by its designers themselves. The professor’s cloned voice was perceived as “shouting” at students. Questions came in bursts despite the instructions. And 83 % of students found the test more stressful than written work, which is easy to understand: 83 % had never sat an oral in their lives. Even so, 70 % acknowledge the format tests genuine understanding. The machine destroyed the paper as evidence, and the same machine now verifies the process. Readers of this series will recognise the pattern. They will meet it again, on a bigger scale, in the final episode.
In Hangzhou, proof does not wait for the assignment. The pen watches the hand write.
The second route goes further, and the previous episode left it in reserve. At the Xiang Lake Future School near Hangzhou, Gao Yilin, fifteen, writes her mathematics formulas with a pen whose tip contains a miniature camera and whose barrel holds a pressure sensor. Every stroke, its shape and the force applied, is captured as data. “The AI analyzes it and immediately tells my teacher how I approached the problem and exactly where I got stuck,” the pupil explains. Moments after a test, her teacher walks over to her desk. The answer was correct, but there was a simpler way to solve the equation, and she should think it through again. Data from the pens allow teachers to identify struggling pupils quickly, down to where they are seated in the classroom. The instrumentation extends to the canteen tray, which counts calories, and to the wristband that tracks heart rate in sports class. “This isn’t surveillance, it’s a form of human-AI cooperation in education,” the school’s principal, Ye Cuiwei, assures. [Quote to be checked verbatim against the archived original before publication.]
The Hangzhou logic goes further than the oral’s. The oral verifies the process after the fact, by questioning the memory of the making. The pen captures the process at the source, while it is happening. Nothing is left to verify, because everything has been recorded. The paper is no more than a by-product of the trace. Nobody in Hangzhou is hunting cheats. The device does not respond to product contamination; it serves the fine-grained steering of pupil progress, aimed at the terminal examination that, as the previous episode showed, stays intact. The two worlds and motives are different, but the displacement of proof is the same.
The researcher does not escape either. The paper comes with its fabrication note.
The third route climbs to the storey the first episode of this series described, the laboratory. In its experiment with papers signed by AI as “first author”, the journal of East China Normal University has set a three-part publication protocol: the text, the editorial committee’s recommendation, and a note describing the writing process. The note is no ornament. It is the admission that the signature no longer establishes provenance, an admission the first episode already flagged. The researcher, last link in the educational chain, is under the same regime as the pupil in Hangzhou and the student in Pennsylvania. What is handed in counts only as documented by the way it was made.
Three storeys, one rule. And everywhere the new instrument aimed at an old objective.
A Chinese middle school, Western universities, an academic journal in Shanghai. Three uncoordinated fields, three populations (pupil, student, researcher), three different motives. And a single displacement: proof leaves the product and moves back to the process. When anything can have made the object, the object testifies to nothing. Only the making testifies: observed live (the pen), reconstructed under questioning (the oral), or declared under protocol (the note). It is, to our knowledge, the most solidly established result of the whole inquiry, precisely because the three fields are independent.
The convergence, though, has a second face. Everywhere, the new instrument aims at an old objective. In the West, the climb back to the process is about catching cheats and restoring the old economy of proof, not founding another one. In Hangzhou, it sharpens preparation for the same examination. In Shanghai, it lends credibility to an institutional experiment. Nowhere does it come from a redefinition of what school must produce. The measuring apparatus of the new regime everywhere precedes the question of purposes. And each route has its price, which none of the three fields displays openly. The generalised oral means paying humans, or handing the judgement of humans over to the machine. Capture at the source installs an infrastructure of complete childhood surveillance, whatever name it is given. The declarative note is worth only the word of the one who signs it. The choice between these prices is political, not pedagogical.
The assignment now counts only as exercise. And the machine is banned there for the pupil only.
The previous episode established, with a randomised trial in support, that delegating the work to the machine destroys the school’s product, learning itself. This episode establishes that the paper handed in proves nothing anymore. Put the two together, and a consequence follows that nobody owns up to aloud: school output has no value as product, and yet stays indispensable as means. The essay nobody will read, the problem solved a thousand times before, the physics experiment with a known result count only through the transformation of the one who makes them. As the dumbbell counts only through the muscle it forces into being, which is why nobody brings a forklift into a weight room.
The consistent school would ban the tool from the pupil’s hand. Against the rest of the world, it would become the one social place where the universal instrument is forbidden, on principle, to those being trained there. No institution owns this in those terms. All approach it in practice, one watertight test at a time: the oral where the machine cannot whisper, the supervised written exam, the work defended in person. The asymmetry is the real face of this school: the machine banned in the pupil’s hand, and omnipresent in the institution’s, questioning the student, capturing the pen stroke, grading on transcript. The ban applies only on one side of the desk. The inversion undoes the question this whole dossier started from. It is no longer how much machine to let into school, but how much school to preserve from the machine, and for whom the preservation is to hold.
The voice agent, the pen and the note all belong to someone. Proof has moved back to the process. Who owns the apparatus that records the process, and, through it, the human being made there? That is what the next episode is about.
All our publications are free and Copyleft. They may be republished, in whole or in part, under following conditions:
→ keeping the title and content
(text, images, editor’s or translator’s notes, boxes)
→ crediting the author(s) and/or translator(s)
→ inserting a clickable link to the source.
Any monetization or commercial use of these publications is prohibited and will be subject to legal action.