England’s higher education regulator says artificial intelligence has made it “close to impossible” to assess some university coursework. In a report covered by the Guardian on 11 November, the Office for Students warned that essays and open-book exams can be finished with a model and still look like student work. Josh Fleming called it “a generational risk to the system as it stands.” The regulator’s answer was blunt: more AI-free assessment, including watched written exams and in-person presentations.
A university degree is a public promise that someone learned something and can show it. Coursework kept that promise by letting lecturers infer the process from the product. Staff now say coursework is often written with AI help. Students describe peers who finish in minutes, get comparable grades, and cannot recall basic details of what they submitted. Lecturers receive stacks that look the same and cite the same sources. The finished essay used to stand in for labor. When it no longer does, it stops being evidence.
I grant that claim. What I want to press is what it is really about. The signs an institution used to judge work can now be produced without the work, so the scarce decision is which signs it will still treat as evidence.
When diligence misses the contribution
The same break shows up where the product is not a student essay but a review, and there the counterfeit does not require fraud.
A paper studying 2,000 submissions to ACL, EMNLP, COLM, and ICLR found that the latest GPT and Claude models draw about half or more of the weaknesses they name from a paper’s appendix. Human reviewers drew 8.4 percent of theirs from the appendix, and those notes mostly concerned presentation and the core idea. The models went after fine-grained technical details that were not central to the contribution. The pattern held whether the model was reached through an API, a coding agent, or the web. The authors also note that model outputs often shape the reviews humans then submit.
An appendix is long, specific, and full of statements that can be challenged without holding the argument of the paper. A model can produce a review that looks thorough because it found things to criticize. The inspection is real. The weaknesses can be real. The review is still a sign that inspection happened. What it may not be is a sign that someone understood the contribution well enough to say whether the paper should stand. Once those appendix comments enter a human draft, the human review can inherit a posture of diligence it did not earn.
That is the deeper version of the Office for Students problem. A signal can fail in more than one way. Coursework often fails when the artifact can be manufactured without the underlying work. Peer review can fail when the activity still happens and the visible result no longer establishes the conclusion the institution wants. The first invites stronger checks on how the artifact was produced. The second demands better judgment about what the artifact actually demonstrates. The visible trace of diligence can detach from the judgment the review exists to support even when no one is cheating.
The exam the regulator prefers is a design response to the first failure. A closed-book test in a room is harder to fake because the student has to produce the answer while someone is watching. Presence raises the cost of the sign. Presence alone does not make the exam a good measure of what the degree claims. Oral exams, credential gauntlets, and elaborate interviews can all be expensive and still miss the capability an institution cares about. The better principle is narrower: a viable signal needs a cost or constraint that remains coupled to the thing being inferred. The regulator is choosing a sign whose production still includes presence. That is coherent. It is also a confession that the old sign broke, and a reminder that the next sign has to be designed, not merely made harder.
The sign, not the signature
Call a counterfeit signal a visible trace that continues to carry evidentiary weight after its relationship to the underlying work has weakened or broken. Counterfeit, here, is about appearance and evidence. It does not require fraudulent intent. A polished essay may be deception. An AI-assisted review may be misplaced effort. A cover letter that no longer sorts candidates is a signal that lost discriminating power. AI accelerates the break by collapsing production costs, but the break is the lost relationship, not the cheaper tool. In each case the institution is still inferring work from a surface that no longer carries that information.
In the Answerability Budget I asked who will put a name on output when generation outruns the people willing to stand behind it. This is a different object. The signature answers for a claim. The sign is what everyone else uses to decide whether the underlying work happened. You can have a name on a pile nobody inspected. You can also have a busy inspection that never touched the part that mattered.
Once a proxy’s cost collapses, institutions either change the metric or ration the signal. Romance authors can meet a monthly Kindle cadence with AI help; pages read remain a harder proxy than books uploaded. Hiring shows the design problem more clearly. For decades a tailored cover letter worked because writing one cost time. Michael Spence’s signaling theory said a signal carries information when it is expensive to send. AI makes the expensive letter cheap for everyone, so the letter stops sorting candidates. Greenhouse, the hiring platform, answered by letting applicants flag one application a month as their top priority. The flag costs something AI cannot reduce: this month’s only flag. Daniel Chait, Greenhouse’s CEO, has said those flagged candidates are hired at roughly five times the rate of everyone else. That is a reported association, not a controlled result. People who spend the flag may already be more selective, better matched, or more motivated. Scarcity can make a signal harder to spray. It does not automatically make it more accurate. The design move still matters: restore a relationship between an observable act and the seriousness being inferred, then keep asking whether that act still tracks the capability you care about.
What a room agrees to treat as proof
Institutions still need signs. People cannot inspect every essay, every paper, every application, every memo. Coordination depends on some visible trace standing in for work that will not be re-done by every person who has to decide.
Once a sign can be produced without the work, keeping it as evidence is a choice, not an inheritance. Some teams will keep the cheap signs and call the resulting pile progress. Some will invent harder ones: an exam in a room, a flag that can be spent once, a review that has to name the core claim before it is allowed to nibble at the appendix. Cost alone will not save them. The cost has to stay coupled to the capability the institution wants to know about.
Mona Lisas and Readymades separated provenance from authorship. A sign is a third thing: neither where the text came from nor who answers for it. It is what a room of people agrees to treat as proof that the work happened. Provenance tells us where something came from. Answerability tells us who owns it. Signals tell us what we are willing to infer from it. The difficult part is that institutions still have to agree on what counts as evidence, even as the technologies producing that evidence keep changing.
Madam I’m Adam
Discover more from Adam Monago
Subscribe to get the latest posts sent to your email.
Leave a Reply